We recommend you first read our original analysis here.

We like to refer to this phenomenon as the short-seller flow effect, and given the positive reception of last week’s article, this felt like a natural follow-up.

To Borrow or Not to Borrow?

One of the toughest obstacles practitioners face when short-selling individual names (and one we have repeatedly encountered in live tests as well) is simply being able to borrow shares in the first place.

When locates are available at all (which is far from guaranteed), their fees can be punishing and significantly erode whatever edge the signal carries: as capital committed to this type of strategy grows, the borrow problem only gets more painful.

But constraints can sometimes be a useful prompt for creative thinking, and the signal we present today is, in fact, rooted in one we initially saw as a dead end.

The Signal

We decided to investigate long-side opportunities in the same ticker that our original signal would flag as promising candidates to fade.

More specifically, the strategy buys stocks at the closing auction using market-on-close (MOC) orders and liquidates them at the following session’s open using market-on-open (MOO) orders, thereby isolating the overnight return component.

Ironically, the very characteristics that make these stocks attractive short candidates intraday may also create the conditions for persistent overnight rebounds.

The stock selection steps remain identical to those described in our original piece, with only the buy/sell rules changing: full details, as always, are disclosed at the end of the article.

Positions are sized at a fixed 10% portfolio weight: this means that, on days when fewer than 10 stocks are identified by our criteria, the portfolio will also be less than 100% invested.

The gross-of-costs results are displayed in the plot below.

Within our backtested sample (2012–2026), we report a CAGR of 37.1%, a Sharpe ratio of ~1.6, and a maximum drawdown around 35%: notably, unlike the original variant, the equity curve shows no signs of structural breaks in recent years.

This signal lends itself to being traded via market-on-close (MOC) and market-on-open (MOO) orders, which minimise slippage by allowing investors to participate directly in the closing and opening auctions of each session.

Considering that this time there are no locate fees to pay, and that the average P&L per share stands at ~$0.042 (against standard-tier IBKR commissions of $0.0035 per share), we can say that the gross-of-cost picture is encouraging enough to justify further work.

We notice that our selection criteria consistently flag numerous qualifying stocks, with the rolling 21-days average number of daily entries never dropping below 5 across our sample (see the chart above).

This suggests there is enough room to investigate additional features (predictive of positive overnight returns) to be combined with our raw signal and used as Boolean operational filters: these trade/no-trade conditions are typically introduced with the aim of enhancing net-of-fee risk-adjusted returns, since they naturally limit trading activity.

A word of caution, however: introducing additional predictive components also means introducing more complexity into the framework. Any new feature should therefore be both logically grounded and accompanied by a meaningful improvement in performance… one significant enough to justify the increased model risk.

Why Does This Signal Work?

We can think of at least three forces that plausibly contribute to the positive overnight returns we observe.

Short Covering Pressure

From studying how active and aggressive short sellers operate within the small-cap space, it appears that many tend to reduce or fully liquidate their positions late in the trading session rather than carry significant overnight exposure.

Many short sellers seek to reduce overnight exposure late in the session. The resulting buy-to-cover flow may create demand into the close, particularly in thinner names where liquidity is limited. If this flow is only partially absorbed, or if some covering activity continues through after-hours and pre-market trading, part of the associated price pressure may be reflected in the following session's opening price.

The Overnight Risk Premium

The second force at play here is the well-known overnight risk premium, a topic we already touched on in this article.

Researchers have consistently documented that a disproportionate share of long-term equity returns accrues overnight rather than intraday, suggesting that investors holding through the close-to-open period are compensated for bearing event risk (i.e. gap risk): precisely the risk that intraday short sellers spend their careers avoiding.

In gaining exposure to the overnight risk premium, we are simultaneously absorbing the event risk that short sellers are keen to offload at the end of the day: a legitimate source of risk that, we believe, may justify a portion of the returns we observe.

(High) Attention Stocks Effect

Haghani, Ragulin and Dewey (2022) suggest that attention stocks tend to earn consistently positive overnight returns. These are defined as well-known names that attract retail interest, media coverage, options activity and strong recent momentum, characteristics that, the authors note, also tend to coincide with higher beta and volatility.

In a sense, the names our signal flags can be considered attention-grabbing and volatile almost by construction. Our selection criteria may therefore be identifying some of the attention stocks that, within the authors’ framework, tend to appear among the strongest overnight performers.

In practice, we do not believe the effect is driven by a single mechanism alone. Rather, the signal may emerge from the interaction between structural short-selling behavior, overnight risk transfer, and the tendency of speculative attention to concentrate in volatile small-cap names.

Conclusion

The idea we presented in our original piece, although strong in historical performance, showed some pain points in practical implementation: namely, the need to account for locate-fee drag and a structural break that appears to have occurred around the start of 2022, which raises some doubt as to whether the signal will return to performing in line with its historical metrics.

That said, we hope this article serves as a useful illustration that even when an idea does not immediately pan out, digging into adjacent territory can sometimes uncover something valuable.

It should also serve as a reminder that identifying market effects is only a portion of the work that quants do: reasoning about what drives them matters just as much, and the answer is rarely attributable to a single factor, but rather to a more complex combination of multiple forces.

If you found this article useful, feel free to leave a comment and reach out via direct message or email at info@concretumgroup.com for any questions.

Full Methodology

In the following table, we describe step by step the rules we used to construct the strategy presented in this article.

Read more

Then stuff breaks. Not the strategy logic - the infrastructure around it.

Over the years, we’ve repeatedly seen traders run into the same production issues after going live. Some are annoying. Some are expensive. Most are completely invisible until they happen for the first time.

Based on those conversations and our own experience running production trading systems, we put together a list of 10 operational pitfalls that every trader should understand before going live.

These examples are drawn from Interactive Brokers specifically, but the broader lessons apply to most broker APIs and automated trading infrastructures.

1. Your Data Can Go Stale and Your App Won't Know

Someone reached out to us after their algo traded all afternoon on a price that hadn’t updated since 11 AM. The app looked completely normal. Connection status said connected, prices were still showing on screen, nothing appeared broken.

Earlier that day, they only had a brief internet disconnect. The broker gateway reconnected automatically, everything came back online, and from the outside it looked like the system recovered fine.

But it hadn’t.

The market data had silently stopped updating.

No error. No disconnect warning. The last price just stayed there looking valid, and the algorithm kept making decisions based on a number that was hours old.

That’s one of the dangerous parts about broker APIs. A connection can look perfectly healthy while the actual market data stream is dead. Your system sees a price, assumes it’s current, and keeps trading unless you explicitly check how old that data is.

The fix is simple in theory: track when each ticker was last updated and stop trading if the data becomes stale. A lot of people don’t realize they need this until it happens in production.

2. Two Algorithms + One Ticker = Cancelled Orders

Two algorithms running on SPY. End of day, Algorithm A submits a MOC BUY for 200 shares, Algorithm B submits a MOC SELL for 150 shares.

IBKR cancelled one of them. Error code 202: “Order would cross.”

Each algorithm tracks its own position correctly. But IBKR sees one account with a BUY and a SELL for the same symbol and rejects the conflicting one. No prominent documentation about this.

The fix: Monitor for MOC cancellations via order status and error events. If a MOC gets cancelled, schedule a fallback market order before close. Or net out positions across all strategies before submitting close orders so you’re only sending one side.

3. Historical Data Rate Limits Are Worse Than Documented

IBKR’s docs basically say: “don’t request too much data too quickly.” In practice, the limits are much easier to hit than most people expect.

A few historical data requests are usually fine. But once several requests happen at the same time, IBKR can start throwing pacing violations or returning empty responses. The dangerous part is that the failure is not always obvious. Sometimes the request simply returns no data and the algo keeps running with missing indicators.

This gets worse throughout the day when indicators like VWAP repeatedly trigger historical data requests across multiple tickers. Even small bursts can eventually hit pacing limits because IBKR tracks request frequency over time.

The real fix is understanding IBKR’s pacing limits extremely well or using a dedicated market data provider instead. IBKR itself explicitly states that it is not a specialized market data provider and recommends using one if your strategy requires heavier historical data usage. Otherwise, you need to aggressively manage request timing, caching, and batching to avoid random throttling or silent failures during live trading.

4. MOC Orders Have Submission Deadlines

Market-on-Close orders can’t be submitted or modified after 3:50 PM ET on the NYSE. If your algorithm tries to close positions at 3:55 PM via MOC, the order gets rejected.

The fix: Submit MOC orders before the cutoff (we use 3:49:55 ET) and have a fallback market order ready at 3:59:58 in case the MOC gets rejected.

5. Early-Close Sessions Can Silently Break Your Logic

A handful of U.S. trading sessions close early each year - typically at 1:00 PM ET instead of 4:00 PM. The day after Thanksgiving, Christmas Eve, and a few others.

Most systems implicitly assume the market always closes at 4:00 PM. On an early-close day, that assumption quietly breaks everything downstream. Your MOC orders get submitted hours after the market already closed - rejected. Your intraday indicators keep computing on a session that ended at 1:00 PM. Your end-of-day position closing logic fires at 3:49 PM into a market that’s been shut for nearly three hours.

Nothing crashes. The system just does the wrong thing at the wrong time.

We’ve also seen that some data providers continue emitting bars after the official close on half-day sessions, while others truncate the session correctly. If your system consumes that extra data, your indicators and signals are contaminated without any obvious error.

The fix: Maintain an official exchange calendar and dynamically adjust your scheduling logic and MOC cutoffs based on the actual close time for that day. Don’t hardcode 16:00.

6. Your Server’s Clock Is Probably Wrong

Your system schedules checks at specific times – “execute at 10:29:58”, “send MOC at 15:49:55”. A typical computer clock can drift several seconds per day. If your VPS hasn’t synced in a while, your MOC order that should arrive at 3:49:55 actually arrives at 3:50:02 – rejected.

The fix: Keep your server clock synced and monitor for clock drift. Even a few seconds off can cause scheduled orders to arrive too late.

7. Multi-Account Routing Fails Silently

If you use a single gateway connection for multiple accounts, every order must explicitly specify the account. If you don’t set it, IBKR routes to the default account. No error, no warning, just the wrong account.

This becomes particularly dangerous in multi-client or advisor-style setups, where a routing mistake can unintentionally allocate trades to the wrong portfolio.

8. Reconcile Against the Broker on Every Restart

Even if you persist your state in a database, you should still reconcile against IBKR’s execution records on startup. Your DB might have missed a fill – maybe the app crashed between receiving the fill event and writing it. Maybe a partial fill came in after the crash. Maybe someone manually traded in the account while your system was down.

The fix: On startup, query IBKR for today’s executions and compare them against your own records before making any new trading decisions. Tag every order with a reference that includes the strategy name and date (e.g., SPY-MyStrategy-20250227) so you can filter which executions belong to which algorithm. If there’s a mismatch between what IBKR says happened and what your DB says happened, trust the broker.

9. Paper Trading and Live Trading Don’t Behave the Same

This one ties back to several points above. IBKR’s paper trading environment does not enforce the same rules as live. For example, conflicting MOC orders (see #2) will be accepted on paper – both sides go through, no Error 202, no cancellation. You test your multi-algo setup, everything works, you go live, and suddenly orders start getting rejected for crossing.

Paper trading will not catch every production issue. Some broker-level validations, order conflicts, and margin checks only kick in on a live account. If your entire testing was done on paper, expect surprises on day one.

Some traders work around the crossing issue by using equivalent instruments on different sides – going long SPY on one algorithm and short VOO or a Vanguard S&P 500 ETF on the other, for example. That avoids the same-symbol conflict at the broker level. But however you handle it, you need to be actively managing how your algorithms interact at the account level. The broker doesn’t care about your internal strategy separation.

These operational issues become even more critical when strategies are deployed across dozens of separately managed accounts, where synchronization, execution consistency, and routing reliability become central.

10. DST and Timezone Mismatches Will Break Your Scheduling

U.S. markets and European countries do not switch to Daylight Saving Time on the same dates. The U.S. springs forward in March, Europe follows a few weeks later. In the fall, the reverse happens. For about two weeks twice a year, the time offset between your server and the exchange shifts by one hour.

If your server is in Europe and anything in your stack references local time instead of market time, your entire schedule silently shifts. MOC orders go out an hour late. Your algorithm misses market open. Signal calculations run on incomplete sessions. Rebalancing triggers at the wrong time.

The dangerous part is that nothing crashes. The system runs normally - just at the wrong hour. And because it only happens for a couple of weeks a year, it’s easy to miss in testing.

The fix: Always use the exchange’s timezone (America/New_York for U.S. equities) for all scheduling and time comparisons. Never derive trading schedules from your server’s local clock. Test your system across DST transition dates specifically - don’t assume that because it works in January it will work in March.

This list is obviously not exhaustive. What breaks depends on your broker, infrastructure, execution architecture, and the complexity of the systems you run.

But one lesson tends to repeat itself consistently: building a strategy is often the easy part. Building a robust production environment around it is usually much harder.

In recent years, the rise of LLMs and AI coding tools has dramatically lowered the barrier to creating trading algorithms. What has not become easier, however, is dealing with the operational realities of live execution: broker-specific edge cases, multi-account routing, order synchronization, failover handling, stale data detection, reconciliation procedures, and execution reliability under real market conditions.

These are not concepts most people learn from textbooks or backtests. In many cases, they are lessons that only emerge after years of deploying and maintaining live trading infrastructure.

At Concretum Group, part of our work involves helping private investors, advisors, and professional allocators automate systematic strategies efficiently across Interactive Brokers environments, including multi-account and portfolio-based execution workflows.

If you are exploring trading automation and would like support in designing a more robust production setup, feel free to reach out to us by email at info@concretumgroup.com

If you found this article useful, feel free to leave a comment and reach out via direct message or email at info@concretumgroup.com for any questions.

It felt revolutionary. For decades, retail investors had paid brokers $5, $10, sometimes even $20 per trade. Suddenly, an entire generation of market participants could buy stocks and options for free, directly from a smartphone.

Wall Street hated it at first…. but later copied it.

Within a few years, essentially every major U.S. broker had moved to a zero-commissions model. To millions of retail traders, it looked like finance had finally been democratized, but beneath the surface, a new business model was taking shape.

The Business Hidden Behind Robinhood

Most retail investors think Robinhood makes money from subscriptions, margin loans, or idle cash balances. Those surely matter, but the real engine behind the entire zero-commission revolution was something called Payment For Order Flow, usually abbreviated as PFOF.

The mechanism is simple.

When a Robinhood user presses “buy,” the order is usually not sent directly to a public exchange. Instead, it is routed to a large market-making firm (a wholesaler) which executes the trade and pays Robinhood for the privilege.

Robinhood gets paid, the trader gets “free” commissions, and wholesaler receives the order flow. Everybody appears to win… at least on paper. In reality, the economics turned out to be far more asymmetric than most people realized.

The Big Three

A remarkable paper “Retail Trading in Options and the Rise of the Big Three Wholesalers” by Bryzgalova, Pavlova, and Sikorskaya (2023) documented what happened next. The authors found that retail options trading exploded during the Robinhood era, eventually reaching more than 60% of total market volume. And even more striking, nearly 90% of the payment-for-order-flow business in options became concentrated in just three firms: Citadel Securities, Susquehanna, Wolverine.

The authors refer to them as the “Big Three Wholesalers”, representing one of the most extraordinary examples of concentration in modern finance.

Millions of retail investors believed they were independently trading markets from their phones, but in practice, an enormous fraction of those trades flowed through the same three firms.

Why Retail Flow Is So Valuable

This is a passage most people misunderstand: the wholesalers are not primarily making money because they are “betting against retail.” That framing is too simplistic: retail order flow is valuable because it is statistically predictable, relatively uninformed, and associated with wide effective spreads, and such combination is financial gold. Imagine these two market participants:

• A hedge fund trading around earnings announcements with sophisticated models.

• A retail trader buying weekly Tesla calls after reading a tweet.

Which flow would you rather internalize? Of course the second one.

Not because retail traders are unintelligent, but because their trading tends to be smaller, more directional, less informed, and above all… less sensitive to execution quality. This last point turns out to be critical; professional traders obsess over spreads, slippage, and execution quality, whereas retail traders mostly focus on one thing:

“Is the commission zero?”

The Invisible Fee

This created one of the most successful psychological reframes in modern finance, through which brokerages were able to remove the visible fee and replace it with an invisible one.

The paper estimates that retail traders lost billions through embedded trading costs hidden inside the bid-ask spread: the authors estimate aggregate trading costs of roughly $6.4 billion in less than 2 years. These numbers are astonishing, and far larger than explicit commissions could ever amount to.

Put in other words, retail investors stopped paying brokers directly and started paying market makers indirectly.

Robinhood’s Defense

To be fair, Robinhood pushed back strongly against criticism of payment for order flow. In a 2021 post on its website, the company argued that retail investors often receive price improvement relative to the quoted NBBO (National Best Bid and Offer), and that PFOF is therefore not necessarily harmful: mechanically, that is true.

A retail investor buying a stock or option may indeed receive a slightly better execution price than the publicly quoted best offer. In that sense, wholesalers sometimes improve execution quality rather than worsen it, and Robinhood’s defense was therefore correct.

But the Bryzgalova-Pavlova-Sikorskaya paper is making a different point entirely, with the key distinction being subtle, but crucial:

“Better than the quoted price” does not necessarily mean “cheap trading.”

Suppose the “fair” midpoint price of an option is $1.00, while the bid stands at $0.90 and the ask at $1.10. If a Robinhood user buys at $1.08, the investor technically received “price improvement” relative to the displayed ask price of $1.10, but economically, the trader still paid a very large spread: that is the deeper issue originally highlighted by the paper. At the same time, retail traders were overly concentrated in products where spreads were already structurally enormous, especially ultra-short-dated options. Therefore, both statements can simultaneously be true:

• Robinhood users often received price improvement relative to quoted markets.

• Retail investors still paid very large implicit trading costs overall.

The debate around PFOF often focuses on whether investors saved a fraction of a cent relative to the NBBO, but the paper suggests this argument may miss the bigger picture entirely, with retail investors being more concentrated than ever before in ultra-short-dated options (where the spreads themselves were extremely expensive).

How Robinhood Created a New Wave of Option Traders

Another important insight from the paper is that the retail-trading boom was not primarily about stocks (as one might expect), but it was about options. Even though only a fraction of users traded them, options generated in aggregate dramatically more payment for order flow than equities.

Why? Because options are vastly more profitable vehicles for market-makers and wholesalers. Spreads are wider, pricing is more complex, and retail traders (as mentioned before) are less sensitive to execution quality: on top of that, the leverage embedded in options arguably fuels addictive trading dynamics.

The paper also finds that retail investors overwhelmingly preferred call options, cheap contracts, and especially ultra-short-dated maturities: the authors report that nearly half of retail option trading volume was concentrated in contracts with less than a week to expiration.

And the average quoted spread on these contracts? About 12.6%.

For comparison, spreads in large-cap equities are often measured in basis points.

The Rise of Lottery Finance

In a sense, weekly options were the perfect product for this new smartphone-trading era: cheap, leveraged, fast-moving and emotionally stimulating. A trader could potentially turn a few hundred dollars into thousands overnight… or lose everything in a matter of days.

Another important point is that these products felt inexpensive because the upfront premium was small (a $150 weekly option feels cheaper than buying $20,000 worth of stock), even though the embedded trading costs were often enormous on a relative basis.

The paper documents that retail investors strongly favored tiny “micro” trades, often below $250, where spreads were especially punitive: seeing the broader picture, it is clear how this approach to markets can’t be considered traditional investing…. it is closer to what one might call “lottery finance.”

From Weekly Options to 0DTE

In hindsight, the paper may have actually understated the long-term trend behind this new way of interacting with financial markets: the studied sample mostly predates the full explosion of 0DTE options (i.e. contracts expiring the same day).

Today, these ultra-short-dated options make up a dominant share of the total liquidity exchanged in option markets. The logic is identical to the weekly contracts cited in the paper, just accelerated:

• lower upfront cost,

• higher leverage,

• more frequent trading,

• stronger gambling dynamics,

• and even more flow for wholesalers.

Weekly options were probably just the transition phase, with 0DTE being the ultimate endpoint.

A Deeper Lesson

It is very tempting to turn this story into a morality tale.

Robinhood bad. Citadel evil. Retail traders reckless.

But we must recognize that reality is more nuanced. The wholesalers genuinely provide liquidity and convenience, retail investors did gain unprecedented market access, and finally trading commissions really did collapse.

On one side it is undeniable that the “system” generated tangible benefits, but it also created a massive transfer of economic rents toward the firms best positioned to monetize uninformed order flow.

This is where we find that the deeper lesson hides: modern finance rarely eliminates friction… it usually hides it.

The visible fees disappear… and the invisible spreads expand.

As trading frequency rises, apps become more entertaining and gamified, while the “hidden” infrastructure layer quietly profits from the economics of this new digital lottery. In the end, what promised to be a retail revolution did not weaken Wall Street intermediaries… it made them more powerful than ever.

This article was written by Concretum Group CEO Andrea Barbon, professor at the University of St. Gallen and an expert in blockchain, decentralized finance, asset pricing, and monetary policy.

If you found this article useful, feel free to leave a comment and reach out via direct message or email at info@concretumgroup.com for any questions.

The series, authored by Jack Schwager, began in 1989: what has made it so enduring is not the trading insights alone, but the human stories behind them, of rigor, discipline, and the ability to learn from failure.

The sixth book in the Market Wizards series is due out this June, and ahead of its release, Jack Schwager and co-author George Coyle joined host Moritz Seibert on the Top Traders Unplugged podcast to discuss what readers can expect.

What caught our attention is that the book profiles a new generation of traders, some of whom built their initial fortunes doing things the original Market Wizards would have likely warned against… like shorting small-cap stocks.

Driven by curiosity, we decided to revisit a signal that has been sitting in our internal archives for some time, one we believe lends itself to some important considerations around intraday short selling.

Identifying Stocks to Fade

Over the past years, some of our internal research has focused on short-term opportunities in single stocks. Notably, one particular family of signals kept emerging as meaningfully predictive of intraday returns, especially within the less covered segments of the US market.

We document a distinct tendency for stocks that experience abnormally large overnight gaps to revert those moves during the subsequent intraday session: at its core, such gap-fade behavior can be interpreted as a short-term mean reversion effect.

This dynamic is also well supported in the academic literature: Akbas et al. (2022) point to a persistent tendency for positive overnight returns to be followed by negative daytime reversals in US stocks, while Stübinger and Schneider (2019) build a statistical arbitrage strategy around the same mean-reverting property, finding it to be particularly significant in the first two hours after market opening.

The specific stock selection criteria we tested are summarized at the end of this article.

Our analysis covers the period from January 2012 through mid-May 2026, and the gross-of-costs results (shown in the chart above) are striking: the headline figures report a CAGR of roughly 98% and a Sharpe ratio of about 2.08, with a maximum drawdown of around 67%.

Therefore, even though this signal can look risky on the surface, it has historically carried substantial predictive power, suggesting that the new generation of market wizards has indeed been exploiting a genuine edge.

But before getting carried away, we think such short-selling framework comes with important caveats, and there are a couple of central points worth discussing in greater detail.

The Bad News

If you are not an experienced quant, the equity curve above is probably already getting you excited about two things: becoming wealthy in a matter of years and earning your place in the next edition of Market Wizards by following a simple set of rules.

Unfortunately, reality is a lot different.

Market Frictions

Identifying a strong signal is not the same as having an economically viable strategy. Our analysis was intentionally kept gross of costs, because that is what the first steps of quantitative research look like: we isolate the raw predictive component before layering on the real-world frictions.

A complete research pipeline would then move on to estimating those costs, namely commissions, market impact, and, in this case, the cost of borrow: unfortunately, this signal we tested ticks every box for facing brutal frictions.

Given that it operates intraday, trading activity is elevated, making fees a meaningful component of the total cost equation.

On top of that, there is market impact. We typically model it through the I-Star framework, which punishes large orders in illiquid and volatile names. We expect extreme gappers, by design, to sit at the volatile end of the spectrum, and although our universe already excludes roughly the bottom 30% least liquid names in the Russell 3000, the remaining stocks still include plenty of names with thin liquidity.

To top it off, shorting introduces its own set of obstacles. Locating shares to short often requires specialist brokers that deal in hard-to-borrow (HTB) names, since every short sale must be backed by a confirmed locate. For some tickers locates may be unavailable precisely when the opportunity to go short presents itself, and when they are available, their cost can be punishing.

Put together, any realistic net-of-cost picture would look meaningfully different from the 98% gross-of-cost figure shown above.

A Structural Break?

From 2022 onwards, something clearly changed. The equity curve peaked in the first half of that year and has since given back a substantial portion of prior gains in a drawdown that continues through today. In the language of academic finance, the abrupt and persistent breakdown of a well-specified signal is known as a structural break.

The most plausible explanation we have found is rooted in crowding effects: in recent years, as short-selling infrastructure has become more accessible and data more widely available, more capital began competing for the same set of (attractive) opportunities.

Adapting for Survival

The traders profiled in the new book were clearly benefiting from an edge that was genuinely strong while they built their track records, and we suspect that their discretionary skills played a major role in amplifying that edge as well as absorbing the market frictions discussed earlier.

For instance, in a separate study of ours, we have documented that a layer of discretionary decision-making by a skilled trader can turn an otherwise unprofitable systematic signal into a profitable one.

Whether the same approach to shorting small-caps can sustain a new generation of wizards is a much harder question to answer optimistically, and it brings us to a broader point on the importance of adaptation.

In the 1980s, the Turtles became wildly successful by operating relatively fast trend-following rules that wouldn’t generate the same impressive results if applied mechanically today: the edge was exceptionally strong when it was first captured, and then it gradually faded.

The same may well be true of the intraday gap-fading approach that some of the new Wizards have harvested for several years.

The ability to adapt as signals become stale is, in our view, a defining quality of the durable trader, whether discretionary or systematic. Markets are a bit like the sea… moving in ebbs and flows: even when the surface looks calm and everything is going well, we must be prepared, as we never know when the tide will change.

We would be glad to hear from traders who have made short-selling a cornerstone of their approach. Feel free to leave a comment, or reach out by direct message or email at info@concretumgroup.com with any questions or thoughts.

Full Methodology

In the following table, we describe step by step the rules we used to construct the strategy presented in this article.

Read more

Regardless of model specifications, we keep observing the same pattern: small implementation details can produce surprisingly large differences in realized performance.

As we like to say… dispersion is in the details.

This theme has come up across several recent articles, from position sizing approaches to reverse-engineering the rules of a legendary portfolio manager in the CTA industry.

You can find our previous articles at the links below.

Today we test something new.

We take a simple intraday trend signal and study four distinct exit policies, treating this experiment as a practical exercise to reason about how a single design choice should (or should not) be made.

Subscribe now

Study Setup

All tests are performed on the iShares Semiconductor ETF (SOXX), an asset class that has been firmly in play for several weeks, having staged a historically remarkable run-up through new all-time highs.

For all tests, we use a signal based on ATR band breakouts, something we already presented in a piece earlier this year.

The core idea is rather simple: at the start of each session, we compute upper and lower ATR bands anchored to the session opening price. When price breaks outside those bands while simultaneously establishing a new session high or low, we enter a position in the direction of the breakout, betting that such directional impulse will persist.

Conceptually, the framework attempts to isolate moments when a meaningful directional imbalance is affecting intraday price formation, and then position in anticipation of that imbalance persisting throughout the trading session.

Trade sizing depends on recent market volatility: the higher the recent volatility, the smaller the position size, and vice versa.

Keeping the entry scheme and sizing unchanged, we test four exit variations:

1. Session Open

2. Session Midline

3. VWAP (Volume Weighted Average Price)

4. PSAR (Parabolic Stop and Reverse)

The following chart provides a graphical illustration of how these exit mechanisms behave during a representative trading session.

Session Open

The most straightforward of the four. Its logic is directly connected to the signal itself: because the ATR bands are anchored to the session opening price, that level becomes the natural reference point for our imbalance measure.

In other words, if the market retraces all the way back to the session open, we consider the measured imbalance as invalidated: the buy or sell pressure that drove the initial move is no longer present, and we exit the position.

Session Midline

We define the session midline as the midpoint between the session’s cumulative maximum and minimum prices at any given point during the day.

If we enter a long position on a new session high or low, it appears reasonable to maintain exposure as long as price action remains closer to directional continuation than reversal.

VWAP (Volume Weighted Average Price)

The Volume Weighted Average Price (VWAP) is an indicator we have studied in depth in a prior research paper. We believe it serves as an elegant exit solution for intraday trend strategies because it defines a natural level of balance and agreement between buyers and sellers.

If the market retraces back to VWAP after a breakout, we interpret this as a sign that the directional imbalance has been absorbed and the earlier directional conviction is no longer supported.

PSAR (Parabolic Stop and Reverse)

The Parabolic SAR is a classic trailing stop mechanism originally presented by J. Welles Wilder in 1978. What sets it apart from the other three exits is its acceleration mechanic: the stop not only trails price, but tightens progressively as the trade extends in our favor, becoming stricter the more the trade moves in our favor.

The stop level updates each bar according to the following rule:

For the first trade of each session, PSAR is initialized at the session open. For subsequent trades within the same session, PSAR is not reset: instead, it picks up from the last stop level reached at the previous exit.

The Results

All four exit policies deliver positive results across the full backtest period on SOXX: Sharpe ratios range from 0.92 to 1.16, CAGRs from 16.2% to 18.3%, and maximum drawdowns are clustered in the -20% to -24% range. Every configuration remains profitable, although we observe some dispersion in performance.

The VWAP exit leads on both Sharpe (1.16) and CAGR (18.3%), with the Midline close behind (Sharpe 1.12, CAGR 17.8%). PSAR and Session Open trail somewhat, at Sharpe ratios of 1.02 and 0.92 respectively.

At this point one might be tempted to simply select the highest-Sharpe exit type and be done with it. Before drawing that conclusion, however, it is worth looking more closely at the trade statistics.

Once we move beyond portfolio-level metrics, the picture becomes considerably more nuanced.

The Session Open exit generates the fewest trades (2,406) and the highest average return per trade at 6.8 bps. On the surface that is appealing, but it comes with by far the worst average intra-trade drawdown at nearly -89 bps, meaning those returns require sitting through significantly deeper adverse moves.

The PSAR exit tends to offer less severe intra-trade drawdowns of (-73 bps) while maintaining a competitive average return of 6.3 bps. The Midline and VWAP exits compress drawdowns further (-65.5 and -64 bps respectively), at the cost of somewhat lower average trade returns (6.2 and 6.1 bps).

This is a pattern we observe consistently in our research: optimizing for one metric tends to come at the expense of another. Once evaluation criteria widen beyond a single metric, what “best” actually means becomes less clear, and every specification carries trade-offs.

The instability becomes even more apparent when looking at annual Sharpe ratio rankings, where we see that there is no consistent winner from one year to the next.

The same strategy that excels in one regime can look poor in the next: rankings rotate, and judging form their average values, no configuration has persistently dominated across the full testing sample.

What to Do Now?

From first principles, if we select one exit policy and discard the others, we are implicitly making a bet that this particular approach is meaningfully superior to the alternatives.

That would be justified if we had very strong confidence that one specific exit type is genuinely and durably better, but the evidence simply does not support that view: different metrics might favor different configurations, and rankings tend to rotate year by year.

Interestingly, the best course of action we have found in these situations is to avoid the choice entirely, or, in other words, choosing not to choose.

This is possible through a technique that is fairly well known in quantitative trading but, in our opinion, not appreciated or discussed nearly enough: ensembling.

Instead of being 100% exposed to a single specification, the more robust approach is to spread exposure across all specifications that, upon logical scrutiny, appear both intuitive and structurally sound. In our experiment, since we have no strong basis to favor one exit policy over the others, we should allocate a portion of the risk budget to each implementation. This eliminates the risk of selecting the worst configuration and removes the implicit assumption that we can reliably foresee which specification will perform best going forward.

Notably, ensembling does not require the individual strategies to be uncorrelated or dramatically different from one another. It simply reflects the fact that our uncertainty about which specification is best is real, and that uncertainty should be expressed in how we allocate rather than suppressed by an overconfident single-point choice.

Conclusion

We find ensembling to be both a robust and underrated approach, one that strongly resonates with the spirit of Occam’s Razor.

Occam’s Razor, in its most practical interpretation for system design, is about parsimony: the simpler a framework is, the fewer assumptions it rests on, and the fewer potential points of rupture it carries.

Every extra assumption we introduce is a place where something can break in the future, either because that assumption eventually stops holding or because it creates a temptation to tinker when performance disappoints. Fewer moving parts means fewer decisions to revisit and fewer levers to pull under pressure.

Ensembling across logically sound specifications is, in this sense, a form of simplicity. Instead of introducing the strong assumption that configuration X is best, we remove that assumption entirely and replace a fragile single-point choice with a robust distributed one. The result is a framework with fewer decisions to second-guess over time, and a design that is inherently more stable across changing market regimes.

Remember… markets rarely reward unnecessary complexity.

If you found this article useful, feel free to leave a comment and reach out via direct message or email at info@concretumgroup.com for any questions.

Last week we published a piece describing the reverse engineering pipeline we used to try and decode his trading rules.

Since then, we have received a number of comments and private messages from interested readers eager to explore more details, as well as several specific requests. Some of these stood out as particularly compelling, therefore we decided to address them in this brief follow-up.

If you have not read our original article, we recommend doing so before continuing. This piece contains several technical references that build directly on the methodology and findings introduced there.

Optimal Lookback Window

As readers may recall, our replication framework was designed to generate thousands of synthetic CTAs, each regressed individually against Mulvaney's monthly returns through OLS, with Mulvaney's returns as the dependent variable and the replica returns as the independent variable. The resulting R² values were then used to rank the synthetic CTAs and identify the trading parameters that best fit his historical returns.

One request we received was to test additional Donchian channel lookback windows. In our original analysis the best-performing parameter was 126 days, which happened to coincide with the lower bound of our tested range, suggesting that shorter lookback periods might produce even stronger fits.

To test this hypothesis, we ran a simulation based on our best-fit replica specification.

Keeping all other parameters fixed (see the table above), we varied only the Donchian channel lookback from 30 to 360 days in steps of 10, producing 34 combinations in total. The chart below illustrates how R² evolves across the different lookback lengths.

The best goodness of fit is clearly achieved for lookback periods between 120 and 140 days, with a peak at 130 days. Notably, R² drops sharply for shorter lookback periods, reinforcing the intuition that Mulvaney does not engage in short-term trend following (as he has consistently stated over the years).

Trade Level Statistics

As some readers have pointed out, Mulvaney has offered glimpses into certain trade-level characteristics of his fund across various interviews and commentaries, including references to win-rates and typical holding periods for his trades:

With this in mind, we extracted trade-level statistics for our best-fit replica (i.e., CTA 1), with one important caveat. Because the framework permits multiple tranches to coexist within the same futures contract, each tranche is treated as a distinct trade for the purpose of all trade-level analyses. In other words, the initial entry and all subsequent pyramiding additions are recorded as separate trades rather than consolidated into a single position.

The overall profile is consistent with what one would expect from a canonical trend-following program: a rather low win-rate (~35%) and a strongly right-skewed PnL distribution (5.98). The negative median PnL reinforces this picture, indicating that most trades are small losers while long-run performance is driven by a relatively limited number of outsized winning trades.

Compared with Mulvaney’s declared statistics, our replica shows a modestly higher win rate, at roughly 35% versus his reported 25%, and a considerably shorter average holding period of ~80 trading days (equivalent to about four months).

Part of this discrepancy could be methodological. We count each pyramid tranche as a separate trade, so a position that is entered and later scaled up appears as multiple trades in our statistics, with the additional tranches naturally showing shorter measured trade durations. If Mulvaney reports holding period from first entry through final exit, the two figures are not directly comparable, and our replica would likely move closer to the reported six-month figure on that basis.

That said, we believe the remaining differences in both holding period and win rate most likely reflect the stop-loss mechanism, as noted in our original analysis. Our simplification, which uses the Donchian channel midline as the stop reference, likely does not capture the probabilistic-stop behavior that Mulvaney has hinted at in a few occasions.

At the asset-class level, trade statistics appear surprisingly stable in terms of hit ratio and average duration, with all sectors clustering around ~35% win rates and ~80 trading days per trade. The main differences instead emerge in the tails of the distribution: rates and commodities display materially higher skewness together with significantly larger best-trade outcomes relative to equities and FX.

Discrete and Continuous Trading

All in all, the way Mulvaney references individual trades and holding periods reinforces the idea of a program that is better approximated by “old-school” discrete signals than by the more modern continuous-forecast approach. In such a model, a position is either on or off, and the system transitions between distinct “states” rather than continuously adjusting exposure as new information arrives.

This distinction has long been debated in the trend-following space, with thoughtful practitioners on both sides.

Continuous forecasts are arguably a better proxy for how new information progressively changes our “view of the world” over time, rather than forcing that evolving view into abrupt binary state changes. A discrete signal, by contrast, models the world in jumps, moving from no position to full position in a single step, which can feel like a crude approximation of a more nuanced underlying process.

That said, Mulvaney is a compelling counterargument.

His program, likely built on discrete trading rules, appears radically simple by design, and we find it fascinating that seemingly unsophisticated rules can survive for decades.

Such track record is a useful reminder that simplicity and parsimony are often underappreciated virtues in quantitative trading frameworks.

A Gift For Our Readers

As a thank-you to our readers for their continued support, we’ve put together an interactive online tool that allows you to build and backtest your own synthetic replica of Mulvaney performance.

Explore the Tool

You can experiment with the same variables we used in our replication study, including different lookback periods, fixed stop losses, as well as adjustable management and performance fee structures. Whether you want to adhere to the parameters we found to best fit Mulvaney, or explore entirely different combinations, the tool gives you full flexibility to test your assumptions.

We’ve made it possible to download a plot of your synthetic CTA, along with a direct link to the relative parameter set. Feel free to explore the tool and share on social media any replica that, in your view, comes closest to Mulvaney’s phenomenal track record.

We would love to hear your perspective, particularly from those who have thought carefully about this trade-off in practice. Feel free to leave a comment and reach out via direct message or email at info@concretumgroup.com for any questions.

“You could publish trading rules in the newspaper and no one would follow them. The key is consistency and discipline.”

Richard Dennis

Not many portfolio managers can claim a track record compounding at roughly 20% per year over more than 26 years of trading.

Paul Mulvaney is one of them.

He is the CEO and CIO of Mulvaney Capital Management, the London-based CTA he founded in March 1999. According to the firm’s own materials, the program has run unchanged since inception, entirely systematic, long-term, and built around a single idea: markets adjust gradually to changes in fundamentals, and those gradual adjustments are the driving force behind trends.

Today, Mulvaney is regarded as one of the most respected names in the CTA industry, with what may be the most impressive long-term record in the field. For context, a $100,000 investment made at inception would be worth almost $10 million today, net of fees. After +89% in 2022, +51% in 2023 and +83% in 2024, the momentum has continued into 2026: as of the end of March, performance is already up approximately +65% year-to-date.

This track record has clearly propelled Mulvaney from a niche name admired by CTA connoisseurs to a household name among trend-following practitioners.

Michael Covel, author of Trend Following, has repeatedly pointed to Mulvaney as an example of how classical trend-following principles (applied with conviction and consistency) can produce outstanding long-term results. On several occasions, he has suggested that reverse-engineering the rules of successful trend managers is one of the most instructive exercises a systematic trader can undertake.

Today, we take on that challenge.

By leveraging a replication framework built around thousands of synthetic CTAs, we attempt to reverse-engineer Mulvaney’s trading rules: if you are curious to see what this exercise tells us about his style, we think this piece will be well worth your time.

What We Know

The starting point of any reverse-engineering exercise is to first establish a set of priors: assumptions that act as fixed points for all subsequent adjustments and calibrations.

In Mulvaney’s case, we are in a relatively comfortable position for two reasons.

First, a substantial body of material is publicly available, ranging from interviews to the firm's institutional presentations and a long series of monthly commentaries dating back to 2000. We have spent the past few weeks going through this archive, and the picture that emerges is remarkably coherent.

Second, Mulvaney’s trading style is clearly identifiable and, judging by what the firm has disclosed, has remained consistent throughout the entire life of the fund. In Paul’s own words, “the system has not changed since inception”.

Against that backdrop, five themes clearly stand out.

1) A Broad Futures Universe

The Mulvaney Capital Global Diversified Program trades a universe of 45 futures markets, spanning equities, interest rates, currencies, energy, metals, softs, and agricultural and livestock contracts.

Critically, this universe extends well beyond the most liquid mainstream contracts. Mulvaney is perhaps most famously associated with the cocoa trade of 2023–2024, one of the most remarkable trend trades of the past several decades, and a position that would have been effectively invisible to managers who restrict themselves to deep, liquid markets.

It appears clear that the willingness to trade less crowded instruments is a defining characteristic of his approach: for our synthetic CTA replicas we construct a universe of roughly 40 futures markets covering the same sectors.

2) Donchian Channel Breakouts

Mulvaney's buy-sell engine is likely a canonical Donchian channel breakout: new N-day highs generate long entries, new N-day lows generate short entries. The firm's own illustrative chart, recently circulated on social media, depicts exactly this type of behavior.

Both the firm’s website and multiple interviews emphasize the long-term nature of the program, with positions held for roughly six months on average, therefore we do not expect the Donchian lookback to be short-dated. Medium to long-term is the only range consistent with Mulvaney’s stated time frame and his repeated preference for filtering out short-term noise.

3) Gradual Capital Commitment

Mulvaney does not seem to deploy his full position at the initial breakout. Instead, exposure is built up in stages as the trend develops, with additional contracts added at what seem to be predefined profit thresholds.

Only once the trend has extended sufficiently does the position reach its maximum size. In the firm’s institutional material, the illustrative winning trade shows exposure scaling incrementally along the trend, while the losing-trade example is stopped out near the initial, pivot-sized entry.

4) A Volatility-Based Trailing Stop

Exits appear to be dictated by a trailing stop that moves only in the direction of the trade: for long positions, the stop can move up but never down; for shorts, the opposite.

Hitting the trailing stop is, crucially, the only reason to exit. There are no profit targets, and Mulvaney has been unambiguous on this point for more than two decades.

The mechanics of the stop itself seem quite sophisticated. In a 2006 interview for Hedgeweek, Paul described the approach as follows:

In the interest of parsimony, we have chosen to use the midline of the Donchian channel as our trailing exit.

We find that this choice has a desirable empirical property: as the trend develops and the channel range expands, the midline moves progressively further from the channel extremes, accommodating rising volatility without any additional scaling parameter (it may grow more distant from one boundary, but it will not reverse its direction).

What we layer on top is a fixed initial stop, set at the inception of the trade, which may sit closer to the entry price than the midline. This is intended to capture the possibility that Mulvaney opens positions with a tighter initial risk threshold before transitioning to the standard Donchian midline as the trailing exit once the trade has room to develop.

5) Risk-Based Position Sizing

Mulvaney suggests he commits only "a small percentage of equity" per trade, which is consistent with a canonical equal-loss sizing methodology: define a maximum acceptable loss at the portfolio level and size each trade such that, if every position hits its stop simultaneously, the total portfolio loss does not exceed that pre-determined threshold.

What We Don’t Know

Despite the volume of publicly available material, a handful of key parameters have never been explicitly disclosed. These are precisely the degrees of freedom our reverse-engineering exercise needs to pin down:

• Per-trade risk allocation - The open question is whether that risk budget is distributed equally across all markets (what we call Loss Parity, or 1/N risk allocation), or first allocated across sectors and then equally across markets within each sector (what we call Hierarchical Loss Parity, HLP).

• Donchian lookback period - Somewhere in the medium-to-long-term range, but exactly how long?

• Fixed stop level - Since position sizing equalizes initial risk, initial stop distance does not affect loss per trade but instead governs stop-out frequency. Tighter stops increase early exits, while looser stops reduce premature termination. At the same time, tighter stops mechanically allow for higher position sizing, and thus greater upside participation when trends persist.

• Pyramiding specification - How many layers is the program designed to add? How aggressively, and at what threshold of open-trade profit does each new layer trigger?

• Directional bias - Are long and short positions treated symmetrically, or is one side deliberately scaled down?

• Execution lag - When a new signal fires, is the trade executed on the close of the same day, the next day's close, or with a further delay?

Reverse-Engineering Mulvaney’s Returns

The starting point is what we call “grid” construction. For each feature of the trend program, we define a range of plausible specifications: every combination across features generates a distinct synthetic CTA specification.

In total, our grid yields 4,320 fully specified trend programs: the parameters swept, along with their tested values, are summarized in the table below.

The rules of the trading program itself are quite straightforward. At its core, the system is a breakout strategy with loss-based position sizing and a pyramiding overlay. The full logic is reducible to a handful of conditions:

ENTRY SIGNAL
  Long   if Close >= Donchian High
  Short  if Close <= Donchian Low

----------------------------------------------------------------------

INITIAL FIXED STOP
  Channel Width = Donchian High - Donchian Low
  Long   = Donchian High - p x Channel Width
  Short  = Donchian Low  + p x Channel Width

----------------------------------------------------------------------

TRAILING STOP
  Donchian Midline = (Donchian High + Donchian Low) / 2

----------------------------------------------------------------------

POSITION SIZING
  Equal-Market Loss Parity:
      Contracts = (15% x Equity / Number of Markets) / DollarStopRisk

  Hierarchical Loss Parity:
      Contracts = (15% x Equity / Number of Sectors / Markets in Sector) / DollarStopRisk

----------------------------------------------------------------------

PYRAMIDING MULTIPLIER
  r = cumulative open-trade profit / initial stop risk
  m = min(cap, 1 + floor(max(0, r) / K))

----------------------------------------------------------------------

POSITION MANAGEMENT
  Contracts stay untouched unless signal changes state or pyramiding event occurs
  Pyramiding events can only increase exposure

----------------------------------------------------------------------

CASH 
  Futures P&L is combined with a collateral cash return equal to
  Fama-French risk-free rate - 100 bps

The resulting AUM path is sampled at month-end. Before any comparison is made, each replica is volatility-scaled to match Mulvaney’s realized volatility over the overlapping sample: this step matters because fees are applied after scaling, meaning the 2% management fee and 20% high-water-mark (HWM) performance fee are netted against a return series running at Mulvaney’s volatility level, ensuring a like-for-like comparison.

Finally, each synthetic CTA time series is regressed against Mulvaney's monthly returns via ordinary least squares:

Every specification is then ranked by the R² of that regression: a higher R² suggests the replica tracks Mulvaney more closely, and its parameter combination is therefore assumed to be nearer to his actual rules.

Note - Because the replica is constructed at daily frequency but identification relies solely on monthly returns, the problem is inherently under-constrained (multiple structurally distinct trading processes can produce statistically indistinguishable monthly series). R² should therefore be read as a coarse filter for identifying plausible candidate models, not as a definitive selection criterion. The output of this exercise is best understood as a region of plausible parameter combinations rather than a single best-fit answer.

What the Data Tells Us

When we narrow the 4,320 replicas down to the top 10 by highest R², the results are remarkably consistent. Put in different words, the parameters that best capture Mulvaney’s returns seem to cluster around coherent values.

The best-fit synthetic CTAs achieve R² values in the range of 0.71 to 0.73, representing a ~45% improvement in explanatory power over the SG Trend Index (a widely used benchmark among trend-following managers), with correlations sitting tightly between 0.84 and 0.85.

At its core, his breakout approach seems to be calibrated on rolling windows of 6 months: waiting for six-month highs to initiate long positions and six-months low to initiate short ones.

The data suggests that Mulvaney does not underweight short positions relative to long ones, and we estimate that execution is likely carried out over a 24-to-48-hour window following a buy/sell signal.

Risk appears to be allocated using a straightforward loss parity configuration, distributed equally across the futures in the traded universe, rather than first splitting the risk budget within sectors. This choice implies that risk allocation is engineered ex-ante at the universe-construction stage, since the number of futures per sector effectively determines each sector's weight in the portfolio.

The ten best-fit replicas suggest that when a position is first opened, a fixed stop is placed at the Donchian high or low, offset inward by roughly one-third of the channel's range. As the position moves into profit, this initial stop is replaced by the midline trailing stop, which then takes over for the remainder of the trade.

Notably, even when Mulvaney is compared against the 10 best-fit replicas before applying the 2% and 20% fee structure, his average annualized alpha relative to them remains slightly above zero in most cases.

There is a final caveat to consider, and it is the most humbling.

Since 2022, Mulvaney has delivered a run of results that our synthetic CTAs do not fully reproduce. Our model captures the bulk of the monthly variance, but the magnitude of the recent run-up is only partially replicated.

This gap becomes particularly visible around 2023–2024, coinciding with the now-famous cocoa trade, where Mulvaney built and held a concentrated long position through one of the most extreme commodity moves in recent memory, something our replicas were unable to capture in quite the same way.

Part of this gap likely reflects a universe effect, as our basket does not track Mulvaney’s exactly. The larger part, however, is almost certainly due to refinements in pyramiding and position sizing that our synthetic replicas do not capture.

Chapeau.

A Word of Caution

Reverse-engineering Mulvaney’s rules is one thing. Trading them is another.

His track record can be reasonably approximated by a simple set of systematic rules, but simple does not mean easy.

Behind his staggering returns lies a turbulent equity curve that would test the conviction of even the most seasoned investors, demanding a rare tolerance for drawdowns.

An analysis of his underwater curve (the chart that tracks how far the strategy sits below its all-time high at any point in time) reveals that he has historically spent approximately 82% of months in a drawdown.

And within that condition, the average underwater value is around -15%, a persistent state of discomfort during which every instinct pushes toward changing something: adding a filter, smoothing the equity curve, or layering on a complementary strategy to make the ride more tolerable.

Mulvaney has faced that pressure for a quarter of a century and never acted on it. Such discipline is not a footnote to his track record: we believe it is a core ingredient of it, and one that cannot be extracted from a ruleset.

Then there is the practical reality.

Trading this system across a universe of 40 or more futures markets with a sufficient capital buffer demands several million dollars of starting equity. At that scale, and given Mulvaney’s volatility profile, a single bad month can easily translate into seven-figure losses.

So yes, we can reverse-engineer the rules. We can build the system, run the backtest, and show you that the logic holds…

What we cannot give you is the stomach to trade it, enough capital to size it properly, or the 25-year track record of sticking with it when nothing seems to work.

A Gift For Our Readers

As a thank-you to our readers for their continued support, we’ve put together an interactive online tool that allows you to build and backtest your own synthetic replica of Mulvaney performance.

Explore the Tool

You can experiment with the same variables we used in our replication study, including different lookback periods, fixed stop losses, as well as adjustable management and performance fee structures. Whether you want to adhere to the parameters we found to best fit Mulvaney, or explore entirely different combinations, the tool gives you full flexibility to test your assumptions.

We’ve made it possible to download a plot of your synthetic CTA, along with a direct link to the relative parameter set. Feel free to explore the tool and share on social media any replica that, in your view, comes closest to Mulvaney’s phenomenal track record.

Conclusion

Mulvaney's approach is, in essence, very coherent with what he has publicly disclosed: a long-term trend-following program built on Donchian breakouts, with exposure to less liquid markets and disciplined pyramiding into winners.

Nothing that any moderately capable systematic trader could not write in a few hundred lines of code.

And yet, as Richard Dennis reminds us… “you could publish trading rules in the newspaper and no one would follow them.”

Mulvaney is living proof that returns don't stem from signals alone. What has made his returns achievable in practice is the willingness to run the same system without tinkering for a quarter of a century, through drawdowns that would have dispatched most of his peers.

If anything, this exercise reinforces a view we have expressed before: in systematic investing, the interesting question is rarely “what are the rules?”, but rather “how do you build a business and a psychology that can execute those rules, exactly as written, for 25 years?”.

A final note on methodology. The replica method presented here is, by design, a first-pass simplification: a deliberately parsimonious model intended to characterize the broad nature of Mulvaney’s strategy. The most important shortcut concerns the trailing-stop mechanism. As discussed earlier, Mulvaney’s public material describes a probabilistically calibrated stop, while our replica uses the Donchian channel midpoint as a proxy. We chose this simplification deliberately, to surface the dominant features of his style without getting lost in the parameterization of the exit.

We would be interested to hear your views, particularly from those who have done their own reverse-engineering. Do our priors hold up against yours? And if you know something we do not, we would genuinely welcome it.

If you found this article useful, feel free to share it with other practitioners, leave a comment with your perspective, or reach out at info@concretumgroup.com

“The first rule of compounding is to never interrupt compounding unnecessarily.”

In theory, this sounds trivial. In practice, compounding is rarely interrupted by a lack of return potential. It is interrupted by investors reacting to volatility. Sharp drawdowns, violent daily swings, headline-driven regimes: these are the moments in which long-term investors quietly turn into short-term risk managers, de-risking at the worst possible time and abandoning plans they had every intention of following.

The lived experience of the market, the day-to-day path of outcomes, shapes behavior. And once behavior becomes pro-cyclical, compounding becomes fragile.

This is precisely the problem volatility targeting is designed to address.

How this research began

In the summer of 2025, we were approached by the structuring team of a large US financial group with a very specific question.

Could you build a volatility forecasting methodology capable of estimating next-day volatility with enough accuracy and stability to serve as the engine of an equity volatility-target index?

The motivation behind the question was not academic.

The underlying goal was to offer clients a better long-term investing experience: a smoother ride through turbulent markets, the kind of risk profile that helps investors stay committed to their plan rather than capitulate at the worst possible moment.

As is typical in the US market, they wanted to package the strategy into a capital-protected note where the end investor gains exposure to equity upside through a call option, typically with a one-year maturity. And in that context, the accuracy and stability of the volatility delivery directly determine the fair price of the option, which in turn determines the participation rate offered to the end investor.

A more stable underlying means cheaper optionality, which means better terms for the end client.

That requirement, not merely forecasting volatility well but delivering a stable risk profile through a wide range of market regimes, is what drove the design of what would become the NextVoL^® framework, developed jointly by Concretum|Research and Professors Walter Distaso and Antonio Mele.

The rest of this piece explains why standard forecasting methods struggle with that objective, and how NextVoL^® addresses the structural limitations we encountered along the way.

Subscribe now

The idea behind volatility targeting

The mechanics are simple. Rather than trying to forecast returns (famously difficult), a volatility-target strategy dynamically scales exposure to keep portfolio risk anchored around a predefined level .

The allocation rule is:

where the numerator is the volatility target, while the denominator represents the estimate of next-day volatility.

When expected volatility is below target, exposure is increased. When it is above target, exposure is reduced. If the volatility target is 10% and your forecast is 20%, you are 50% invested. That is it.

The appeal of this approach rests on two well-documented stylized facts.

First, volatility is persistent: high-volatility days tend to be followed by high-volatility days, and calm clusters with calm.

Second, and more importantly for equity investors, risk-adjusted returns tend to deteriorate as volatility increases.

In SPY, the highest-volatility quartile historically delivers a negative return-per-unit-of-risk.

Scaling exposure inversely with volatility is therefore not just a risk-stabilization overlay. It can improve long-run compounding efficiency.

But none of this works if the volatility forecast is poor. The scaling rule is mechanical. Its effectiveness rests entirely on the quality of the volatility estimate.

Why the forecasting method matters

Over the past four decades, a rich landscape of volatility forecasting approaches has emerged:

• Rolling historical volatility. Transparent but slow and mechanically jumpy.

• GARCH-type models. Elegant econometrically, but sensitive to misspecification.

• RiskMetrics EWMA. Simple, but backward-looking with a fixed decay.

• HAR models. Multi-scale and parsimonious, but can lag at regime transitions.

• VIX / implied volatility. Forward-looking, but structurally biased upward by the volatility risk premium.

• High-frequency realized volatility. Responsive, but exposed to microstructure noise and unstable overnight adjustments.

Each of these approaches has strengths, but each also has a structural limitation that becomes visible when embedded in a volatility-target index.

Some models react too slowly and fail to stabilize risk during sharp regime shifts.

Others react too aggressively, generating excessive turnover and leverage instability.

Many rely on a constant scaling factor to approximate the overnight component of daily volatility, an assumption that does not hold when the relative contribution of overnight and intraday variance shifts across regimes.

The NextVoL^® framework

NextVoL^® is our attempt to address these structural limitations jointly, rather than one at a time.

The framework starts from an intraday-based volatility estimate, which gives it the responsiveness of high-frequency data. On top of this base, it incorporates three mechanisms:

• Short-horizon smoothing, which attenuates the impact of transient volatility spikes and prevents the forecast from generating noisy day-to-day leverage swings.

• Dynamic overnight adjustment, which abandons the constant scaling multiplier and instead accounts for shifts in the relative weight of intraday versus overnight variance across regimes, reducing the systematic bias embedded in static corrections.

• A stabilization mechanism, which allows the forecast to adapt when realized risk persistently deviates from the objective, reinforcing consistent target delivery without sacrificing responsiveness.

The design philosophy is straightforward: keep the informational advantage of high-frequency data, but structurally discipline the well-known failure modes that come with it.

Empirical results

We tested all seven methodologies (the six standard approaches plus NextVoL^®) on a 10% volatility-target index constructed on SPY, daily rebalanced, from January 2008 to December 2025.

No risk-free remuneration on the uninvested portion, no leverage cap. We deliberately stripped out frictions in order to isolate the pure effect of the forecasting choice.

The full-sample numbers are reported below:

A few observations stand out:

• NextVoL^® delivers the most accurate target, with average rolling 252-day volatility of 10.11%, essentially on target, and the narrowest distribution of all methods (1st percentile at 9.58%, 99th percentile at 10.75%).

• Volatility-of-volatility is nearly half that of the next best method (0.26 vs 0.48 for plain RV, 0.58 for HAR and RiskMetrics, 0.91 for the VIX-based approach).

• The VIX-based strategy systematically undershoots the budget (average realized vol of 7.82% against a 10% target), an under-utilization of roughly 22% driven by the embedded volatility risk premium.

• NextVoL^® also produces the highest Sharpe ratio (0.79) among the volatility-target approaches.

Hitting the target is not enough

Averaging 10% realized volatility over the full sample is a necessary condition, but it is not a sufficient one. A method can hit the target on average while fluctuating wildly around it year after year, and a method that stays tightly anchored to the target through time is a fundamentally different product from one that averages 10% by alternating phases of over- and under-delivery.

The full distribution of rolling realized volatility for each method makes this point visually:

Three things jump out of this chart:

• NextVoL® is a tall, narrow spike centered almost exactly on the 10% target. It is the only distribution that does not meaningfully cross into either over- or under-delivery regions.

• The VIX-based approach sits far to the left, with its mode around 7.8%, a structural under-utilization of the volatility budget driven by the implied-vol risk premium. Over an 18-year horizon, that is about 22% of the intended risk left on the table.

• Historical Volatility sits far to the right, with a wide distribution centered above 11.5%, systematically overshooting the target, and doing so inconsistently.

• The other methods (RV, RV HAR, GARCH, RiskMetrics) cluster reasonably well around the target but with distributions 2–3× wider than NextVoL®.

To rank the methods rigorously, we need a single summary statistic that penalizes both dimensions of miss at once: how far a method sits from the target on average, and how much it fluctuates around it through time. The natural candidate is the mean squared error of the rolling 252-day realized volatility relative to the 10% target:

• The squared deviation penalizes bias. A method that averages 8% or 11.5% is structurally off-target, and that offset gets weighted in.

• It also penalizes dispersion. A method that averages exactly 10% but swings between 7% and 13% is not truly delivering the target either.

Ranked by this metric, NextVoL® sits comfortably in first place across the full sample. But a full-sample ranking is only half the story.

Persistent, not sample-specific

The aggregate MSE ranking could, in principle, be driven by a handful of favorable years.

To check for this, we decomposed the ranking year by year:

NextVoL^® ranks first or second in sixteen out of eighteen calendar years, with an average annual rank of 1.3.

The methodologies that look close in the aggregate table (RV at 2.9 and GARCH at 3.3) turn out to be considerably more variable on a year-by-year basis. Both occasionally drop to rank 4, 5, or 6. This suggests that their aggregate performance partly reflects offsetting periods of over- and under-delivery, rather than consistent tracking. The NextVoL^® advantage, by contrast, is structural rather than sample-specific, and that is precisely what you want from a methodology embedded in a live product.

Why this matters beyond the Sharpe ratio

One could read the results above and focus on the Sharpe improvement. But the Sharpe story, while real, is the less interesting half of the picture.

The more important advantage of NextVoL^® is the stability of volatility delivery itself, captured by the vol-of-vol of 0.26 and the remarkably tight distribution of rolling realized vol around the 10% target. This is not a cosmetic property.

For anyone building structured products, volatility-controlled indices, or insurance-linked solutions such as Fixed Index Annuities (FIAs), the stability of the underlying’s realized volatility is directly priced into the derivative overlay.

A lower vol-of-vol means lower uncertainty on future variance, which translates into tighter option pricing on the wrapper, cheaper hedges for the issuer, and, crucially for the end investor, higher participation rates on the embedded optionality.

A volatility-target index that delivers its objective with a vol-of-vol of 0.26 is a materially better foundation on which to build a structured product than one that delivers with 0.70 or 0.90. On paper, their average volatilities may look similar. In practice, the gap translates into sharper option pricing, higher participation rates, and a better deal for the investor holding the note

If you are interested in understanding this study in greater detail, including the precise construction of the volatility-target indexes, the full empirical methodology, and the extended results across all forecasting approaches, we invite you to read the white paper we have just released.

Volatility Targeting for Long-Term Compounding, by Walter Distaso, Antonio Mele, and Carlo Zarattini.

Read the Paper

If you are working on volatility-controlled index design, structured product pricing, or volatility forecasting more broadly, we would love to hear your perspective. You can reach out at info@concretumgroup.com.

As always, feel free to leave a comment and reach out via direct message or email at info@concretumgroup.com for any questions.

Every signal family within the trend-factor umbrella is, at its core, a different way of determining whether an asset is going up or down, and then betting on the persistence of that behavior.

No amount of mathematical sophistication or complex modeling changes the fact that the time-series momentum effect is the main PnL driver for any trend-following portfolio manager. One might therefore expect that different versions of fundamentally similar signals, all harvesting the same market effect, would lead trend managers to achieve rather similar, homogeneous results.

The data tells a very different story.

In practice, the signal component is really just step zero. What ultimately determines a trend-following program’s outcome isn’t just the signal itself, it’s the dozens of decisions layered on top of it:

• Universe composition

• Position sizing

• Risk budgeting

• Long-short versus long-only orientation

• Rebalancing frequency, and more…

Each layer compounds on the last, and over time, these choices cause managers to drift substantially from one another.

This phenomenon is commonly referred to by practitioners as trend-follower dispersion. It is what causes two highly successful trend managers, such as Mulvaney, which runs at a historical volatility of approximately 40%, and Quantica (which operates at a fraction of that volatility) to be placed in the same broad category despite delivering markedly different return profiles.

Notably, Quantica has maintained its relatively low volatility with great consistency over the past 20 years (as shown in the picture below), while Mulvaney's realized volatility has exhibited considerably wider swings around its long-run average.

In our experience building trend-following programs, both for our own portfolios and for external clients, position sizing is one of the most impactful decisions a manager can make.

Whether driven by mandate restrictions, VaR requirements, or the deliberate construction of a high-volatility trend sleeve, the same signal applied to the same universe of futures can give rise to very different outcomes.

Drawing inspiration from a piece we published in 2025 (link), we revisit the question:

What is the best approach to position sizing?

As we will show with a simple case study, the answer is more nuanced than one might expect.

The Strategy

Buy/Sell Engine

Although our focus today isn't on the buy/sell engine itself, this is a good opportunity to introduce a different entry/exit signal model that first emerged in the 1980s.

The Aberration strategy is a trend-following framework developed by Keith Fitschen in 1986 (Fitschen, 2013)1, built on what are effectively Bollinger Bands, first formalized by John Bollinger around the same period (Bollinger, 2001)2.

In our specification, the bands are computed from a 50-day moving average and ±2 standard deviations of prices.

Entry signals are generated when the price moves outside the bands: a breakout above the upper band triggers a long position, while a breakout below the lower band triggers a short position. The position is held until the price crosses the moving average in the direction opposite to the prevailing trend, which serves as the exit signal.

Our sole modification to the standard rules is to impose a monotonic adjustment of the moving-average-based stop. For long positions, the stop is permitted to move only upward; for short positions, only downward: the stop never shifts in a direction adverse to the trade, effectively acting as a trailing exit constraint.

In practice, the strategy can be summarized as follows:

• Band construction

  • UB = SMA(50days) + 2 st.dev.

  • LB = SMA(50days) - 2 st.dev.

• Entry logic

  • Long → price breaks above the UB

  • Short → price breaks below the LB

• Exit logic

  • Long → price crosses below SMA(50days)

  • Short → price crosses above SMA(50days)

• Stop dynamics

  • Never move the stop against the direction of the trend

Sizing Approach

Given the objective of this study, we assess the strategy under three different sizing policies:

• Volatility Targeting (VT). This approach dynamically scales position sizes to maintain a constant level of daily volatility, ensuring that risk contribution remains stable and homogeneous across portfolio holdings. It is grounded in the empirical observation that volatility, while noisy, exhibits a degree of predictability. Its main limitation, however, is structural: VT tends to scale down exposure precisely when trends accelerate, as strong directional moves are often accompanied by rising volatility. As a result, it can systematically reduce participation in the most profitable phases of a trend (this effect is less pronounced in equity markets, where volatility and returns are typically negatively correlated).

• Volatility Parity (VP) is essentially a set-and-forget approach; position sizes are determined at entry based on prevailing market volatility, and then left unchanged throughout the life of the trade. In other words, risk is calibrated ex-ante, but not rebalanced ex-post. The intuition is simple. By fixing position sizes, the strategy allows profits and losses to evolve naturally, without mechanically interfering with the trade. This has two important consequences. First, it minimizes turnover and implementation costs. Second, it allows large winners to compound and gradually dominate the portfolio, as their notional weight increases over time. The main drawback, however, is that risk is no longer controlled after entry. If volatility expands materially during the life of a position, the strategy can end up carrying significantly more risk than originally intended.

  Note - It is worth mentioning that this approach does carry a minor methodological inconsistency: successful long positions take a large portfolio portion if left untouched, while successful short positions actually shrink in size. This asymmetry is rarely acknowledged and should merit some further investigation and clarification.

• Volatility Parity with Pyramiding (VPP). This approach starts like Volatility Parity at inception, but allows positions to be scaled up as the trade moves into profit. Additional units are added at predefined profit thresholds, effectively pyramiding into winners. VPP introduces convexity to the P&L path: exposure increases when trends persist, amplifying large winners, while added risk is given back if the trade reverses. This makes it particularly suited to capturing the right tail typical of trend-following strategies. In our implementation, exposure is capped at 4× the initial position size, ensuring that risk remains bounded.

Results

We backtest the three portfolio constructions from 1978 to 2026 across a universe of nearly 40 futures markets spanning Agriculture & Livestock, Energy, and Metals, using data from Norgate Data.

Consistent with the objective of this study, results are reported gross of fees, and fractional contracts are allowed. These simplifying assumptions are intentional: they abstract from capital and execution frictions, ensuring that differences in performance are driven by the allocation methodology itself rather than by starting AUM or trading constraints.

Before turning to the full portfolio results, it is useful to build intuition by examining how the different sizing methodologies would have behaved during a well-known trend.

The chart below illustrates the Cocoa trade of 2023–2024, a move that most trend followers will remember.

The lower panel highlights the key differences in position dynamics. Under VP, the number of contracts remains constant throughout the trade. Under VT, position size gradually declines as volatility expands, reducing exposure in order to respect its constant risk target. VPP takes the opposite approach: exposure increases as the trend develops, scaling from roughly 20 contracts to more than 80, thereby amplifying returns and capturing a larger share of this outlier move.

The equity curves presented below show that volatility targeting (VT) and volatility parity (VP) trace relatively smooth upward paths, while Volatility Parity with Pyramiding’s (VPP) equity line is punctuated by sharp spikes wherever pyramiding increased exposure in a handful of positions. In some cases, such levered bets paid off substantially, while others reversed quickly, giving back profits and driving up the portfolio's realized volatility.

The visual intuition is confirmed by the performance metrics reported in the table below.

Volatility Targeting delivers the lowest realized volatility (10.8%), followed by Volatility Parity (13.1%), while VPP reaches a substantially higher level (25.4%). This is particularly notable given that all approaches start from the same position sizing at trade inception.

On the return side, the ranking reverses. VPP achieves the highest CAGR (16%), followed by VP (11.5%) and Volatility Targeting (10.3%).

If one were to select a “best” approach using conventional criteria, the natural choice would be the configuration with the highest risk-adjusted return.

From this perspective, Volatility Targeting clearly stands out, with a Sharpe of 0.96 and a Sortino of 1.37. Volatility Parity follows, with a Sharpe of 0.89 and a Sortino of 1.16, while VPP, despite its higher raw returns, ranks last on a risk-adjusted basis (Sharpe 0.71, Sortino 0.74).

So, does this imply that VPP is structurally inferior, and that Volatility Targeting is the dominant approach?

Not necessarily.

A Different Lens (Trade-Level Analysis)

The conclusion above relies on path-dependent, daily return metrics, where intra-trade fluctuations directly impact the NAV trajectory. However, if we shift the perspective and analyze performance at the trade level, the picture changes meaningfully.

Specifically, we reconstruct the NAV by accounting for each trade only at its exit, effectively removing intra-trade P&L volatility from the path. In this representation, returns are realized in discrete steps, rather than continuously marked-to-market.

As shown in the chart below, VPP continues, as expected, to deliver the highest cumulative returns. However, once intra-trade noise is stripped out, its volatility (particularly on the downside) appears much closer to that of the other two approaches.

To quantify reward efficiency more precisely, we compute, for each sizing specification, the Profit Factor, defined as the ratio between total profits and total losses across all trades.

This metric measures how much gross profit is generated per unit of loss incurred. Intuitively, it treats losses as the “cost of doing business” and evaluates how efficiently each strategy converts that cost into gains. In other words, it captures how many dollars of profit are produced for every dollar lost.

Notably, under this metric, the ranking across the three sizing policies reverses.

VPP emerges as the most efficient configuration, with a Profit Factor of 1.74, followed by VP at 1.55 and VT at 1.48.

This result is consistent with the trade-level statistics reported above: although VPP exhibits a lower hit ratio and larger intra-trade drawdowns, it generates substantially larger winners (107 bps vs ~74–78 bps), while its average loss remains broadly in line with the other approaches.

We refer to this as the observer effect: the apparent “optimal” strategy depends critically on the metric through which performance is evaluated.

The Observer Effect

Consider two investors: a trend-following portfolio manager at an institutional firm, and a well-capitalized retail investor running a similar strategy independently. While they may trade the same signals, their objectives, and therefore their utility functions, are fundamentally different.

The institutional manager typically aims to maximize AUM. To attract and retain capital from institutional allocators, he must operate within well-defined risk constraints. In practice, this means delivering stable volatility and strong risk-adjusted returns, as measured by metrics such as Sharpe. From this perspective, Volatility Targeting becomes the natural choice, as it stabilizes risk over time and aligns the strategy with the expectations of external investors.

The retail investor operates under a very different set of constraints. He may subscribe to the idea that “you don’t eat Sharpe ratios”, focusing instead on absolute returns. Moreover, his evaluation framework is often less path-dependent. Rather than monitoring daily equity fluctuations, he may assess performance at the trade level, once positions are closed.

For this investor, approaches such as Volatility Parity or Pyramiding may be more appealing. Volatility Parity offers simplicity and lower turnover, while VPP, as shown earlier, maximizes reward efficiency under the Profit Factor metric by amplifying exposure to the right tail without materially worsening average losses.

Curiously, these two practitioners observe the exact same strategy from different angles, and the ideal sizing implementation depends strictly on their point of view.

This is what we refer to as the observer effect. In quantitative finance, the apparent optimum is not absolute, but conditional on the metric, the constraints, and the utility function of the observer.

More broadly, this challenges a common belief in systematic investing: that sufficiently rigorous analysis will always lead to a single, unambiguous answer. In reality, even in a fully quantitative framework, the “right” choice depends not only on the data, but on the context in which that data is interpreted.

Conclusion

All in all, empirical evidence supports the idea that position sizing in trend strategies is critical: it can dramatically shape how each trade contributes to the overall portfolio, and therefore it dictates the behavior of the trend program as a whole.

What stands out is that the problem of optimal position sizing in trend following may simply not have a unique solution.

All approaches we tested appear valid, depending on the metrics used to evaluate them. And which metrics make sense depends on the observation angle of the person running those strategies and making those decisions. Consequently, we suggest that no universally correct answer exists.

Regardless of these considerations, history shows that success stories can emerge from very different philosophies around sizing. As always, we encourage investors or aspiring quants to run their own numbers, think deeply about first principles and methodological implications, and then, whatever their answer might be, have the discipline to stick with it.

We know that this topic is often at the centre of debate among trend-following practitioners, and we would be very interested to hear your views.

Leave a comment

If you found this article useful, feel free to share it with other practitioners, leave a comment with your perspective, or reach out at info@concretumgroup.com

Share

In quantitative trading, backtesting data quality is a critical but often overlooked factor. Most attention goes to strategy design, execution, and risk management, yet the reliability of the underlying data provider can materially affect results.

While expanding our Opening Range Breakout (ORB) research across multiple data sources, we expected consistent results, but that is not what we found. Using identical code, parameters, and time periods, the same strategy produced materially different outcomes depending on the data provider.

This raises a fundamental question:

How much of a backtest is driven by the strategy itself, and how much by the data?

This article presents the analysis and its key findings, with full reproducibility details to be published separately.

Strategy Overview: Opening Range Breakout

Before analyzing the data, we define the strategy used throughout this study: the Opening Range Breakout, as described in Can Day Trading Really Be Profitable? (2023), co-authored with Andrew Aziz.

The formulation is intentionally simple and fully deterministic, making it suitable for cross-provider comparison.

Rules

• Opening range: first 5 minutes

• Long if 5th minute Close > Today Open

• Short if 5th minute Close < Today Open

• Otherwise Flat

• Entry: Open of 6th minute

Stops

• H/L: range low / high

• ATR: fraction of 14-day ATR

Exit

• Profit target (10R) or market closure

Sizing

• 1% risk

• Max leverage: 4×

Because the strategy relies on intraday highs, lows, and threshold-based decisions within a narrow time window, small differences in the data can propagate into materially different outcomes.

The Discovery: Same Strategy, Different Results

The key question is not whether differences exist, but how large they are.

To measure this, we ran the same backtest over the last 10 years across four different data providers:

• Massive (Polygon)

• Alpaca

• IQFeed

• Interactive Brokers

All tests used identical dates, code, and parameters, as well as a starting capital of $25,000.

Two stop configurations were evaluated:

1. H/L Stop – An opening range high/low (H/L) stop based on intraday extremes,

2. ATR Stop – An ATR-based stop derived from daily volatility.

H/L Stop Results

Under the H/L configuration, results diverge sharply. Final portfolio values range from $226k to $726k, representing more than a threefold difference from identical strategy logic.

ATR Stop Results

In contrast, the ATR-based configuration produces far more consistent results across providers. While performance differences remain, the overall behavior converges and the dispersion is significantly reduced.

This divergence reflects how the strategy interacts with the data. When stops rely on intraday highs and lows, small differences in price can compound into large performance gaps. When stops are based on daily volatility, sensitivity to intraday variation is reduced.

This leads to a more fundamental question: what specific differences in the data drive these outcomes?

Before analyzing each issue in detail, we summarize the main sources of discrepancy identified in the data:

• Phantom Highs and Lows — isolated price spikes not confirmed across providers

• Stale Bars — sequences of unchanged prices indicating missing or improperly aggregated data

• Early-Close Leakage — data extending beyond official market hours on shortened sessions

• Tick-to-Bar Assignment — Differences in how providers assign trades occurring at boundary timestamps (e.g., 09:35:00) to adjacent bars.

• Venue Coverage and Trade Aggregation — differences in the set of underlying trades used to construct bars, due to varying data sources (e.g., SIP-based consolidation, direct exchange feeds, or partial synthetic aggregation), leading to discrepancies in intraday highs, lows, closes, and volumes

We now examine each of these issues in turn.

Data Quality Issues

Issue 1: Phantom Highs and Lows

When comparing minute-level High and Low values across providers, we observe that most series align closely, but isolated spikes appear in a single dataset and are not confirmed by the others (e.g., Polygon in the examples below).

Although these discrepancies are small and infrequent, they directly affect strategies that rely on intraday extremes. A slightly higher reported high or lower low can shift stop levels, alter position sizing under fixed risk, and ultimately change trade outcomes. For H/L-based stops, even minor differences can materially impact backtest results.

These discrepancies are typically not driven by actual market activity, but by differences in how providers construct their datasets. U.S. equities trade across multiple exchanges and off-exchange venues such as dark pools, where trades may be reported to the tape with a delay. As a result, some transactions can appear out of sequence relative to the prevailing market. Depending on the provider’s filtering and consolidation rules, these prints may be included, excluded, or adjusted, leading to isolated highs and lows that are not consistently observable across datasets.

Issue 2: Stale Bars

A second issue observed in the data is the presence of stale bars, defined as 1-minute candles where Open = High = Low = Close. While occasional occurrences are expected, extended sequences of such bars in a highly liquid instrument like TQQQ are unlikely to reflect actual market activity and instead point to missing or improperly aggregated data.

For intraday strategies, this has direct consequences. During the opening range, stale bars can collapse the range to a single price, reducing stop distance and destabilizing position sizing. Outside the opening window, they suppress volatility and can prevent stop or target conditions from being detected, leading to delayed execution and price gaps. In both cases, the time series no longer reflects continuous trading, and backtest results become unreliable.

A Case Study on Stale Bars: IBKR Data Used in the Original Paper

The original ORB study was written in early 2023 and used intraday data downloaded from Interactive Brokers (IBKR The figure below shows the equity curve reported in that study.

At first glance, these results appear inconsistent with the IBKR backtests presented earlier in this article. This raises a natural question: if the same data provider is used, why do the results differ?

To investigate this, we compare the dataset used in the original study with the same trading days, re-downloaded from IBKR on March 24, 2026.

The difference is not subtle. The earlier dataset shows continuous intraday price variation consistent with active trading, while the more recent download contains extended sequences of flat, unchanging prices. In some sessions, more than 350 out of 390 one-minute bars show no price movement.

These differences are structural. The same provider, queried at different points in time, can produce materially different datasets, leading to different backtest outcomes.

For this reason, a trading strategy that appeared profitable in a backtest conducted in 2023 may lead to completely different conclusions when evaluated using data downloaded three years later. The issue is not necessarily the strategy itself, but the underlying dataset, which has changed over time. This highlights the critical importance of validating data integrity before drawing any conclusions from historical simulations.

Issue 3: Early-Close Day Leakage

On certain trading days, U.S. markets close early (typically at 13:00 ET). However, some data providers continue to report intraday bars beyond the official session close. These bars do not necessarily reflect actual tradable conditions during the regular session, but rather extend the time series into periods that fall outside standard market hours.

For backtesting, this introduces a structural inconsistency. The strategy implicitly assumes a different trading regime, effectively allowing execution outside regular trading hours. As a result, key implementation assumptions break down: market-on-close orders can no longer be used as intended, and the strategy may be exposed to liquidity conditions that are not representative of the regular session.

This issue also creates cross-provider inconsistencies. Some providers truncate the session correctly on early-close days, while others include additional bars beyond the official close. As a result, identical strategies can produce materially different outcomes depending on how each dataset handles these sessions.

Although early-close days are relatively infrequent, the impact is systematic and can bias results if not explicitly addressed. In practice, this requires enforcing a consistent session definition across the dataset. Fortunately, several Python libraries and data-cleaning workflows allow researchers to detect and correct these discrepancies, ensuring alignment with actual tradable hours.

Issue 4: Tick-to-Bar Assignment

A more subtle, yet highly impactful, source of discrepancy arises from how data providers assign individual trades (ticks) to time bars. In particular, trades occurring exactly at time boundaries (e.g., 09:35:00) may be included in either the preceding bar or the subsequent one, depending on the provider’s aggregation convention.

While this difference may appear negligible, it has direct consequences for strategies like ORB. The signal depends on the relationship between the opening price and the close of the 5-minute bar. A single trade allocated differently can slightly shift the bar close, potentially flipping the signal from long to short, or vice versa.

To quantify this effect, we extended the analysis by adding a fifth provider, Databento, and restricted all datasets to the same overlapping window of approximately 750 trading days.

Across this common sample, four providers produce broadly consistent results, while Databento exhibits significantly higher performance.

H/L Stop

ATR Stops

The source of this divergence is highly concentrated.

On just 16 out of 753 days (2.1%), Databento generates a different trade direction compared to the other providers. In most cases, the difference is extremely small, often as little as one cent in the 5-minute close. However, given the binary nature of the signal, this is sufficient to alter the trade.

The following examples illustrate these boundary cases:

When these discrepancies occur on strongly trending days, their impact is amplified. A small number of signal flips, concentrated in high-return environments, can meaningfully reshape the equity curve and create the appearance of superior performance.

The implication is not that one provider is correct and another is wrong, but that differences in tick-to-bar assignment can propagate into materially different outcomes for threshold-based strategies.

Issue 5: Venue Coverage and Trade Aggregation

A further, often overlooked source of discrepancy lies in how the underlying trades are sourced before being aggregated into time bars. U.S. equities trade across multiple venues, including exchanges such as NYSE, Nasdaq, and Cboe, as well as off-exchange facilities. Since no single venue captures all transactions, different data providers rely on different inputs, such as consolidated SIP feeds, proprietary direct exchange feeds, or internally aggregated subsets of venues.

Even when focusing only on executed trades, these choices lead to different datasets. SIP-based data aims to capture a broad cross-venue view, while direct-feed-based datasets may include only a subset of venues. For example, the Databento US Equities Mini package we have used is built by aggregating, through a proprietary methodology, a subset of trading venues and therefore represents only a synthetic “mini” market view rather than a fully consolidated one. Databento also offers direct feeds from multiple venues, but these require significantly more complex aggregation to reconstruct a comprehensive view of the market.

These differences directly affect the construction of 1-minute bars. Missing or additional trades within a time interval can change the observed high, low, close, and volume of a bar. For strategies sensitive to intraday extremes, these discrepancies can translate into materially different signals, even when each dataset is internally consistent.

Conclusion

This article set out to answer a simple question:

If you run the same backtest over the same period using different data sources, do you get the same results?

The answer is NO!

Using the Opening Range Breakout (ORB) strategy from our paper Can Day Trading Really Be Profitable? (2023), and applying identical code, parameters, and trading days across multiple providers, including Massive (Polygon), Alpaca, IQFeed, Interactive Brokers, and Databento, we find that results can diverge materially depending on the data source.

In fact, identical code applied to different datasets can produce significantly different outcomes, with performance dispersion exceeding a threefold difference under certain configurations. These discrepancies are driven by a combination of data issues, including phantom highs and lows, stale bars, early-close leakage, venue coverage and trade aggregation and differences in tick-to-bar assignment, where small variations in how trades are aggregated can flip threshold-based signals.

Once these data issues are identified and addressed, results across providers converge to a consistent conclusion: independent data sources confirm the validity of the backtest results presented in our original paper.

The broader implication is that data is not a neutral input. Differences in how it is constructed, aggregated, filtered, or updated can materially alter conclusions about a strategy, even when the logic is unchanged.

For this reason, backtest results should never be taken at face value. Proper validation requires not only robust strategy design, but also a systematic approach to data quality. Knowing where to look, and how to detect and correct these issues, is an essential part of any serious quantitative research process.

If you found this article useful, feel free to leave a comment. For any questions, reach out via direct message or email at info@concretumgroup.com.

Scope and limitations

This study evaluates the ORB strategy under a specific set of assumptions, instruments, and market conditions. Results are not universal and depend on data quality, execution assumptions, and implementation details. Small changes in these inputs can lead to materially different outcomes. Independent validation is recommended.

Disclosure

This analysis is conducted independently for research and educational purposes. We are not affiliated with, sponsored by, or compensated by any data provider referenced in this study.

The first rule of intraday trading is simple: every position must be closed before the end of the trading session, with no risk carried overnight.

It sounds so obvious that very few traders stop to ask whether such an unquestioned rule can actually hide exploitable information behind it.

In other words, what if an intraday trend signal does not fully exhaust its efficacy by the closing bell?

At first glance, deciding to carry a trend trade overnight can look like a minor implementation tweak. In reality, it changes the character of the strategy... so much so that this single design choice appears to introduce exposure to an entirely different risk premium, one that has been extensively documented in the academic literature

Full Article

For quantitative traders, we believe this piece can serve as a useful reminder: no matter how established certain rules are, sometimes the most valuable insights come from deliberately breaking them.

If you found this article useful, feel free to leave a comment and reach out via direct message or email at info@concretumgroup.com for any questions.

Many long-term investment plans do not fail because markets stop compounding.

They fail because investors do not live inside backtests and historical simulations. They live through ugly drawdowns, frightening headlines, and the temptation to make a terrible decision at exactly the wrong time.

That is why today we want to try to frame volatility targeting differently.

Not as a return-enhancement method, but as a way to protect the compounding process from the behavior that most often interrupts it.

As researchers, we are aware that returns are notoriously hard to predict. Volatility is not easy either, but at least we know it tends to cluster. Turbulent periods often remain turbulent for a while, and calmer periods tend to persist. We find that this can give investors something they can systematically respond to, without pretending they can forecast where markets are headed.

And the behavioral implications can be rather valuable: a smoother path does not merely look better on a chart. Sometimes it can make the difference between staying invested and capitulating in exactly the regimes where discipline matters for capturing the market’s long-term returns.

Full Article

If that framing is right, then we argue that volatility targeting should not be seen merely as an abstract set of rules that quants enjoy tinkering with.

It can be about making the investment journey survivable enough that compounding is allowed to do its work.

If you found this article useful, feel free to leave a comment and reach out via direct message or email at info@concretumgroup.com for any questions.

Cryptocurrency markets trade 24/7. No opening bell, no regular trading hours, no holidays or weekend closures.

Driven by this peculiarity, academic literature has shown interest in investigating day-of-week seasonal effects in both Bitcoin and broader crypto markets.

With this piece, we wanted to explore a related yet distinct question. Instead of examining passive Bitcoin returns across the week, we asked whether particular seasonalities exist within an actively managed framework.

This means we are less interested in discovering if Bitcoin is more likely to move up or down in a given window, and more in whether it moves with enough persistence for trend-following models to successfully exploit such price action.

In our study, we construct a volatility-managed benchmark as an ensemble of several high-frequency trend models, and map how its behavior shifts (on average) across the week. The results point to something we find genuinely interesting, something we have come to refer to internally as the Monday Asia Open Effect.

Full Article

In our article, we explain why we believe this phenomenon may reflect positioning flows tied to institutional participation, potentially pointing to the rebalancing habits of larger investors that have become increasingly relevant in crypto markets in recent years.

For anyone building or evaluating systematic crypto strategies, we think such considerations deserve attention.

If you found this article useful, feel free to leave a comment and reach out via direct message or email at info@concretumgroup.com for any questions.

Researchers often talk about volatility as if it were a clean number hidden inside the market, waiting to be measured with absolute precision.

It is not. Volatility is inherently latent: there is no single instrument that reveals it as one exact number. The best we can do is rely on proxies: close-to-close returns, intraday realized measures, range-based estimators, GARCH-family models...

Each captures something useful, but at the same time none provides a definitive answer.

Once we accept that, a natural follow-up question emerges.

When someone claims one volatility model is better than another, better relative to what exactly?

In practice, it is structurally impossible to forecast “true” volatility: we are forecasting tomorrow’s value of a chosen benchmark proxy. Change that proxy and the target changes, change the loss function and you may start preferring an entirely different model.

On top of that, certain models can be systematically biased, as implied volatility often is, running persistently above realized due to the volatility risk premium. Or sometimes the benchmark itself can be noisy enough to make genuine forecasting skill look unstable: these are different problems, and they call for different responses.

Full Article

For anyone working with volatility targeting, options pricing, or risk models, this article is not an academic exercise.

It is a practical reminder that forecast quality only makes sense once you are explicit about the target you are trying to hit, the horizon you care about, and the loss function you are willing to score against.

If you found this article useful, feel free to leave a comment and reach out via direct message or email at info@concretumgroup.com for any questions.

For many quantitative traders, a standard operational workflow can sound something like: compute a signal at the end of the trading session, execute at the next open, and keep live trading neatly aligned with an end-of-day backtest.

It is simple in practice and easy to explain, but our experience suggests it is often suboptimal.

If a signal decays slowly, executing it at the next day’s open should have limited impact. However, when the alpha is driven by short-term mean reversion or microstructure effects, the roughly seventeen-hour gap between 4:00 PM and 9:30 AM can erode a substantial portion of its predictive power.

Using intraday-resolution data, we compare two implementations: one that generates the signal at the close and executes at the next day’s open, and another that forms the signal 15 minutes before the close and executes via a market-on-close order on the same day. This allows us to address a question that research notes often avoid.

At what point does execution timing stop being an implementation detail and start becoming a first-order driver of performance?

Something worth noting is that execution policies are not just about speed.

Modern technology makes near-instant signal computation feasible, but certain windows of the trading session face particular liquidity challenges, with spreads widening and slippage risk increasing meaningfully.

Full Article

We find that the balance between signal decay, signal-to-execution lag, and liquidity (particularly for short-term trading programs) is not a secondary concern: it can determine whether a compelling research result survives first contact with live execution.

If you found this article useful, feel free to leave a comment and reach out via direct message or email at info@concretumgroup.com for any questions.

Many trading ideas can appear promising based on canonical trade-level statistics. High average returns per trade, strong hit rates, and attractive full-sample results: that is often where many researchers decide whether a signal is worth pursuing.

But what if the most attractive trades appeared in a form that no real portfolio could fully exploit?

When designing mean-reversion programs, signals often do not arrive smoothly. Instead, they tend to appear in concentrated bursts, especially during market selloffs, when dozens or even hundreds of names can trigger short-term oversold conditions at the same time.

Empirically, we find that these days tend to produce some of the strongest winners across the full trade sample. However, several critical considerations remain.

Capital is finite. Leverage limits are finite. Execution capacity is finite.

What happens when we impose realistic capacity and operational constraints on the days when the strategy generates an unusually large number of attractive signals?

And right when the opportunity set explodes, simple trade-level statistics can depict a situation that would be practically impossible to harvest at full size.

We call this distortion opportunity set bias.

Given the effect it can have in inflating simulated results relative to achievable real-world performance, we decided to explore this phenomenon in greater depth using a cross-sectional mean-reversion strategy as a case study.

Full Article

If you found this article useful, feel free to leave a comment and reach out via direct message or email at info@concretumgroup.com for any questions.

An opening-range breakout can look clean on a chart and elegant in simulation.

Then the market opens.

Suddenly you need to scan hundreds of stocks, rank them in real time, size positions consistently, place multiple orders at once, manage stops, and do all of it without hesitation or emotional interference.

That is where many promising signals risk to fall apart.

In this piece, we look at the part of intraday trading that is often underestimated: not the signal itself, but the technological stack required to turn that signal into something practically tradable.

Unfortunately, there’s no place to hide: the faster the signal you trade and the more names you monitor, the less room there is for manual execution.

Therefore, the real question we should ask ourselves is not whether automation is useful: it is whether some intraday strategies are even tradable without it.

For anyone keen on intraday strategies, relative-volume setups, or other short-term signals, this article can give you some context of what needs to happen behind the scenes…

Read the Article

If you found this article useful, feel free to leave a comment and reach out via direct message or email at info@concretumgroup.com for any questions.

In trend-following, most of the debate goes to signals: traders argue about breakouts, moving averages, filters, momentum definitions, and exits.

And yet, position sizing is often dismissed as a mere footnote.

Our evidence suggests this is a mistake. In our research practice, both for ourselves and for institutional clients, we consistently find that much of what a strategy looks like in practice depends on how risk is scaled once a trade is on, and trend-following programs are no exception: the same trend-following signal can produce a very different experience depending on how positions are sized.

In this article, we want to compare three sizing approaches applied to the same core strategy, looking for an answer to a very simple question.

How much of long-run performance is driven not by the entry itself, but by how exposure is calibrated once a signal has fired?

That question matters because each sizing choice reshapes the personality of the program: its return distribution, drawdown profile, and the kind of patience it demands from the investor.

Full Article

Certain sizing methodologies stabilize risk through time, some leave the initial allocation largely untouched, and others can become more aggressive only after the market has already moved in our favor.

This is why position sizing is far more than an implementation detail. It fundamentally shapes how a strategy behaves and what it feels like to live through, and by the time performance reports make that visible, the decision has long since been made.

If you found this article useful, feel free to leave a comment and reach out via direct message or email at info@concretumgroup.com for any questions.

Traditional finance teaches a simple principle: more risk should always be rewarded with more return.

Yet since the discovery of the low-beta anomaly (Black, 1972), investors have known that less volatile stocks tend to offer more attractive returns than classical theory would predict.

Counterintuitive, yet empirically persistent. In this piece, we seek an answer to a related and practical question.

Is there a point where rising stock volatility stops being productive and starts acting as a drag on expected returns?

Using survivorship-bias-free data on Russell 3000 constituents from 1990 to 2024, we find that the answer is yes. Historical stock returns peak when annualized volatility sits within a certain range, beyond which the relationship flattens and eventually inverts. Past a certain threshold, additional volatility tends to subtract from performance rather than add to it.

On top of that, we also investigate whether extremely volatile stocks are inherently poor investments, or whether part of the problem lies in how capital is allocated to them. By applying a volatility-targeting approach and dynamically reducing exposure as stock volatility increases, we find that even highly volatile names can be tamed.

Full Article

For anyone working with factor-based equity portfolios, the question is hard to avoid: when volatility rises, are you actually being paid for it, or are you simply absorbing more risk?

If you found this article useful, feel free to leave a comment and reach out via direct message or email at info@concretumgroup.com for any questions.

Every market cycle creates its stars.

Someone posts extraordinary returns. A track record starts circulating. A story forms around rare talent or unusual genius. By the time the narrative reaches a wide audience, the idea that those returns were simply a product of luck is typically out of the question.

Today, we are conducting a simple but revealing experiment.

What if a large group of “traders” just picked stocks at random, year after year, and we ranked them on a leaderboard the way the real world often ranks investors? How impressive could the winners look if nobody had any actual edge?

The uncomfortable answer is… very impressive: probability alone can generate outliers that look extraordinary.

We find that if the sample is large enough, the right tail of the distribution will inevitably produce performances that invite admiration and mythmaking.

That does not mean every exceptional trader is merely lucky. Rather, it suggests that exceptional returns, by themselves, are far weaker evidence of true skill than most people would like them to be.

Full Article

Once we accept that, the real question changes. What should count as genuine evidence of skill? Persistence across time? Controlled drawdowns? A coherent and testable process? The ability to explain the source of returns in a way that survives known factor exposures?

For anyone who follows fund rankings, social-media leaderboards, or famous trader lore, this is a useful corrective. Before admiring the winner, maybe you should ask how many invisible losers had to exist for that winner to appear.

If you found this article useful, feel free to leave a comment and reach out via direct message or email at info@concretumgroup.com for any questions.