Our AI is a better seller than a buyer

We had a suspicion. Somewhere around the sixth month of live operation, a pattern kept surfacing in the weekly numbers that nobody had designed for: the desk seemed to make money when it sold and struggle when it bought. Suspicions are cheap, and a trading desk that runs on suspicion is a trading desk that eventually blows up. So we went and checked it properly — every trade the system has taken since January, replayed minute by minute against the real tape, at every stop distance we could reasonably have used.

The suspicion was right. It was also more lopsided than we expected, and the reason turned out to have almost nothing to do with the market being bullish.

First, a word about win rates

The obvious way to compare our short book against our long book is to look at which one wins more often. That measure is close to useless, and it is worth explaining why before any number in this article means anything.

A trade with a tight stop and a distant target should lose more often than a trade with a wide stop and a near target. That is geometry, not skill. If you place a stop 10 points below and a target 30 points above, a market that moves entirely at random will stop you out roughly three times for every once it pays you. Win rate on its own tells you where somebody put their lines, not whether they were right about direction.

So we use a different yardstick throughout. For every single trade we compute what a driftless random walk would produce given that trade's own stop and target: stop ÷ (stop + target). That is the hit rate the trade would earn from a market with no opinion whatsoever. Beating it is the only thing that counts as directional skill. Every "edge" figure below is the gap between what actually happened and what a coin-flipping market would have produced on the exact same levels.

The answer, in one table

Seven and a half months. 237 trades entered and closed — 108 short, 129 long. Here is the whole finding before we take it apart.

Longs are a coin flip, and a slightly losing one. What makes that genuinely strange is the market they were flipped in. Over this window NAS100 rose 22.7%, US500 rose 17.2%, and US30 rose 14.5%. A long book operating in a tape like that has the wind squarely behind it, and it still could not clear a baseline that assumes the market has no opinion at all.

The short book did clear it, by 11.3 points, while trading against that rally the entire time.

The boring explanation, and why it fails

Before reaching for anything interesting, the obvious objection has to be cleared: maybe the long trades were simply harder. Maybe the system was asking more of them — further targets, tighter stops — and the lower hit rate is just that difficulty showing up.

It is the opposite. Longs were handed the easier geometry on both axes: a closer target and a wider stop, a mean target-to-stop ratio of 0.952 against the short book's 1.128. They were being asked to travel less distance with more room to be wrong, and they still hit less often.

Sorting each side into quartiles by how demanding the trade was makes the shape unmistakable.

That bottom-left cell is the one worth sitting with. Long Q4 hit 27.3% where noise would have produced 42.3%. Fifteen points below random is not the signature of an edge that has gone missing. Random is what you get when your signal contains no information. To land fifteen points beneath it, across a full quartile of trades, the entries have to be systematically arriving at bad moments.

In plain terms: on its most ambitious long setups, the system was reliably buying strength that was about to reverse. That is a specific, diagnosable failure, and it is a very different problem from "the long model is weak."

Same trade, opposite direction

The cleanest test available: group every trade by its target-to-stop ratio and compare longs against shorts within the same band. Same geometry, same difficulty, only the direction differs.

Thirty-eight percent against seventy-five, on the same geometry, in the same markets, over the same seven months. Whatever is happening here, it is not about where the lines were drawn.

Which markets, exactly

Broken out by instrument, the pattern is close to universal — with one loud exception that changed our roadmap.

Read down the short column and there is no argument to have: seven instruments, seven positive edges. Read down the long column and the picture inverts everywhere except NAS100.

NAS100 longs are not a rounding error either. Thirty-four trades, a 71% hit rate against a 56% baseline, +7.29R — the single best long book we have, and better than NAS100's own short book. US30 longs finish +3.20R with a two-point edge, which is real money but statistically a shrug. Everything else on the long side is a leak.

So what if we had just stopped taking them?

This is the question the whole study exists to answer, and it is the one a subscriber actually cares about. Not "is there an asymmetry" but "what would it have been worth to act on it." We reran the same seven months as four separate portfolios, changing nothing about entries, stops or targets — only which trades get taken.

Two things in that table deserve to be said plainly rather than left in a cell.

Trading only shorts would have beaten trading everything, on every measure at once. More return (+27.70R against +20.62R), roughly half the drawdown (7.00R against 13.14R), and not a single negative month in eight. The combined book carries nearly double the pain for three-quarters of the return. That is not a trade-off; the diversification we thought we were buying by running both directions was costing us on both axes simultaneously.

The long book is not uniformly bad — it is one good book carrying ninety-five bad trades. Strip NAS100 out and the remaining 95 long trades return −14.37R on their own. Put NAS100 back and the best configuration in the study appears: +34.99R across 142 trades, with the longest losing streak cut from seven to four.

Month by month

Totals can hide a lot. A configuration that makes all its money in one lucky month is not a configuration, it is an anecdote. Here is every month of the study across the three portfolios worth comparing.

July is the month that makes the case on its own. The desk as actually run finished +0.89R — a month that looked, from the outside, like treading water. Underneath it, the short book made +9.28R and the long book lost most of it back. The published result was not a quiet month. It was two books pulling in opposite directions and roughly cancelling.

March and August are the honest counterweight. In both months the shorts-only portfolio is the only one that stays positive, but the shorts-plus-NAS100 configuration goes red — because NAS100 longs had a bad March (−1.83R) and a bad August (−3.00R). Adding a genuinely good book back to the portfolio still bought us two losing months. There is no configuration in this study that wins everywhere.

Why would a machine be better at selling?

The data says it happened. The data does not say why. What follows is our analysis rather than our measurement, and we want to be clear about which is which — every table above is reproducible from the tape, and everything in this section is an argument.

1. The bull-market intuition is a horizon error

The first reaction most people have to Table 1 is that a rising market should have helped the longs, so something must be broken. That intuition quietly assumes we are holding positions long enough to collect the rise. We are not. Ninety percent of our shorts open and close the same day, with a median hold of forty-eight minutes.

Over forty-eight minutes, a 22.7% annual rally contributes almost nothing. Drift needs weeks to matter; intraday it is a rounding error against the noise. What an intraday system actually harvests is not direction, it is volatility — and volatility is not symmetric. Down moves in equity indices are faster and larger than up moves of equivalent significance; this is the leverage effect, and it has been in the literature for forty years. A long and a short with identical target-to-stop geometry are therefore not identical bets, because the price paths available to them are not mirror images. The short simply gets to its target sooner, more often.

2. Selling is an urgency trade; buying is a patience trade

When a market falls, liquidity thins out ahead of it and resting stops cascade — the move feeds itself and travels. When a market rises, it meets supply at every step: profit-taking, covered calls, rebalancing flows. Buyers have to be recruited one at a time. That asymmetry in the order book is why declines look like events and advances look like grinds, and an intraday system with a fixed target and a seventy-two-hour horizon is structurally rewarded for trading events.

3. Exhaustion is visible; continuation is not

This is the mechanism we think matters most, and Table 2 is the evidence for it. Overextension is a measurable state — distance from VWAP, momentum divergence, position within the session range. A model reading a chart can see it. Continuation is not a chart property at all; it depends on whether new buyers arrive in the next thirty minutes, which is information the chart does not contain.

Our system reads state well and infers flow poorly. On the short side those are the same skill: it sells into exhaustion, which is exactly what exhaustion signals are for. On the long side they diverge — it sees the same overextension, reads it as strength, and buys. That is the fifteen-points-below-random cell in Table 2, and it is why the failure concentrates in the most ambitious long setups rather than spreading evenly.

4. Which explains NAS100

If the long problem were simply "our model cannot buy," NAS100 longs would not be our best-performing book. The distinguishing feature of NAS100 is momentum persistence: extreme concentration in a handful of mega-caps, and a dealer-hedging and same-day-options complex that mechanically extends intraday moves rather than damping them. On NAS100, up moves actually travel within the session. It is the one market on our roster where continuation is a real, tradable property rather than a hope.

Which reframes the conclusion into something more useful than "we are better at shorting". Shorts work everywhere. Longs work only where structural drift and intraday momentum persistence coexist — and on our roster that is NAS100, alone. Note that both currency long books are negative, and currencies have no structural drift at all.

What does not fit

A study that only reports what confirms the thesis is marketing, not research. Three things in this data cut against the story above.

The long book was fine for three straight months. April, May and June each show a healthy long-side edge of +9 to +13.5 points — during the most bullish stretch of the entire sample. If longs were structurally broken they should have been broken then too. They were not.

The collapse does not line up with fear. The tidy version of this article says the system profits from panic. July supports it beautifully: the only down month, the highest-volatility month, and the short book's best edge of the study at +22.9 points. But August rose 1.8% on low volatility and the long book posted its worst month of the year at −24.9 points of edge. A calm, rising tape did not rescue the longs. Fear is part of the story; it is not the whole of it.

February and March break the narrative arc entirely. February was the long book's best month on record (+5.73R). March was its second worst (−4.28R). Those sit at the very start of the sample, long before anything we could point to as a regime change. The honest reading is not that the long side worked and then broke — it is that the long side has always been erratic, and the short side has quietly been the stable engine the whole time.

One further caveat we would rather state than have found: August carried known infrastructure problems on our side, including repeated API credit exhaustion and a mid-month trend-model rollback. August is the worst month in the sample and it is also the month we trust least. It is the first place we are looking, and until that diagnosis is finished we are treating August's numbers as suspect rather than as evidence.

What we are building because of this

Research that does not change anything is a blog post. Three things are moving from this study into the product, and we would rather describe them before they ship than after.

Splitting the AI traders by direction. Today a trader is one instrument. We are separating each into a long book and a short book that can be enabled independently, so a subscriber can run all seven short traders and only the NAS100 long — the exact configuration in Table 5's highlighted row — without hand-filtering signals. If the numbers move and long books elsewhere start earning their place, they get switched on. The point is that this becomes a setting rather than a rewrite.

Underperforming traders get paused, not deleted. We are formalising something we have been doing ad hoc: a trader whose book stops clearing its baseline goes into storage. It keeps running against live data and keeps accumulating a record, but stops issuing entry signals. If it recovers, it comes back with its evidence attached. This is how we handled the retired instruments earlier this year, and Table 4 is a list of candidates.

Direction-aware risk in the Automated Trader. The bot currently applies one risk setting to every position. Given a short book at +11.3 points of edge and a long book at −2.8, that is plainly the wrong shape. We are adding separate risk allocation for longs and shorts, so the size follows the evidence instead of the convention.

There is one more variable that interacts with all of this, and it deserved its own investigation: where the stop actually sits. The same replay engine was run across twelve stop widths, and the answer turned out to depend on direction too — our short books want more room than the model gives them, while NAS100 longs want exactly what they already have. We wrote that up in Moving the stop loss? Let's find out, along with the configuration that combines both findings.

None of that is a promise that the next seven months look like the last seven. One hundred and eight short trades is a real sample and a thin one, and the confidence interval on almost every figure in this article spans zero. What we can say is narrower and more useful: over the period we have actually traded, the asymmetry was large, it was consistent across seven instruments, it survived every attempt we made to explain it away as geometry, and acting on it would have produced 70% more return on substantially less drawdown. That is enough to change the product. It is not enough to stop checking.

Please read this part before acting on any of it

This article is published for educational purposes only. It is a description of what our own system did over one specific stretch of market history, and an argument about why. It is not financial advice, not a recommendation to trade in any direction, and not a suggestion that you should short markets or stop buying them. Nothing here accounts for your circumstances, your risk tolerance, or your obligations.

The dataset is young. Seven and a half months and 237 trades sounds substantial in a blog post and is thin as evidence. The short book is 108 trades. A single unusual quarter could move any figure in this article materially, and the honest statistical position is that we cannot reject the possibility that some of what we measured is chance. Past performance — replayed or live — does not correlate reliably with future performance, and the smaller the sample the weaker that already-weak link becomes.

The overfitting risk here is real and we are the ones most exposed to it. We ran many configurations against one history and are now reporting the ones that performed best. That is the exact procedure that manufactures results which evaporate on live data. Some of the guard rails we used are described above — a random-walk baseline rather than raw win rate, testing whether the effect survived matching on geometry, checking that it held across seven separate instruments rather than concentrating in one — but no guard rail eliminates the problem. The more selective a configuration is, the more suspicious you should be of it, including ours. A rule that only works on the exact subset of instruments and the exact period we tested is not a discovery, it is a description of the past written in a flattering font.

Which is exactly why we are pausing the long books rather than deleting them. The asymmetry we measured is a property of a particular market regime — one specific stretch of volatility, drift and liquidity conditions — and regimes turn. A long book that lost money against a rising tape may work well in a range, or after a volatility reset, or when the flow structure that currently punishes intraday continuation changes shape. Equally, the short edge that looks robust today could compress the moment enough participants position for the same thing. Turning a book off is a decision about today, not a verdict for all time, and the paused books keep running against live data and keep accumulating a record precisely so that we can see it flip and bring them back when the evidence justifies it. Anyone applying this thinking to their own trading should hold it the same way: as a live hypothesis under continuous review, not a conclusion.