Our live track shows an 89% win rate. Our honest expectation is a profit factor of 1.12.

Both numbers are ours, both are published, and the smaller one is the one to plan around. Here is exactly why they differ — and how we check that we are not fooling ourselves.

The two numbers

On the dashboard you will find the live paper track: several thousand closed trades, a win rate around 89%, and a profit factor well above 5. Every one of those entries was published in the public channel before it closed, so none of them was chosen after the fact.

Alongside it we publish a second, much duller figure: a median profit factor of 1.12 per semester across 17 consecutive semesters of walk-forward testing, 2017H2 to 2026Q1. Worst semester 0.97. Best 1.35. Sixteen of seventeen profitable.

A reasonable person seeing 89% next to 1.12 concludes that one of them is a lie. Neither is. They measure different things, and the gap between them is the most useful thing on this website.

Why the live number is so much prettier

Three reasons, in order of how much they matter.

1. One regime is not the distribution. The live track covers months, not years. The walk-forward test covers 8.75 years, including the 2018 bear market, the 2021 alt-season, the LUNA and FTX collapses, and the flat, liquidity-starved stretch of 2023 that produced our single losing semester. A few good months tell you what the strategy does in those months. They do not tell you what it does in 2023.

2. The feed that produces the public posts uses an optimistic fill model. The engine enters with limit orders. The leg that feeds the channel assumes those limits fill more readily than they would in a real order book, and it runs without some of the filters our other legs apply. We run parallel legs with harsher, more realistic fill assumptions specifically to measure that gap, and their results are visibly worse. This is a known bias in the published live figure, not a discovery someone else made about us.

3. Win rate is close to meaningless on its own. Ours is high because the exit logic banks partial profit early and trails the remainder; that converts many trades into small wins. It says almost nothing about whether the strategy makes money, because it says nothing about the size of the losses. On the leg with honest fill assumptions, 36.5% of closed trades lose money, and the deepest single loss on record is about −26%. A strategy can win 89% of the time and still bleed if the 11% is large enough. That is precisely what profit factor measures and win rate does not.

What walk-forward testing is, and what ours says

A normal backtest lets you tune a strategy on the same data you then grade it on. That is how you get a beautiful curve and a losing account. Walk-forward does the opposite: fit on one period, test on the next period the strategy has never seen, then roll forward and repeat, across the entire history. It cannot borrow information from the future, which is why its answers are so much less flattering.

Results depend heavily on account size, because position sizes run into the liquidity of the pairs being traded. We publish the retail tier on the dashboard, and here is the whole set:

Per-semester profit factor and, in the last column, the maximum drawdown when the semesters are compounded into one curve. Win rate barely moves across the whole table — 72.9% to 73.1% — which is another way of saying the win rate is not where the differences live.

Two things are worth staring at. Small accounts do not do better; they do worse per unit of pain, because a fixed risk budget forces them into thinly traded pairs. And the drawdown column is the number most services never print. On the retail tier, planning around a fifth of the account being underwater at the worst point is realistic. Planning around 2.8% — the current live figure — is not.

How we check that we have not fooled ourselves

Walk-forward alone is not enough. If you try enough variants, one of them will look good on out-of-sample data by luck. So we measure that risk directly.

Probability of backtest overfitting. We took 39 configurations across six parameter grids over 49 months and ran a combinatorially symmetric cross-validation — thousands of train/test splits in every possible order. It returns the probability that a configuration which ranked best in-sample is actually below median out-of-sample. Ours came out at 0.018. The deflated Sharpe ratio, which penalises the result for the number of things we tried, stays at its ceiling even when we tell it we ran thousands of trials.

And the part nobody publishes. We also run an anytime-valid sequential test on the live legs — a confidence sequence that can be checked every day without invalidating itself. On the leg with honest fill assumptions, the mean return per unit of risk is currently +0.006 with a confidence interval that still straddles zero. In plain terms: on live data alone, our edge is not yet statistically proven. The test says we need on the order of 9,500 closed trades to settle it, and we are not there. The multi-year walk-forward evidence is much stronger than the live sample, and we are not going to pretend the live sample says more than it does.

What to actually expect

• Plan around a profit factor near 1.1–1.2, not 5. A profit factor of 1.12 means that for every dollar lost, about $1.12 is won. It is a real edge and a thin one.
• Expect roughly a third of trades to lose, and expect the occasional loss far deeper than the intended stop when a thin market gaps through it.
• Expect drawdowns in the tens of percent, not single digits.
• Do not size from the live win rate. It is the least stable number we publish.

If that reads as an anti-advertisement, consider what the alternative advertisement is worth. Every figure here is checkable: entries are timestamped publicly before they resolve, losses stay up, and the numbers are net of fees, slippage and funding. We would rather you arrive with correct expectations than leave with disappointed ones.

Check it yourself

The live record, updated continuously, is on the dashboard. Real trades with their reasoning, including the twelve deepest losses, are on Trade breakdowns. How the engine decides is on How it works, and the numbers are defined in the FAQ. Every entry appears in the free channel before its exit — that is the part you do not have to take on trust.

Educational information about a trading system, not financial advice and not a solicitation. Crypto futures carry a high risk of loss. The track described here is paper trading: no real-money orders are placed. Past performance does not predict future results — that is, in fact, the entire point of this article.