My Trading System Has a 73.5% Win Rate and Loses Money: A 16-Year Backtest Audit

I spent a few days auditing a live FX system I built, over 16 years of history. The result was negative in a way I did not expect, and the diagnostic that found it is general enough to be worth writing up — it applies to any parameter search, not just trading.

Every number below comes from an actual measured run. Nothing is illustrative.

The setup

17 configurations, 4-hour entry bars, exits simulated on 1-hour bars, 2010-05-28 to 2026-07-08, 7,641 trades. Crucially I fed the unmodified production functions historical CSV instead of a live data feed, so the code path being tested is the code path that runs live.

Bug 1: the simulation charged costs to the wrong place

Platform OHLC bars are BID prices. The backtest placed entry at the bar close, put the take-profit and stop-loss at close ± n·ATR, checked touches against bid highs and lows, then subtracted the spread from the final P&L.

Live, a long position fills at the ASK, the bracket sits relative to that ask, and it closes on the BID. So live needs price to travel one extra spread to reach the target, and reaches the stop one spread sooner.

Subtracting cost from the result is not the same as charging it to the trigger. The first changes how much you win. The second changes which trades win at all.

Measured: 2–5 percentage points of win rate, always against you.

This generalises well beyond trading: if your simulation applies costs as a post-hoc adjustment rather than modelling the mechanism that generates them, your event counts are wrong, not just your totals.

Bug 2: the win rate was a design choice, not evidence

Targets were 0.5–0.8×ATR, stops 1.5–2.0×ATR. That is a reward:risk of 0.25–0.4, which fixes the breakeven win rate by arithmetic:

breakeven win rate = SL / (TP + SL) = 75–82%

Out-of-sample, using walk-forward test slices only: 4,947 trades, 73.5% win rate, -4,669.9 pips, -0.944 pips per trade.

A 73.5% win rate that loses money is not a paradox. Small targets get hit often. That is what small targets do. The high win rate was purchased by accepting a loss four times the size of the win — it was never evidence of an edge.

The diagnostic worth stealing

Before re-optimising anything, I asked whether the parameter surface was learnable at all.

For each walk-forward fold: evaluate the entire parameter grid on the training window and on the test window, then take the Spearman rank correlation between the two.

if rho > 0  -> training rank predicts test rank; selection is meaningful
if rho ~ 0  -> the surface is noise; NO selection rule can help

That second line is the valuable part. It separates "my optimiser is bad" from "there is nothing here" — two situations that look identical from the outside and demand opposite responses.

Result across 119 folds:

• median rho: -0.024
• mean rho: -0.030, 95% CI [-0.090, +0.030] — includes zero
• folds with rho > 0: 47.9% — a coin flip

Zero information. Consistent with that, walk-forward optimisation with plateau smoothing produced worse out-of-sample results than the parameters I never touched, and selected the grid boundary in 11 of 17 configurations — the signature of an optimiser with nothing to grip.

Then I checked whether the effect exists anywhere

38 instruments, non-overlapping holding periods, volatility-scaled, pooled by asset class so 1,140 tests collapse to 6, split first-half versus second-half:

Asset class	Sharpe @0bp	Breakeven cost	1st half	2nd half	Stable
FX major	+0.08	3bp	+0.06	-0.09	no
FX cross	+0.02	1bp	-0.05	-0.06	no
Index	+0.07	5bp	+0.04	+0.02	yes
Metal	+0.12	12bp	+0.13	+0.03	yes
Energy	+0.13	23bp	+0.15	-0.02	no
Crypto	+0.42	100bp	+0.54	+0.15	yes

Two things I would not have guessed:

Breakeven transaction cost is more decision-useful than Sharpe. FX majors break even at roughly 3bp and crosses at roughly 1bp — both below the actual spread. The question "is there an edge here" is settled before you write any strategy code.

The significant result was the fake one. Energy shows t = 3.73 on the full sample. It is flat-to-negative in the second half. Without the split it would have looked like the strongest finding in the table.

Crypto was the only survivor, and it decays hard (0.54 to 0.15). At realistic CFD spreads the second half turns negative. Historical effect, not a current one.

The actual lesson

Build the falsifier before the strategy.

The reason a long series of configurations looked "validated" is that no component in the pipeline had the job of saying nothing here. Every part was designed to find something, so it found something, every time.

The rank-correlation test above took an afternoon and invalidated months of parameter work — including the one lead I was most confident about. That is the test doing its job, and it is the cheapest insurance I have added to a research process in a long time.

---

This is research documentation, not investment advice. Past results do not indicate future results. Trading CFDs carries a high risk of loss.