How to read a backtest

What a backtest is

A backtest runs a strategy's rules over historical data: where it would enter, where the stop and target were, how each trade ended. The output is a list of trades and metrics. It answers "did this work before", not "will it work next". Its job is to kill bad ideas before they cost money.

Five numbers to look at

1. Number of trades. 20 trades is an anecdote, 200 is statistics. The calculator above shows how wide the true win rate range is with few trades.
2. Expectancy — the average result per trade, in R or % after fees. Near zero means the strategy doesn't work, however pretty the curve.
3. Profit factor — total wins divided by total losses. Below 1 loses, 1.2–1.5 is a working strategy, above 3 on a large sample is a reason to look for a bug.
4. Max drawdown — the deepest fall from a peak. Ask yourself whether you'd sit through it with real money. Live drawdowns are usually bigger.
5. Win rate together with win/loss size. Win rate alone means nothing — see Why signal channels lose.

Overfitting

The main backtest trap is overfitting. Try a thousand parameter sets and one will look profitable on history purely by chance. Such a strategy describes past noise, not a pattern, and fails on new data. The more parameters and the fewer trades, the easier it is to overfit: five conditions tuned on 40 trades are almost certainly fitted.

Red flags in someone else's backtest

• No fees or slippage — for frequent trading that is the difference between profit and loss.
• Bull market only — almost any long strategy looks brilliant on 2020–2021.
• Look-ahead bias — entering at the close of a candle that hasn't closed, or repainting indicators.
• Survivor coins only — testing today's top coins ignores the ones that died.
• A perfect curve — smooth growth with no drawdowns is more often a code bug than genius.
• No out-of-sample test — parameters tuned and tested on the same period.

How to validate honestly

1. Out-of-sample data. Tune on one part of history, test on another the strategy never saw. That result is the one that counts.
2. Several independent periods. Profitable in each, not on average.
3. Random-entry comparison. Same stops and targets, random entry times. If random entries do as well, the edge is the market, not the signal.
4. Parameter robustness. If a 1.5 ATR stop profits but 1.4 and 1.6 lose, it's overfitting.
5. Demo or minimum size before real money.

Our example: the strategy we threw away

We tested order-flow ideas — absorption, delta divergence, buy/sell imbalances — on our own Bybit data for 20 coins. About 2,800 parameter sets, 0.13% round-trip fees and slippage. Parameters were tuned on the first 60% of history and tested on the last 40%.

The best variants showed +0.4% per trade and a profit factor near 2 on the first part. On the held-out data the same settings made −0.2% to −0.6% per trade. Not one of 2,800 variants stayed profitable. Textbook overfitting — which is why we don't run such strategies. The scans in our bot were validated the same way: several periods, random-entry comparison, 0.055% fee per side.

FAQ