How to Backtest a Trading Strategy (And How Many Trades You Actually Need)

To backtest a trading strategy properly, you replay its exact rules over enough historical trades to know whether its edge is real — and the mistake that ruins most backtests is stopping too early. You need a minimum of 50 trades to say anything meaningful, and 100 or more before risking real capital. Below 50, randomness dominates: a 30-trade sample cannot statistically distinguish a 45% win-rate strategy from a 60% one. The number that matters is per setup, not per account — and the sample must span different market regimes.

That's the answer. The rest of this post is why those numbers, shown with arithmetic you can check, plus the two validation mistakes that sample size alone doesn't fix.

The statistics, in plain English

Your backtest win rate is an estimate of the strategy's true win rate. Estimates come with error bars, and the error bars shrink slowly — with the square root of the number of trades, not linearly.

The standard error of a win-rate estimate is sqrt(p × (1−p) ÷ n). At a true win rate near 50%, the 95% confidence interval (roughly ±2 standard errors) works out to:

Trades (n)	95% confidence interval on win rate
30	±18 percentage points
50	±14 percentage points
100	±10 percentage points
300	±6 percentage points

Read the first row carefully. After 30 trades showing a 55% win rate, the honest statement is: "the true win rate is somewhere between 37% and 73%." That range contains both a losing strategy and a world-class one. You know almost nothing.

At 100 trades, 55% observed means "between 45% and 65%" — still wide, but now the downside scenario is survivable rather than catastrophic. At 300 trades the picture is real. This is why "it worked for two weeks" is not evidence; two weeks of day trading is usually 15–40 trades, the worst part of the curve.

The common guidance — 50 minimum, e.g. TradeZella's futures backtesting guide, 100+ preferred — is that table translated into a rule of thumb.

Win rate isn't even the right number

Expectancy is:

Expectancy = (Win rate × Average win) − (Loss rate × Average loss)

A 40% win-rate strategy with 2R winners is profitable (0.40 × 2R − 0.60 × 1R = +0.20R per trade). A 60% strategy with 0.5R winners is a slow bleed (0.60 × 0.5R − 0.40 × 1R = −0.10R). Sample size applies to both inputs — average win is estimated from your winners only, so a 50-trade sample with 20 winners has a 20-observation estimate of your average win. Outlier winners distort small samples hard: if one trade contributes a third of total profit, your expectancy estimate is one lucky fill wide of reality. (Prop firms encode this exact suspicion as the consistency rule.)

Practical rule: alongside 100+ trades, you want 30+ winners and 30+ losers before trusting the R-multiple averages.

The two mistakes sample size doesn't fix

Counting trades across setups instead of per setup

"I have 120 logged trades" — 50 opening-range breakouts, 40 VWAP fades, 30 news scalps — is three undersized samples, not one validated strategy. Every count in this post is per setup. If you trade three setups, you need three validations. This is the strongest argument for trading one setup until it's proven.

One regime, one month

100 trades collected in a single trending month tells you how the strategy performs in that regime. NQ alternates trend weeks and chop weeks; ranges compress on Mondays and expand into Thursday/Friday; CPI and FOMC days behave like a different market entirely (see the session breakdown). A valid sample deliberately spans:

• Trending weeks and ranging weeks
• Different months (different volatility regimes)
• Scheduled-news days, or an explicit rule that you skip them

If your 100 trades all come from one regime, treat the count as 100 with an asterisk: validated for that regime only, and the opening range breakout post shows how regime-dependence looks in practice for one concrete strategy.

How to actually collect 100+ trades without burning a year

Live trading at 2–3 setups a week means a 100-trade sample takes most of a year — which is why almost nobody has one, and why most prop-firm attempts fail on untested edges.

Market replay compresses the calendar. Replaying real historical sessions candle-by-candle at 10–50x speed, one evening covers 5–10 full trading days honestly: you see the same information the market showed in real time, place simulated orders, and log the outcome. The discipline requirements:

1. Same session window every time — if you'll trade the 9:30–11:00 open live, backtest exactly that.
2. Log every signal, including the ones that would have lost. Skipping "obvious losers" with hindsight is the classic way replay samples lie.
3. Record entry, stop, target, R-multiple, MAE/MFE, timestamp — the per-hour and per-regime breakdowns come from this.
4. Pre-commit the rules. If entry/stop/target aren't written down before the replay session, you're practicing discretion, not validating a system.

Two to three weeks of evenings produces a 100-trade, multi-regime sample — the thing that takes a year live.

A validation checklist

Gate	Threshold
Trades logged (per setup)	≥100 (50 = absolute floor)
Winners / losers separately	≥30 each
Regimes covered	Trend + chop + news handling defined
Expectancy	Positive after commissions, computed from the log
Outlier dependence	Top trade < 15–20% of total profit
Rule definition	Written before the sample, unchanged during it

Pass all six and you have something worth funding — your own account or an evaluation. Fail any and the next 100 trades are cheaper than finding out live.

FAQ

Is 20 trades enough to test a strategy?

No. At 20 trades the 95% confidence interval on a ~50% win rate is roughly ±22 percentage points — an observed 60% win rate is consistent with a true rate below 40%. Twenty trades is a pilot that tells you whether the idea deserves a real sample.

How many trades do professional quants use?

Orders of magnitude more — thousands of signals across years and regimes, plus out-of-sample and walk-forward testing. Retail day traders can't match that, but 100+ trades per setup across multiple regimes captures the same principle at achievable scale.

Does a higher win rate mean I need fewer trades?

Slightly — the standard error sqrt(p(1−p)/n) peaks at p=0.5, so extreme win rates are estimated a bit more precisely. But the improvement is small, and expectancy still depends on average win/loss sizes, which need their own 30+ observations per side.

Can I combine backtest and live trades in one sample?

Yes, if the rules were identical and you log them identically — in fact a replay-validated strategy should keep accumulating the same log live. Watch for execution differences (slippage, fills) showing up as a live expectancy drop; that gap is real information about your assumptions.

How long does it take to backtest 100 trades?

Live: often most of a year at a few setups per week. In accelerated market replay at 10–50x, roughly two to three weeks of evening sessions — one evening covers 5–10 historical trading days with honest bar-by-bar information flow.

Once you have a sample, run the numbers through the free win rate calculator — it converts wins, losses, and average win/loss into expectancy, profit factor, and the breakeven win rate your reward-to-risk requires.

Build your 100-trade sample this month

TestMax replays years of real NQ, ES, and EURUSD data candle-by-candle at up to 50x, logs every simulated order, and computes win rate, expectancy, profit factor, and equity curve for you — the whole checklist above falls out of the futures backtesting tools automatically. Start free. Simulated results don't guarantee live results.