Quant
How many trades is enough?
A backtest comes back with 34 trades, a 62% win rate, and a tidy profit. It feels like evidence — you can see every trade, they're right there. But a strategy with no edge whatsoever produces exactly that result surprisingly often. Before you ask whether a strategy is good, you have to ask whether you have enough trades to tell.
Small samples lie loudly
Start with a coin. Flip it 30 times: how often does a fair coin — a strategy with zero edge — produce 18 or more heads, a 60% "win rate"? Roughly one time in five. Not a fluke, not a rare event. One in five. Which means if you and four friends each test a completely worthless strategy over 30 trades, one of you walks away convinced.
Now recall that you probably tested more than five variants. This is where sample size and the multiple-testing trap compound each other viciously: small samples produce impressive-looking noise, and searching many strategies guarantees you'll find the most impressive noise available.
The rule of thumb worth memorising
The statistic that matters is roughly your per-trade edge divided by your per-trade noise, multiplied by the square root of the number of trades. Two consequences fall straight out of that square root, and they're the whole post:
- Halve your edge and you need four times the trades. Evidence gets expensive fast as the edge gets subtle — and real edges are subtle.
- An annual Sharpe of 1.0 takes about four years of data to become statistically convincing. A Sharpe of 0.5 takes around sixteen. That is not a data problem you can code your way out of.
Which reframes the Sharpe ratio in a useful way: it isn't only a quality measure, it's a measurement speed measure. A higher Sharpe means you need less history to know you have something. Strategies that trade rarely aren't just slower to earn — they're slower to verify, and you may never accumulate enough trades in a lifetime to be sure.
Thirty trades can't distinguish a real edge from a coin flip — a worthless strategy hits 60% wins over 30 trades about one time in five. You typically need hundreds to low thousands of trades before the numbers mean anything.
When you can't get more trades
Most strategies can't simply produce more history. The legitimate moves:
- Trade more instruments, not more often. Running the same logic across 50 uncorrelated markets multiplies your sample without changing the idea. This is why breadth is prized — it buys statistical power honestly.
- Pool similar assets. If the effect should hold across a family of instruments, test it as one hypothesis across all of them rather than fitting each separately.
- Shorten the holding period — if, and only if, the edge genuinely exists at that horizon. Otherwise you're just paying more costs for fake precision.
- Accept the uncertainty and size for it. A thin sample doesn't forbid trading; it demands smaller size. This is exactly what Kelly formalises — bet proportionally to your confidence, and half-Kelly exists because your edge estimate is itself uncertain.
The move that isn't legitimate: over-trading to inflate the count. Costs scale with trades while edge doesn't, so you'll reach statistical significance on a strategy that's now unprofitable after fees. Congratulations on proving it precisely.
Sample size is a claim about the future
Underneath all of this is a simple, humbling point. Your backtest doesn't measure your edge — it produces one noisy estimate of it, and the noise shrinks embarrassingly slowly. That's why Monte Carlo is so useful: reshuffling your trades shows you the range of equity curves your own strategy could plausibly have produced, and it's usually far wider than the single curve you've been staring at.
So before optimising anything, count your trades and ask what that number can support. Against the random walk, a small sample isn't weak evidence — it's usually no evidence at all, dressed up as a chart.