Essays · Checklist

10 signs your trading strategy is overfit

By Daniel Roache · 14 July 2026 · 9 min read

Every overfit strategy shares a family resemblance: it fits the noise of one specific slice of history and mistakes that fit for skill. Here are the tells, roughly in the order they're worth checking.

1. The equity curve is smoother than its own trades

Individual trades are noisy; a real strategy's equity curve should inherit some of that noise. A curve that glides upward in an almost straight line, while its trade-by-trade results scatter wildly, has usually been fitted to produce that smoothness. Real edges are lumpy.

2. The Sharpe ratio is implausibly high

Professional quant funds live on Sharpe ratios of roughly 1 to 2. A backtest showing 3, 5, or 10 isn't a discovery — it's a warning. Extraordinary risk-adjusted returns almost always shrink dramatically the moment they meet data they weren't fitted on.

3. Out-of-sample performance falls off a cliff

Split your history: build on the first portion, then test — once — on a portion you never touched. If the Sharpe drops from 1.2 in-sample to –0.2 out-of-sample, you didn't find an edge; you found the training set. A large in-sample/out-of-sample gap is the single clearest symptom of overfitting.

4. The rules contain suspiciously specific numbers

"Enter when RSI crosses 67.3 and price is above 1.23× the 17-period average." Why 67.3 and not 70? Why 17 and not 20? Oddly precise thresholds are usually the residue of a grid search that found the exact values that happened to work on the test data. Real parameters survive being rounded.

5. Performance is a peak, not a plateau

Nudge each parameter up and down by a notch and re-run. A robust edge sits on a plateau: nearby values still work. An overfit one sits on a peak: profit craters on either side of one magic value. If your strategy only works at exactly 20 periods, 20 was luck.

6. Removing the top few trades kills it

Strip out the best three to five trades and re-check. If the whole edge evaporates — or goes negative — you weren't trading a system, you were holding a lottery ticket that happened to hit. Real edges are carried by the many, not the few.

7. It has more parameters than it needs

Every additional rule, filter and threshold is another degree of freedom to fit noise. The more knobs a strategy has, the more ways it has to look good by accident. Favour a handful of robust rules over a baroque machine of exceptions.

8. The "many trades" are really one trade in costumes

Clustered signals fire together during the same move. Counting them as independent inflates the trade count and fakes statistical significance. Ask how many genuinely independent bets there were — the honest t-statistic is computed on those, not on every fill.

9. It only works in one kind of market

A trend-follower is a genius in a bull run and a donor in a range. If your strategy was built and tested in a single regime, it has learned that regime, not the market. The cure is to hold out entire regimes it never saw — see Leave-One-Regime-Out testing.

10. The gross edge is smaller than the fees

Some strategies are genuinely predictive and still lose money, because the signal is smaller than the round-trip cost of trading it. Always net fees and slippage. A "profitable" system living below the fee wall is structurally negative; it just hasn't paid the bill yet.

None of these tests is exotic. Most take minutes. The reason overfit strategies keep costing people money is not that the checks are hard — it's that a beautiful backtest is emotionally convincing, and running these tests risks taking the beauty away. That's exactly why they're worth running. For the deeper causes behind these symptoms, read why most backtested edges are statistical mirages.

Keep reading