Quantprove Glossary / Statistics
3 min read

Overfitting

When a strategy is tuned so tightly to the past that it only works on the past.

Charly del SolCo-Founder

Overfitting is when you tweak a strategy until it fits your history almost perfectly, and along the way you fit the noise. The backtest looks incredible. Then it fails live, because the awesome hot streaks never repeat.

How does a strategy get overfit?

Every time you add a rule, a filter, or tweak a parameter to make the equity curve prettier, you risk fitting the past instead of the market. Test 200 parameter combos and keep the best one, and you've hand picked the luckiest version, not the truest. A backtest with 15 conditions on 300 trades isn't a strategy. This is simple: it wins beautifully in the sample and fails the moment it meets data it never saw.

How do you catch it?

Hold your data. Build on one slice, then test on a slice the strategy never touched, your out of sample data, and see if the edge survives. If the in sample makes +0.4R a trade and the out of sample makes +0.05R, that gap was fitting, not skill. Walk forward testing repeats the trick across the whole record, which is much harder to cheat on than a single split. More trades help too, an edge measured over 760 trades is far tougher to fake than one from 40.

Can you underfit too?

Yes. Strip a system down to two rules and it can be too blunt to catch anything, robust but edgeless. The goal was never the simplest possible strategy. It's the simplest one that still holds up out of sample. And a little curve fitting is unavoidable, honestly, since you tuned something to end up here. The sin isn't fitting, it's fitting so hard you fool yourself into trading a backtest.

Frequently asked questions

Tuning a strategy so closely to past data that it captures random noise instead of a real edge. It backtests great and then underperforms live, because the noise it learned doesn't come back.
Test it on data it was never built on. A big drop from in sample to out of sample, say +0.4R falling to near zero, is the classic sign. So is a system that needs a long list of rules to look good.
Keep the rule count low, hold out data for out of sample testing, favour more trades over clever parameters, and lean on walk forward checks. Simple systems that survive fresh data beat complex ones that only shine on history.
It helps a lot but doesn't cure it. A bigger sample makes a fake edge harder to manufacture, which is why Quantprove leans on sample size in your Edge Score. Torture enough parameters and you can still overfit a huge dataset.

References

  • Expectancy — What one average trade is worth to you, after the wins and losses cancel out.
  • Out of Sample — The data your strategy never saw while you built it, kept back to test if the edge is real.
  • Walk Forward — Out of sample testing done over and over, rolling through your whole history instead of one split.
  • Monte Carlo — Reshuffling your trades thousands of times to see the range of outcomes luck could have handed you.

Turn trading knowledge into evidence.