Glossary

Overfitting

Updated 2026-07-25

Overfitting is when a model captures the accidents of the data it was built on rather than a durable pattern. It shows up as a backtest that looks superb and live results that do not resemble it. More parameters, more filters and more hand-tuning all make it easier, because every added degree of freedom gives the model another way to memorise noise.

The most common route to it in trading research is not one bad model but many attempts: test enough variants and one of them will look excellent by chance alone. The defences are unglamorous. Keep the parameter count low, test out of sample with walk-forward validation, count how many hypotheses you tried, and require a result to clear its interval lower bound rather than just its point estimate. Also see survivorship bias.

See the numbers before you pay: our walk-forward accuracy scorecard is public, with sample sizes and 95% confidence intervals per symbol.

Get the daily BTC & ETH liquidation levels, free

One short email a day: the strongest clusters above and below price. Delayed published data, no trade signals, unsubscribe instantly.