Also called Overfitting

Curve fitting

Tuning a strategy until it memorises one stretch of history. It produces beautiful backtests and no future returns, which is why results are held out and re-tested.

Updated 2026-09-03

Curve fitting, or overfitting, is tuning a strategy until it describes the past beautifully and the future not at all. It is not a mistake made by careless people. It is the default outcome of ordinary, well-intentioned optimisation, and avoiding it takes deliberate effort.

How it happens without anyone deciding to do it

You test a moving-average crossover. It loses money. You try a different pair of windows. Better. You add an RSI filter. Better again. You restrict trading to certain conditions. Better still. At every step you were reasoning sensibly, and at the end you have a rule set whose parameters were selected by the noise in one particular history.

The tell is fragility. A robust strategy shows a broad plateau of parameter values that all work reasonably. An overfitted one shows a sharp peak: the exact numbers chosen are dramatically better than the numbers either side of them, which is not how a real market edge behaves.

The defences are: fewer parameters, a hypothesis formed before the test rather than after it, an out of sample period held back and looked at once, and honesty about how many variants were run. The last is the one people skip, and it is the one that matters most.

See also

  • Out of sample — Testing on data the search never saw. Angie hides the last quarter of history, then re-runs the winners on it, and flags any whose result collapses.
  • Backtest — Running a set of rules over history to see what they would have done. It shows what did happen, never what will, and it flatters any rule tuned on the same data.

← See this one in the deck

Educational content, not investment advice. Angie explains how a valuation is built so you can judge it yourself. What you do with that is your call.

IsiofiaIsiofia