Overfitting

Overfitting occurs when a strategy's parameters are calibrated so precisely to the quirks of a historical dataset that the model has effectively memorized noise rather than learned a genuine edge. The result looks brilliant on the training data and collapses the moment it touches live markets or even a fresh out-of-sample period.

Tell-tale signs:

• An implausibly high in-sample Sharpe ratio — real edges rarely produce Sharpes above 2-3 at daily resolution without extraordinary conviction.
• Performance degrades sharply as soon as a new date range is used.
• The strategy required dozens of parameter combinations to find, but only one was "kept."
• Adding more rules always improves the backtest.

Why it happens. With enough free parameters you can fit any finite dataset perfectly. Each extra rule or threshold is an opportunity to capture a historical accident rather than a structural edge.

The cure. Reserve an out-of-sample period before you touch the optimizer, and use walk-forward analysis to build an equity curve made entirely of unseen data. If the Sharpe collapses out-of-sample, the strategy was overfit.

Quantle research outputs include out-of-sample metrics — treat in-sample numbers as exploratory only. Past patterns are not a promise of future results.