What is overfitting?

Overfitting, explained

The more parameters a model has and the more combinations are tried, the easier it is to find something that fitted the past by chance. Machine-learning models are especially prone to this when trained on limited market history.

Common defences are testing on out-of-sample data the model never saw, walk-forward testing, keeping rules simple, and checking that results hold across different periods and markets.

For users of AI stock tools, the practical check is the gap between back-tested and live results. A tool whose live, dated picks track well below its published backtest may be showing the effects of overfitting.

Example

A strategy that buys only on Tuesdays when a stock's 13-day RSI is between 41 and 44 might show strong returns from 2015 to 2020 purely by chance, then fail once traded live.

Related terms

• BacktestingBacktesting is testing a trading or investing rule on historical market data to see how it would have performed had it been followed in the past.
• Survivorship biasSurvivorship bias is the error of judging results only from the stocks, funds or strategies that still exist, ignoring those that failed or disappeared, which makes past performance look better than it was.
• AI stock scoreAn AI stock score is a single number or grade that a model assigns to a stock to summarise how favourably it rates that stock's outlook over a stated time horizon.
• Chart pattern recognitionChart pattern recognition is the automated detection of recurring shapes in price charts, such as head and shoulders, triangles, flags or double bottoms, that technical analysts read as possible continuation or reversal setups.

Where this matters

Overfitting comes up most when choosing AI stock pickers. See Best AI stock pickers for the tools we compare, with dated prices.