quantitative-research

A backtest is likely overfit if the Sharpe ratio exceeds 3, train performance far exceeds test performance, or the strategy has more than 5 parameters. Use walk-forward validation and require at least 20 observations per parameter to reduce overfitting risk.

Walk-forward analysis repeatedly optimizes parameters on a rolling training window and validates on the subsequent out-of-sample period. Strategies passing this test historically survive live trading about 60% of the time, versus 5% for simple backtests.

Look-ahead bias appears as negative shifts like shift(-1), using today's index constituents historically, or trading on same-day close data. Audit every data point by asking whether the information was actually available at trade time.

The gap usually comes from transaction costs, market impact, and overfitting. Realistic costs range from 3.5 bps for SPY to 80+ bps for crypto, and in-sample Sharpe ratios typically degrade by 50% or more in live trading.

Run a cointegration test requiring p-value below 0.05, then calculate the spread's half-life using OLS regression. Pairs with half-lives of 10-40 days that remain cointegrated out-of-sample historically perform best.

Avoid ML when working with low signal-to-noise data, limited samples, or short histories like 3 years of daily data. Simple regularized models with walk-forward validation outperform deep learning on price data in most cases.