backtesting-frameworks

Use a vectorized backtester with pandas for fast signal-based strategies, or an event-driven backtester for realistic order and fill simulation. Always shift signals by one bar to avoid look-ahead bias and deduct commission plus slippage from returns.

Walk-forward analysis splits historical data into sequential train and test windows, optimizing parameters on each training window and validating on the following unseen test window. It prevents overfitting better than a single train/test split by simulating realistic re-optimization over time.

Avoid look-ahead bias by using point-in-time data and shifting signals so trades execute on the bar after the signal is generated. In vectorized backtests, apply .shift(1) to signals before multiplying by returns.

Vectorized backtesting is fast and suits simple signal-based strategies on a single instrument. Event-driven backtesting is slower but models orders, fills, positions, and cash accurately, making it better for multi-asset portfolios and complex order types.

Common causes include overfitting parameters to historical data, ignoring transaction costs and slippage, and survivorship bias from testing only on securities that still exist. Use out-of-sample testing, Monte Carlo simulation, and realistic cost models to detect these issues.

Monte Carlo drawdown analysis bootstraps historical returns by resampling them with replacement across thousands of simulations, then computes the maximum drawdown of each simulated equity curve. The result is a distribution showing expected, median, and worst-case drawdowns.