Understand backtest results

Language: English · 简体中文

The default run_backtest result is a five-element tuple:

stats, net_returns, gross_returns, turnover, periods = result

Summary statistics

stats commonly includes total return, annualized return, annualized volatility, Sharpe ratio, maximum drawdown, and average turnover. Use the current implementation and tests as the authority for exact fields. A common access pattern is:

summary = {
    "total_return": stats["total_return"],
    "annualized_return": stats["ann_return"],
    "annualized_volatility": stats["ann_vol"],
    "sharpe": stats["sharpe"],
    "max_drawdown": stats["max_drawdown"],
}

A metric may be None or NaN when there is not enough data to calculate it. Check the sample length and metric definition before comparing experiments.

Net and gross returns

• gross_returns reflects returns along the portfolio path before modeled costs.
• net_returns uses the same path after modeled transaction costs and slippage.
• Their difference helps show whether turnover and execution assumptions materially reduce returns.

For example:

cost_drag = gross_returns - net_returns

This cost drag is a model result, not an estimate of actual fills. Actual execution requires evidence such as orders, fills, and account ledgers.

Holding-period details

periods is a list of dictionaries organized by holding period. Common fields include:

Field	Meaning
entry_date	Start date of the holding period
exit_date	End date of the holding period
net_return	Net return for the holding period
gross_return	Gross return for the holding period
turnover	Turnover for the holding period
positions	Target positions used for the holding period

Backends and configurations may add fields. For the full output contract, see backtest outputs and position outputs.

Ask three questions before interpreting results

1. Did the return come from the portfolio path, or from omitted costs and pricing assumptions?
2. Were the signal, price, and tradability inputs available at the time represented by the backtest?
3. Is this a diagnostic result, or is it supported by complete execution evidence?

Backtest metrics alone cannot answer these questions. Review the input data, configuration, position artifacts, and execution evidence together.