Reading Backtest Metrics: Which Numbers Actually Matter for a Trading Bot

Anyone evaluating backtest metrics gets almost the same set every time: net profit, win rate, profit factor, maximum drawdown. Alongside them circulate rules of thumb for when a strategy counts as good. Those thresholds are folklore with no firm foundation. What matters more is understanding what each number actually measures - and where it misleads for an automated strategy.

Win rate: meaningless on its own

A high win rate feels good, but says nothing until you know the ratio of average win to average loss. A strategy that wins seven trades out of ten but gives back three times as much on each loss is losing money. One with four winners out of ten can be profitable. The meaningful number is the expectancy per trade, which combines both.

Profit factor: compact, but sensitive

Profit factor divides the sum of all wins by the sum of all losses. It's compact and easy to compare, but it has two weaknesses. With few trades it swings wildly - a single large winner can push it up dramatically. And it's sensitive to costs: if the average profit per trade sits only just above execution costs, a solid value can drop below one with just a little more slippage.

Maximum drawdown: a floor, not a ceiling

Maximum drawdown is the largest decline from a peak in the equity curve to the subsequent trough. It's often read as the worst the strategy can do. In fact it's only the worst that happened on this one historical path. The same trades in a different order could have produced a deeper decline, and the future follows no particular order. It makes more sense to treat the historical value as a minimum and plan risk so that a considerably deeper decline remains bearable.

The metrics hardly anyone looks at

• Number of trades: few trades make every other metric uncertain, however good it looks
• Longest losing stretch, in trades and in calendar time: weeks without a new high often decide whether you stick with it more than the depth of the drawdown does
• Distribution over time: does the result come evenly from the whole period, or almost entirely from a single market phase?
• Time in the market: how long positions stay open and exposed to risks the backtest doesn't model, such as overnight gaps

Three cross-checks before you trust a report

• Remove the best trades: is the strategy still profitable without its five biggest winners? If the result hinges on a few outliers, it's hardly repeatable
• Raise the costs: what does the result look like with the assumed cost per trade doubled? A strategy that can't survive that has little room for reality
• Split the period: do both halves of the test period paint a similar picture, or does the overall result live off one of them?

For why a result that's too good can itself be a warning sign, see the article on overfitting in trading strategies.

For how realistic execution costs feed into the evaluation, see the article on slippage in a trading bot.

No single metric answers whether a strategy is any good. Together with the cross-checks, though, they show fairly reliably whether a good result rests on a broad base - or on a few lucky trades and optimistic assumptions.