Produce one interpretation that keeps every metric attached to its basis.
By the end, you should be able to state which sample was measured, what P/L unit and cost convention were used, how many trades contributed, what the path looked like, and why no single statistic is carrying the entire conclusion.
Do not compare values until the measurement basis matches.
- Sample: full history, development sample, and holdout answer different evidence questions.
- Unit: money, percent, pips, or another basis changes scale.
- Costs: spread, commission, slippage, and other charges change trade P/L.
- Trade count: a small sample can be dominated by a few observations.
- Selection history: a best-of-many result is not equivalent to a prespecified single test.
Ask one question of each statistic.
When the metrics disagree, investigate the structure instead of voting.
Positive aggregate P/L can coexist with weak trade average, deep drawdown, concentration in a few trades, or poor holdout behavior. A ratio can look attractive because its denominator is unusually small. Conflicting statistics are often a prompt to inspect the underlying trades and path, not a reason to average the verdicts.
Ask whether the metric itself is stable across runs, periods, or perturbations.
A single Profit Factor, Expectancy, or Return/DD value hides how that statistic changes under different samples or plausible perturbations. When comparable repeated results exist, inspect their distribution rather than reporting only the best or average value.
Metric shopping can make almost any backtest look better than it is.
- Comparing Net Profit across different capital, sizing, currency, or cost conventions.
- Treating Profit Factor above 1 as a robustness verdict.
- Ignoring how a small loss denominator can inflate Profit Factor or a small drawdown denominator can inflate Return/DD.
- Using positive Expectancy without checking trade count or the influence of unusually large trades.
- Reporting maximum drawdown without duration, recovery, or equity-path context.
- Mixing full-sample and out-of-sample statistics without labeling the sample.
- Selecting thresholds after inspecting many candidates or many metrics.
Before copying a metric into a report, carry these fields with it.
- Metric name and exact definition used by the software that reported it.
- Sample and time range.
- Unit and initial-capital basis where relevant.
- Cost and position-sizing assumptions.
- Trade count or observation count.
- Candidate/selection context.
- Path or distribution evidence needed to interpret the summary.