Trade-Level Versus Portfolio-Level Metrics

Two research reports can describe the same backtest and disagree about whether it's good. One averages statistics over individual trades: average return per trade, hit rate, average holding period. The other looks at the equity curve produced when all the trades are sized and combined into one book: Sharpe ratio, maximum drawdown, volatility. Both are legitimate — and they can point in opposite directions, because sizing, timing overlap and correlation between trades live entirely in the second view and are invisible in the first.

Why the two disagree

A trade-level metric treats every trade as one data point, regardless of how large it was or when it happened. A portfolio-level metric looks at the realized dollar (or percentage) P&L path once trades are combined with their actual sizes and actual timing. The gap between them opens up in three common ways: if bigger trades tend to be the losers, average trade return can look fine while the equity curve is ugly; if many trades are open simultaneously, portfolio-level volatility depends on how correlated they are, something a trade-by-trade average cannot see at all; and if a strategy trades rarely, trade-level statistics can be noisy — one lucky or unlucky trade dominates a small sample — while the smoother portfolio curve better reflects a full backtest period.

Worked example: same trades, opposite sizing

A strategy places four trades in a quarter with returns of $+2\%, +2\%, +2\%, -30\%$. Average trade return is $(2+2+2-30)/4 = -6\%$ — a poor trade-level number. But suppose position sizing put only 5% of the book on the losing trade (a name the strategy was already trimming) and 25% each on the three winners. Portfolio-level contribution is $0.25(2) + 0.25(2) + 0.25(2) - 0.05(30) = 0.5+0.5+0.5-1.5 = 0\%$ — flat rather than sharply negative. The trade-level average understates the outcome because it implicitly weights every trade equally; the portfolio-level number reflects what the book actually earned.

The reverse also happens: a strategy with a strong 60% average trade return can produce a mediocre Sharpe ratio if its biggest trades cluster on the same days and move together, so portfolio volatility is far higher than the trade-level statistics would suggest.

What this means in practice

Read both levels together, never one alone. Trade-level statistics are the right tool for diagnosing the signal itself — is the underlying edge real, trade by trade, independent of how it happens to be sized right now? Portfolio-level statistics are the right tool for judging whether the strategy is investable — what an allocator's capital would actually have experienced, including correlation and sizing decisions. A common failure mode is a research team optimizing trade-level hit rate while sizing decisions (made downstream, or by a different desk) quietly erode the portfolio-level outcome, with nobody looking at both together to catch it.