Average winners and losers: what a hit rate leaves out

How often is half the question

A hit rate says how often a rule made money. It says nothing about how much it made when it was right, or lost when it was wrong. A rule that wins nine times in ten can still lose money over time if the tenth loss is large enough, and a rule that wins less than half the time can come out ahead on a few large gains.

The average trade already contains both parts. It equals the share of winning trades times the average winner, plus the share of losing trades times the average loser. Splitting it back into those parts shows what the average hides.

Bigger losses than gains, on both rules

Take two once-a-year SPY rules: the 15 sessions from October 31 (2000–2025) and the November window (1999–2025). Each year is one trade, so trades and years are the same count. October 31 rose in 18 of 26 years and November in 23 of 27. Their losing years were the larger ones.

November's average winner, +3.79%, is close to October 31's, +3.45%, and its average loser is a little smaller. The real difference is the count of losing years. With 4 losses against 23 gains, November's gains summed to 4.30× its losses; October 31's 8 losses brought its ratio down to 1.40×.

Two ratios, and what they assume

Payoff ratio = the average winning observation ÷ the absolute average losing observation. Below one, the typical loss is bigger than the typical gain, and the rule needs to win more often than it loses just to break even.

Sum ratio = the sum of the winning observations ÷ the absolute sum of the losing ones. It folds in how often each side happened. On gross trade returns it reads like a profit factor, but only under one assumption: every trade starts with the same amount of money. With compounding or changing position sizes, a sum of percentage returns is not what an account would have earned.

Zero observations count as neither a win nor a loss. These two rules have none, in gross returns or in excess. When a rule has no losing observation at all, both ratios are undefined and are shown as not available, never as a large number.

Gross returns and excess answer different questions, so they are kept on separate rows. A trade can make money and still trail the same-length hold, which is why October 31 has more gross winners than excess winners.

A few losing years carry the average loser

Both rules' worst trade came in 2008: −18.53% for October 31 and −9.58% for November. Without that one year, October 31's other 7 losing trades averaged −3.71%, and November's other 3 averaged −3.57%.

That is the practical weakness of these ratios on a once-a-year rule: the average loser rests on 8 and 4 observations. One more bad year, or one fewer, moves it a long way. Removing a year here is a sensitivity check, not a way to avoid it: the year happened, and a trader holding the rule then would have taken the loss.

Why these numbers do not size a position

Formulas such as the Kelly criterion turn a win rate and a payoff ratio into a fraction of capital to risk. Fed with these figures, they would treat two dozen historical years as if they were the true odds of every future year, and would know nothing about how far a trade fell before it closed: these are entry-to-exit returns, and the loss inside a window is not measured.

The ratios are descriptive. They show the shape of what happened, in gross returns and against a holding benchmark that is not a trade anyone could place. They are not a statistical test and not a guide to how much to put on.

Reading winners and losers

• Read the hit rate together with the average winner and the average loser.
• Check which is bigger. Below a payoff ratio of one, the rule depends on winning often.
• Count the losing observations. A handful of them cannot pin down an average loser.
• Keep gross returns and excess over holding apart.
• Look at the worst observation and the year it happened, not only the averages.

The October 31 research and the November research list every trade. Two related guides: win rate against excess over holding and mean against median.