Why a 90% Hit Rate Can Still Lose Money

One number can hide two very different questions

A sports forecast can ask how often an event occurs. A financial decision asks whether the price offered for that event compensates for its risk. These are related, but not identical, questions.

“Nine correct predictions out of ten” sounds strong. If each successful selection pays so little that the occasional loss consumes the accumulated gains, even a genuinely 90%-accurate strategy can lose money.

The break-even calculation

For a simple decimal-odds outcome with a full win or full loss, expected return on one unit staked is:

Suppose the true win probability is exactly 90%. The break-even decimal price is 1 ÷ 0.90 ≈ 1.1111. Below that price, the expected financial return is negative. Above it, the arithmetic becomes positive — but only if the probability estimate is accurate and the price is actually available.

Hypothetical arithmetic only: no historical bookmaker prices or verified AIONORA model probabilities are implied.

Even positive theoretical expected value is not the same as a dependable monthly profit. Real outcomes vary, and estimation errors can erase small projected advantages.

A concrete example: 90 wins, 10 losses

Imagine 100 equal one-unit wagers, each at odds of 1.07. Exactly 90 succeed and 10 fail. The observed hit rate is 90%.

The 100-unit total stake returns only 96.30 units. The outcome is a −3.7% realised return on turnover, despite a 90% hit rate. Real outcomes will not always equal the assumed percentage, but this illustration shows why counting wins without recording prices is incomplete.

The reverse matters too: a lower hit rate at sufficiently higher prices may be economically stronger. Comparing strategies by headline hit percentage alone can reward the wrong method.

Three reasons a backtest can look better than reality

1. Favourable events are not necessarily valuable prices

Goal-line rules such as “over 0.5 goals” may be correct very frequently. But their prices can be so short that high frequency provides no durable edge. In AIONORA's retrospective football research, simpler high-occurrence goal markets produced strong historical hit percentages without establishing profitable executable odds.

2. Missing results can bias a sample

Our archive audit identified a data-quality problem: some 0–0 football results were encoded differently from other scores and were excluded by an earlier parser. Omitting those matches artificially improves the measured success of bets that require at least one goal. After identifying that issue, the retrospective calculations had to be reassessed.

That is why reproducible ingestion rules, duplicate checks and explicit handling of score formats matter just as much as sophisticated forecasting models.

3. Selecting the best-looking rule after the fact is not forecasting

It is easy to search hundreds of historical patterns and show the most flattering one. But picking a rule using the results it is later evaluated on leaks information from the future. Proper evaluation locks the rule before testing it on an untouched later period, then checks it again on genuinely new events.

What better research should report

A meaningful evaluation should include the exact market and its line, decimal price and capture time, estimated win probability, break-even probability, alternative independent estimates, sample coverage, uncertainty, historical calibration, and whether the original forecast existed before the event.

When enough prospective observations exist, proper scoring rules such as Brier score and log loss can assess probability quality. Closing-line comparisons may supply another diagnostic, but neither scoring improvements nor a lucky winning streak guarantee future returns.

For AIONORA, no certified profit or model calibration is claimed until the necessary independent prospective evidence is actually available. We deliberately publish reasons to reject uncertain research hypotheses as well as reasons to investigate them.

To check forecast quality yourself, explore our interactive Brier Score calculator. For more on probability interpretation, see our model probability research guide. For dated research editions and their limitations, consult the research archive.