Backtests are the most dangerous sales tool in sports betting. A model that wins beautifully on old data can lose steadily on new games, and the reason is usually overfitting: the model memorised the past instead of learning patterns that repeat. Spotting overfitting before you trust a model is a skill, and it starts with knowing the warning signs.

The clearest warning sign is a large gap between training and test performance. A model that wins on 70 percent of its training games but only 50 percent of unseen games is not discovering an edge; it is remembering answers. A healthy model shows a smaller, stable gap, because real betting data is noisy and no genuine pattern wins everywhere.

Perfect or near-perfect training scores are another red flag. Real sports data contains injuries, weather, referee decisions, and sheer luck, so a model claiming near-perfect accuracy on training data has almost certainly fitted the noise rather than the signal. Treat any claim of extremely high accuracy with scepticism, and ask how the model performs on data it has never seen.

Data leakage is overfitting's sneakier cousin. It happens when a model is fed information that would not have been available at betting time, such as closing odds when the bet was supposedly placed earlier, or statistics from the very match being predicted. The classic test is simple: at the exact moment the bet would have been placed, could the model have known this number? If not, the backtest is fiction.

Model complexity relative to data size is another warning. A model with dozens of features trained on a few hundred matches can find apparent patterns in almost anything, including patterns that are pure coincidence. Simpler models with a handful of well-chosen features, such as recent form, home advantage, and injuries, are usually more robust than complex models that use every statistic available.

Cherry-picked time windows are a subtler trap. A model tested only on one season, one league, or one team's winning run may look strong while failing everywhere else. Demand results across multiple seasons, leagues, and market conditions, and be wary of any vendor who only shows the best window.

Repeated testing on the same data quietly destroys independence. Every time a modeller tweaks a model after seeing test results, the test set stops being a fair judge. The fix is to keep a truly held-out set that is only checked once, or better, to test on genuinely future games after the model is frozen.

Time order matters more in betting than in most machine learning. Randomly shuffling matches into training and test sets lets future games inform predictions about past ones, which never happens in real betting. Proper evaluation trains on earlier matches and tests on later ones, in chronological order, because that mirrors how the model will actually be used.

Protecting yourself comes down to a short checklist. Ask for out-of-sample results on unseen, chronological data. Ask what features the model uses and whether each was available at betting time. Ask how many times the model was adjusted after seeing test results. Prefer simple models, regularisation techniques that penalise complexity, and results reported across many seasons rather than one glowing window.

No model beats the market reliably without surviving these checks, and most that fail them were never edges at all, only memories of a past that will not repeat. Treat a backtest as a claim to be verified, not as evidence, and never stake money you cannot afford to lose on a model whose testing you have not examined. Please gamble responsibly.