Evidence

Random selection inside the same odds band, on the same fixtures, at the same best-of-bookmaker prices. It lands near zero rather than at a loss because shopping for the top price already strips out the bookmaker margin. The model finishes below that line, which means its selections are actively harmful rather than merely weak: it keeps backing the side the market prices as too short, and the market keeps being right. Every time this model 'finds value', what it has really found is its own largest error.

Brier score over every scored match, lower is better. This is the cleanest test there is: it uses all matches rather than only the ones the model liked. The closing price is the better forecaster. If the model cannot out-predict the market, no staking rule or filter will turn it into profit — the edge has to come from forecasting, not from bet selection.

Filters like “drop this league” or “cap the edge” are chosen by a human looking at results, which is exactly how a backtest gets flattered. So the sample is split by date: the filters were picked on the first half and are scored here on the second, which never influenced them. The gain survived, which is the encouraging direction — a rule that only works on the data that produced it is a description of the past. Note the out-of-sample interval still includes zero, so this is a promising direction, not a demonstrated edge.

Every bet split into ten equal groups by the edge the model claimed. The ranking is inverted: the model's *small* disagreements with the market pay, and its large ones lose. That is not a paradox — it follows from the market being the better forecaster. When this model says it is twenty points better than the closing line, the overwhelmingly likely explanation is that the model is wrong, not that the market is. A big claimed edge is a error signal, so capping edge from ABOVE is the counter-intuitive move that survives out-of-sample testing.

The same model and edge threshold, run over different price ranges. The band is not an arbitrary restriction — it is a filter on the model's own worst errors. Removing it entirely (1.01-99.00) drops ROI to -6.92% with an interval that excludes zero, i.e. a reliably losing strategy: at extreme prices the model backs long-odds home underdogs, which is exactly its known bias. A narrow band buys accuracy at the cost of sample size. Note this table is a diagnostic, not a tuning dial — picking the best-scoring row would be fitting the parameter to the test set.

Totals (Over/Under) and 1X2 are separate bets on the same model. Totals produce far more qualifying selections, because a two-way market prices closer to even money more often — which means a much tighter confidence interval for the same seasons of data. Neither market's interval clears zero, so neither has demonstrated an edge — but the totals interval is narrow enough that a real edge of a few percent would have shown up by now.

Each competition is modelled separately, so these are independent-ish samples of the same strategy. If profit comes from one league only, you have found a quirk of that league — or noise — rather than a method. Per-league samples are small; treat the ordering as indicative.

Bets grouped by the edge the model claimed at the time. If the edge number carries information, ROI should climb down this table. It does, which is weak evidence the estimate is meaningful.

Every model probability bucketed against how often that outcome actually happened. A calibrated model sits on the diagonal. Miscalibration matters more than accuracy here: a model that says 45% when the truth is 40% will invent edge on exactly the bets it feels most confident about.

The sample split into four equal blocks by date. A real edge is boring — similar returns throughout. The periods broadly agree, which is what you want to see.