Overfitting: We Tested 168 Betting Rules. The Best Ones Lost Next Month
Try enough rules and one looks brilliant by accident. We trained 168 simple rules on August–September and tested the winners on October. Four of the five best lost money.
Last reviewed November 3, 2026
Key takeaways
- >168 simple rules on six leagues, trained on matches before 1 October 2026
- >Not one of the 64 judged rules beat its luck band in training
- >The five best in training returned −9.9% in October; four of the five lost
- >Of 20 rules with no edge, about one lands outside its luck band by chance, half of those looking like winners
- >Count your tries, keep data back, prefer a reason over a result, then paper-bet
Try enough betting rules on the same matches and one of them will look brilliant. Not because it found something, but because you tried enough. This is called overfitting, and it is the single most common way a backtest lies. We ran the experiment on our own data to show how big the effect is.
The experiment
We took six leagues (Eredivisie, Premier League, La Liga, Bundesliga, Serie A, Ligue 1) and defined simple rules: back the home win, the draw, the away win, the over, the under, or both-teams-to-score yes or no, at one of four price ranges (1.01–1.50, 1.50–2.00, 2.00–3.00, 3.00–6.00), one bet per match, €10 flat, Toto's prices before kickoff.
That is 168 rules. We then did what anyone hunting for a strategy does:
- "Train": run every rule on the matches played before 1 October 2026.
- Pick the winners: the rules with the best ROI (and at least 30 bets).
- "Test": see what those winners did on the matches played from 1 October to 31 October.
The results
In training, 64 of the 168 rules had enough bets to judge. 24 of them were in profit, and not a single one beat its luck band. Six lost by more than luck explains.
Then the five best rules from training, in the month that followed:
Rule | Training (before 1 Oct) | Test (October) |
|---|---|---|
Bundesliga, over at 2.00–3.00 | +31.5% (33 bets) | −25.1% (29 bets) |
Eredivisie, over at 2.00–3.00 | +25.2% (56 bets) | +14.9% (26 bets) |
La Liga, over at 1.50–2.00 | +23.3% (45 bets) | −4.2% (23 bets) |
Premier League, draw at 3.00–6.00 | +21.7% (47 bets) | −13.6% (37 bets) |
Eredivisie, over at 1.50–2.00 | +19.1% (46 bets) | −19.0% (22 bets) |
Four of the five "best strategies" lost money in the next month. Together, the top five returned −9.9% in October, worse than all the judged rules together (−5.6%), which is roughly what you would expect from betting at random against the bookmaker's margin.
Being the best in training predicted nothing.
Why this happens
Each rule's result is partly signal (if any) and partly noise. When you sort 64 rules by their result, the ones at the top are mostly the ones with the luckiest noise. Their luck does not carry over to the next month, so they fall back.
The rule of thumb: of 20 rules with no edge at all, about one lands outside its luck band by chance, and half of those on the winning side. Try 100 and two or three will look like winners. "I found a strategy that beats the luck test" means very little unless you also say how many you tried.
How to protect yourself
- Count your tries. Our backtester counts how many settings you have tried in a session and shows it next to the result. If you tried 40, expect one or two flukes that look like winners.
- Keep data back. Find the rule on one period, then judge it only on a period you did not look at while searching. That is what the "test" month above did.
- Prefer a reason over a result. A rule with a plausible cause (the league started the season scoring far more than last season) can be checked against that cause. A rule that is "just in the numbers" usually is not.
- Then paper-bet it on matches that have not been played yet. See Paper betting.
The honest takeaway
The interesting result in this experiment is not a strategy. It is that, across 168 simple rules on six leagues, none showed an edge beyond luck, and the best-looking ones did worst afterwards. That is what a fair market looks like, and it is the background every claimed "system" should be measured against.
Method: Edge Terminal backtester, 6 leagues × 7 selections × 4 price ranges = 168 rules, one bet per match, €10 flat, Toto prices before kickoff (backfilled closing prices where Toto had none). Training: matches before 1 October 2026; test: 1–31 October 2026. Judged: rules with ≥30 training bets. Luck band: ±2 standard errors of the per-bet return. Run 3 November 2026. Past results only; nothing here is a tip.