Why is beating football odds so hard?

The hypothesis: fixture “echoes”

Cortenscore's Link Scanner checks whether the run of fixtures around a match repeats the run around an earlier match that ended in a surprise. For example, if a team's opponents two and three weeks before the target match are the same as the opponents before one of its surprises in an earlier season, we call it a sequence echo. A single repeated opponent is a single echo, a repeat of the opponent's previous score is an opponent score echo, and the previous and next opponents repeating together is a sandwich echo.

“Surprise” means one thing here: a HT/FT comeback, where the team losing at half-time wins the match (1/2 or 2/1). These results happen in roughly one match in twenty and are priced high. See HT/FT explained for the market itself.

The idea came from examples our users spotted in screenshots. The question was simple: in matches with these echoes, do comebacks happen more often than the market prices them?

Pilot result: it looked promising

The measure was observed comebacks ÷ comebacks the market expected. The market expectation comes from several bookmakers' closing prices per match, with the margin removed. A ratio of 1.00 means the market priced it correctly; above 1 means more surprises than the market expected.

In the Turkish Süper Lig, the English Championship and the Premier League (2019–2026), matches with at least two different echo types showed a ratio of 1.40 [1.13, 1.69]. In matches without echoes the market was spot on: 1.00. The gap grew steadily with the number of echo types. On paper, it looked like a pattern the market had missed.

We were not convinced. Several groups and periods had been examined (multiple comparisons), the idea itself came from observed examples, we did not know at what time the archived prices were available, and a team's past tendency to produce surprises was a confounder. The honest way to separate these was to repeat the same test on data we had not seen.

Pre-registration: rules before results

Before looking at any result in the new leagues, we wrote down and froze the following:

• Primary hypothesis (H1): in matches with at least two echo types, the observed ÷ expected ratio is above 1.
• Decision rule: “replicated” requires the ratio to be above 1 and the lower end of the 95% interval to be above 1. A ratio of 1 or below means “not replicated”.
• League selection by coverage only: at least 900 matches since 2019, full HT/FT closing prices from at least three bookmakers in at least 80% of matches, and at least 90% of matches joined between data sets. The pilot leagues are excluded from the primary analysis.
• Not allowed: dropping leagues based on the result, changing thresholds or groups, or switching to opening prices and saying “it worked there”.

15 of 43 candidate leagues passed these rules. The 28 excluded leagues were dropped without looking at results; most failed the join threshold because our season archive does not cover some of their seasons.

Replication: it did not hold

2. Bundesliga, Bundesliga, A-League, Belgium, Ligue 1, Eredivisie, Portugal, Greece, the Scottish Championship and Premiership, LaLiga, LaLiga2, Switzerland, Serie A and Serie B: 30,954 matches in total. Results at closing prices:

The primary test gave 0.95; under the rule written in advance, the result is not replicated. The same test at opening prices gave 0.96 [0.83, 1.09]. Because some of the ideas came from Serie A examples, a sensitivity test without Serie A (added before the run) did not change the result: 0.94 [0.80, 1.07]. League-level ratios ranged from 0.60 to 1.50; no single league's interval sat clearly above 1.

Comeback prices also carry a heavy margin: even where the market probability is right, every price includes that share.

What we learned

1. The market prices this outcome correctly overall. Across 30,954 matches the observed comeback rate was 4.95% and the market expected 4.96%: a ratio of 1.00.
2. The 1.40× in the pilot was most likely luck, or specific to those three leagues. It is easy to notice the most striking of many groups and periods; an effect that does not repeat on new data is not evidence.
3. The small excess in echo matches fits a confounder. An echo only forms if the team has past surprises, so surprise-prone teams collect more echoes. That is not a difference the market missed.
4. This is not an isolated case. In the same research line we also tried adding team strength (Elo, Dixon–Coles), schedule congestion and referee tendencies to the market consensus. None improved the forecast statistically, and adding team strength made the log-loss clearly worse. The market appears to price team strength already.

Impact on the product: the Link Scanner is not presented as a prediction tool and does not rank matches; every echo is shown with its own track record. This result supports that stance.

Limitations

• We do not know exactly when the archived prices were available, and some closing records we inspected were written after kick-off. The results are therefore a “historical association”, not a real-time test.
• The season archive and the odds records spell team names differently; names were matched only from fixture co-occurrence (same day ±1, mutual unique match), and leagues below the join threshold were excluded.
• This is the failure of one hypothesis; it does not prove that no method can beat football markets. It does give an honest starting point: a pattern claim should not be taken seriously until it has been tested with rules written in advance, on new data.

Method summary

• Market probability: the nine-way HT/FT closing price vector from at least three bookmakers per match; margin removed with the power method, bookmakers averaged with equal weight. We explain the margin in how to calculate bookmaker margin.
• Outcome label: half-time and full-time scores from a source independent of the archive used to compute echoes.
• Uncertainty: block bootstrap by calendar week, 5,000 resamples, fixed seed; square brackets are 95% intervals.
• Time leakage: every feature was computed only from matches on earlier days; a same-day result never fed another match.
• Data: raw match and price data are not shared because of licensing and terms of use; every number on this page is an aggregate. Questions about the method and corrections are welcome through the contact page.

Why sample size and confidence intervals matter so much: try the sample size guide and the sample interval calculator. Our general approach is on the methodology page.

Create account Membership and pricing (Turkish) ↗

Published 28 October 2026 · Last updated: 28 October 2026 · Responsible: Cortenscore editorial team · Spotted an error? Report a correction. This page is a statistics education and research summary; it is not a prediction or betting advice. Past results do not indicate future results.