Do Weather and Fixture Congestion Improve Football Predictions?

Does rain improve match-result predictions?

We studied Premier League results from 2019–20 onward using two separate weather tracks. ERA5 reanalysis describes weather retrospectively; it is useful for an explanatory test but cannot be used as if it were a pre-match forecast. The GFS archive supplies nominal 48-hour forecasts, covering a shorter period from 2024 onward.

Four additions were specified in each track: precipitation, wind and gusts, temperature and humidity, and all weather signals together. Precipitation combines the six hours before kickoff, the first two match hours and an indicator for an overlapping hour with at least 2 mm. It includes rain, showers and snow, rather than measuring rain alone.

The forecast precipitation addition improved overall test log loss by 0.004277, but its adjusted interval includes zero. Its improvement in the later half was only 0.001780, below our required 0.005. The all-weather addition was selected on tuning in both tracks; neither selected model passed.

We also tested over 2.5 goals and both teams to score. No addition passed for either target. The full-time result, goals and BTTS tests form one family of 24 comparisons; finding one promising number among them is insufficient.

Do short rest and consecutive away games help?

The schedule study used 4,894 test fixtures across six league-specific samples, from 2019–20 onward. We tested differences in rest days, 48- and 72-hour turnarounds, league games in the preceding 7 and 14 days, consecutive away fixtures, short rest during an away sequence, and a combined adjustment.

The combined adjustment was selected on tuning in all six leagues. Five overall test scores worsened; the Bundesliga difference was almost zero. None passed the statistical gate.

Of 4,894 test fixtures, 4,774 had prior same-season league history for both clubs. There were 393 fixtures with a measured turnaround of at most 72 hours, 819 second-or-later away fixtures, and only 35 third-or-later away fixtures. These categories overlap.

This did not test literal next-day back-to-back games. There were 0 measured turnarounds of at most 48 hours in the test sample. Only earlier kickoffs strictly before the forecast's 24-hour cutoff enter the history, and cup matches are absent. An away sequence is a sequence of league venues, not proof that players stayed away from home.

How we tested the additions

The control is a fixed Dixon–Coles goals model, with a 365-day rating half-life and shrinkage setting of 4. Forecasts are made as of 24 hours before kickoff and team ratings are refitted every 28 days. The control excludes player-availability adjustments and is not a replay of every historical production setting.

Each sample is split chronologically: roughly 40% for tuning, 30% for calibrating confidence and 30% for testing. Unavailable match outcomes are purged at the stage boundaries. Addition weights are fitted on tuning only; control and candidate confidence are calibrated separately. The final block is evaluated without choosing settings from its results.

Weather additions adjust goal rates, so we can test match results, over 2.5 goals and BTTS. Schedule additions adjust relative home/draw/away probabilities; this experiment does not test schedule effects on total goals. Missing same-season schedule history produces no adjustment.

Our adoption rule requires log loss to fall by more than 0.005 and Brier score to fall in each chronological test half. We also require the upper end of an uncertainty interval to be below zero. We resample whole weeks 20,000 times, with Bonferroni adjustment separately over 24 weather comparisons and 36 schedule comparisons. Lower scores are better; a log-loss improvement is not a percentage-point increase in accuracy or evidence of betting profit.

The two experiment families share football matches. Weather test tracks also overlap. Their match counts cannot be added to describe independent evidence, and all results are retrospective on historical seasons studied previously.

What we cannot conclude

• Heavy-weather evidence is limited. Only 7 reanalysis test fixtures and 4 forecast test fixtures met the hourly precipitation threshold. The grids do not measure pitch drainage, roofs, surface wetness or a local shower directly.
• Real fatigue is incompletely measured. Domestic cups, European matches, national-team duty, extra time, travel distance and individual player minutes are missing. A league-only rest gap can overstate actual recovery time.
• Historical availability is not fully auditable. The forecast archive does not return the original publication and ingestion timestamps used by this study. Fixture records contain final played kickoff times, without historical announcement or postponement timestamps. These gaps independently prevent live adoption, even if a future retrospective score looks better.
• Association is not causation. These tests ask whether specified signals improve this control on these samples. They cannot establish that rain never changes matches, that fatigue is irrelevant, or that a different feature design would fail.

Decision: keep the published model unchanged. Better all-competition schedules, player workload data and forecasts saved before kickoff would make a stronger next experiment. A fresh prospective sample is needed before treating an apparent improvement as confirmed.

Data, sources and citation

Download all 60 comparisons as CSV, including exact log-loss and Brier differences, adjusted intervals, sample dates and tuning selections. Tables round values to six decimals; the CSV retains the stored precision.

Weather data: Open-Meteo, using ECMWF ERA5 reanalysis and NOAA GFS forecasts, under CC BY 4.0. See the historical weather and previous runs documentation. Stadium coordinates come from the linked Wikipedia sources in our venue register; Everton's ground change is handled by date. Match-level neutral venues are not verified.

Football scores and kickoff records use this project's historical football-data.co.uk, football-data.org and OpenFootball archives. These sources are combined with the coverage rules described in the model methodology.

Audit the weather report, schedule report and experiment code. Reports retain source and snapshot hashes. Reproducing the full run requires the matching archived inputs; raw API caches are not included in the repository.

Cite: Football Edge Lab (2026). Do Weather and Fixture Congestion Improve Football Predictions? Experiments weather-v1 and schedule-v1, model dc-v4. https://footballedgelab.com/research/weather-and-fixture-congestion.