Time Decay in Football Prediction Models Explained

Time decay in football prediction models is a method for giving recent matches more influence than older matches. Instead of treating a result from last week and one from last year as equally informative, the model reduces an observation’s weight as it ages.

The speed of that reduction is often described through a half-life: the time required for an observation’s weight to fall to 50%. A short half-life makes a model more responsive to current changes but also more sensitive to noise. A long half-life preserves sample size but may leave outdated performances with too much influence.

There is no universally correct half-life. It should be selected with chronological, out-of-sample testing and treated as a model assumption rather than a definition of “true form”.

What Time Decay Means in a Football Prediction Model

A football model usually estimates team strength from historical observations such as goals, expected goals, shots, results or player availability. Time decay changes how much each observation contributes to that estimate.

Without decay, every eligible match might receive a weight of 1.00. With decay, the newest match may retain almost all its weight while increasingly old matches receive progressively smaller values.

This is more systematic than selecting a team’s last five matches and calling them form. A decay function can use the full available history while allowing influence to decline smoothly. It can therefore be added to a transparent baseline when analysts build a simple football prediction model.

The idea has a long history in football modelling. The Dixon–Coles football-score model addressed the changing nature of team performance by applying time-dependent weights within its likelihood calculation.

How Exponential Recency Weighting Works

A common approach is exponential decay expressed through a half-life:

w_i = 2^(-t_i / H)

Where:

• w_i is the weight assigned to observation i.
• t_i is the observation’s age, measured in days, weeks or another consistent unit.
• H is the chosen half-life in the same unit.

The interpretation is direct:

For a weighted average of a metric y, the estimate becomes:

Weighted estimate = Σ(w_i × y_i) / Σw_i

In a statistical model, the same weights can instead multiply each observation’s contribution to the loss function or likelihood. The model is then penalised more for fitting recent matches poorly than for fitting distant matches poorly.

A Worked Before-and-After Forecast Example

Consider a hypothetical home team with eight attacking-performance observations. The values below are illustrative expected-goals figures, not data from a real club. A 56-day half-life is used solely to demonstrate the calculation.

The unweighted mean is 1.45. After applying the example weights, the weighted estimate is approximately 1.60 because the stronger recent performances receive more influence.

Suppose a basic independent Poisson model holds the opponent’s expected-goals estimate at 1.10. The forecast changes as follows:

These percentages are model estimates, not observed facts. The example shows the mechanism rather than proving that the shorter-memory forecast is better. If the recent improvement is genuine, decay may help. If it is mainly variance, injuries to opponents or an unusually easy schedule, the revised estimate may be too aggressive.

How to Choose a Half-Life With Walk-Forward Testing

A half-life should not be selected because its output looks convincing for one team. It is a hyperparameter that must be tested on forecasts made with information that would genuinely have been available at the time.

Time-series cross-validation, also called rolling-origin or walk-forward evaluation, preserves chronology. Each training set contains only observations that occurred before its test matches.

1. Define the target. Decide whether the model predicts scorelines, goals, 1X2 probabilities or another clearly specified outcome.
2. Choose candidate half-lives. For example, compare 28, 56, 84 and 112 days with a no-decay benchmark. These are test candidates, not universal recommendations.
3. Set the first forecast date. Fit the model using only matches completed before that point.
4. Recalculate every observation’s age. Weights must be measured relative to the historical forecast date, not today.
5. Generate genuine forecasts. Predict the next match or block of matches without using future results, revised data or later team information.
6. Roll the origin forward. Add newly available matches, refit the model and repeat the process.
7. Compare probability quality. Use an appropriate proper scoring rule such as log loss or Brier score, while also checking calibration, stability and performance by season or competition.
8. Lock the choice before final testing. The final test period must not be repeatedly reused to tune the half-life.

The best-performing candidate should also beat or meaningfully complement the no-decay model. A small historical improvement that disappears across seasons, competitions or plausible implementation changes is weak evidence for a durable advantage.

Time Decay, Rolling Windows and Recent Form Are Different

Time decay is therefore not a synonym for form. It is a mathematical assumption about how quickly historical evidence loses relevance.

Practical Implementation Decisions

Use calendar time or football time consistently

Days measure how long information has aged in the real world. Matches or rounds measure how many new observations have arrived. Those choices are not equivalent when clubs have postponements, international breaks or different fixture congestion.

Calendar time is often appropriate when the concern is tactical or personnel change. Match count may be useful when learning depends mainly on new competitive observations. The choice should be tested rather than mixed accidentally.

Preserve stable base information

Recent evidence should usually update rather than erase sensible base rates in football analysis. League scoring levels, long-term home advantage and promoted-team priors may need slower decay than volatile team-specific features.

Consider different speeds for different inputs

A player’s current availability can change within days, while a competition’s scoring environment may change over seasons. Applying one half-life to every variable is convenient but may be poorly matched to the underlying processes.

Multiple decay rates create additional flexibility, however, and every extra parameter increases the opportunity to fit historical noise.

Track the effective sample

A model may contain hundreds of historical matches but behave as if it has far fewer once aggressive weights are applied. A common diagnostic is:

Effective sample size = (Σw_i)² / Σ(w_i²)

This does not capture every dependency in football data, but it helps reveal when a responsive model is relying on a thin amount of effective information. The wider principles are covered in GoalIQAI’s guide to sample size in football analytics.

Limitations and Common Errors

• Tuning on the final test set: repeatedly choosing the decay rate that performs best on the same period turns the test set into training data.
• Using future information: revised datasets, closing prices, confirmed line-ups or later matches can leak into historical forecasts if point-in-time controls are weak.
• Double-counting recency: a model may use decayed observations and recent-form features derived from the same matches, giving the newest data more influence than intended.
• Confusing change with noise: several strong or weak matches do not necessarily establish a new team level.
• Ignoring opponent quality: recent performances against unusually strong or weak opponents can distort a recency-weighted estimate.
• Assuming a manager change requires a reset: a new coach can matter, but squad quality, inherited tactics and implementation time do not disappear instantly.
• Choosing a half-life by intuition alone: a plausible football story cannot replace chronological validation.
• Reporting only average accuracy: a decay setting may appear acceptable overall while failing badly in particular seasons, leagues or promoted-team cases.
• Treating weights as probabilities: a match weight describes influence on estimation. It is not the probability that the match remains relevant.

GoalIQAI Interpretation: Time Decay Is a Bias–Variance Choice

Time decay is best understood as a trade-off between stale information and unstable estimation.

• A short half-life reduces dependence on old regimes but increases sensitivity to recent randomness.
• A long half-life produces steadier estimates but may underreact to transfers, tactical changes, injuries or shifting competitive strength.

The correct question is not “How much recent form should count?” It is “Which decay structure produces the most reliable future probabilities under a leak-free evaluation process?”

Sensitivity also matters. If moving from a 56-day to an 84-day half-life changes a match probability substantially, the forecast is dependent on an uncertain modelling choice. That should reduce confidence rather than encourage false precision.

A broad experimental review of score-based football models found similar overall performance among several leading methods. Time decay should therefore be treated as one useful component of a complete model, not as an automatic source of predictive superiority.

Key Takeaways

• Time decay gives recent football observations more influence than older observations.
• Half-life is the time required for an observation’s weight to fall to 50%.
• Shorter half-lives improve responsiveness but reduce the effective sample and can amplify noise.
• Exponential weighting uses old data gradually rather than discarding it at an arbitrary rolling-window boundary.
• The decay rate should be chosen through chronological walk-forward validation, with no future information entering historical forecasts.
• A no-decay model remains an essential benchmark.
• Large forecast changes across plausible half-lives indicate model uncertainty, not extra confidence.

Related Guides

• Backtesting a Football Betting Model Explained
• How Professional Bettors Validate Their Models
• Football Betting and Analytics Knowledge Base

Stay Ahead of the Market

Subscribe to GoalIQAI for evidence-based football analytics, probability education and practical guides to building and evaluating better prediction models.