A market-initialised accelerated failure time model for in-play football forecasting
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "A market-initialised accelerated failure time model for in-play football forecasting".
Jane: Here's the gist of the paper. In-play football forecasting models have struggled to match the accuracy of betting exchange prices, which aggregate information from many market participants The calibrated Weibull model exhibits a similar classification accuracy to Betfair across 140 evaluation matches, while retaining interpretable team-level parameters and covariate effects that the market does not provide Model Extensions and Calibration The authors close the gap by combining two extensions to a Weibull accelerated failure time model: calibrating team strength parameters to Betfair Exchange prices at kick-off to capture pre-match market information, and including post-shot expected goals as a time-varying covariate to capture in-play information The calibration approach involves jointly fitting team-strength parameters to 1X2 and over/under betting markets via squared-error minimisation which applies to any intensity-based goal arrival model and enables stronger in-play forecasting They evaluate all models against Betfair Exchange in-play prices across 140 matches from the second half of the 2024–25 season, using log-loss, Ranked Probability Score, and classification accuracy at minute intervals How the training signal is built The model incorporates several adaptations for in-play prediction to evolve beyond static pre-match estimates These adaptations include: 1.
Tom: Time decay, where team strengths are re-estimated after each gameweek using exponential time decay based on previous matches 2.
Jane: Half-specific shape parameters, which show that the hazard increases substantially faster in the second half 3.
Tom: Conditional Weibull sampling, which accounts for the fact that the distribution is not memoryless by conditioning on elapsed time since the last goal 4.
Title and authors: Tom: We started by talking about the title, "A market-initialised accelerated failure time model for in-play football forecasting," and what that actually means in plain English.
Jane: It essentially means they are using a model designed to predict when a goal will happen, but they’re initializing that prediction not just from pre-match data, but by calibrating the fundamental team strength parameters to what the betting exchange prices are doing at kick-off.
Lu: So, instead of guessing how strong Team A is before the game based only on historical stats or simple models, they are using actual market pricing information to set their starting point for how good that team is.
Meng: It’s like if you were building a prediction tool and you used the exchange's current odds for a team as your initial guess instead of just using the historical win percentage.
Tom: Right, and it’s about getting that pre-match market context into the in-play predictions, which is where most models fall short because they only look at what happened before.
Jane: And this paper focuses on how that pre-match information stays relevant throughout the game by adding a dynamic component.
Lu: They do that by including post-shot expected goals as a time varying covariate, which lets the model adjust its scoring rate based on what’s happening in the current moments of play.
Meng: So it's not just a static pre-match guess; it’s an evolving prediction that constantly gets updated by live actions.
Tom: And they tested this against one hundred forty matches, and the key result is that this calibrated model almost matches Betfair’s classification accuracy of seventy point six percent <ref:2605.16066#pg1,calibrated model almost matches Betfair’s classification accuracy>.
Jane: That's a high number because it shows that their method is very effective at capturing the dynamics of in-play football outcomes.
Lu: And they keep team-level parameters interpretable, so we don't just get a black box and have to guess why the prediction shifted.
Meng: That’s good for us because if something breaks in production, we can look at those parameters and see which part of the input is causing the drift.
Tom: So what they are really showing is that incorporating market calibration at the start and dynamic in-play features makes a big difference compared to older methods.
Jane: It’s about making the model smarter by feeding it more relevant context throughout the entire game, not just before it starts.
Lu: And this whole approach confirms that calibrating team strengths to oneXtwo and over/under markets via squared error minimization is a powerful way to make stronger in-play forecasting <ref:2605.16066#pg1,to 1X2 and over/under>.
Meng: That sounds mathematically heavy, but if the results hold up on real traffic, it’s worth exploring.
Tom: It certainly seems like a robust direction for improving our existing models, especially when accuracy against market prices is the benchmark.
The paper's summary: Jane: Now that we know *how* they do it, let's talk about what the actual summary of "A market-initialised accelerated failure time model for in-play football forecasting" says about their findings.
Tom: The core message is that combining pre-match market calibration with post-shot expected goals as a time varying covariate closes the gap between theoretical models and real betting exchange prices.
Lu: They are summarizing that this combination allows them to capture both the pre-match context and the immediate, in-play dynamics of goal arrival, which is something previous in-play models struggled with because they ignored pre-match market prices entirely.
Meng: Basically, they’ve fixed the constraint where models are only looking at final scores by allowing the scoring rate to evolve based on live events.
Tom: And their evaluation shows that this calibrated model almost matches Betfair’s classification accuracy of seventy point six percent across one hundred forty matches evaluated at minute intervals <ref:2605.16066#pg1,calibrated model almost matches Betfair’s classification accuracy>.
Jane: That accuracy is what’s impressive because it proves that this method is not just theoretically sound; it performs well on real-world data from the second half of the two thousand twenty-four–twenty-five season.
Lu: They also confirmed that this approach, when compared against other continuous-time scoring models calibrated to the same pre-match odds, shows that market calibration is the dominant driver of predictive accuracy.
Meng: So they are essentially saying that if you want better in-play forecasts, start by calibrating your team strengths to market prices first.
Tom: And on the profitability side, they found that Kelly staking with this calibrated model yields a return on investment of four point five percent over seventeen thousand four hundred fifty-eight bets.
Jane: That’s a concrete result showing that this isn't just about accuracy; it actually translates into better trading performance when you use the right staking strategy.
Lu: And they also showed that the edge in betting simulation comes from an information timing advantage—the model reacts immediately to a goal, while exchange prices might lag by one or two minutes.
Meng: That’s a crucial detail; it means we aren't exploiting stale prices, but rather using the model's speed to react to events.
Tom: It’s not about being better at predicting the exact final score in isolation; it’s about having a better mechanism for updating your belief moment by moment.
Jane: So, they are showing that this specific combination of calibration and time-varying covariates is what unlocks superior predictive power in dynamic sports environments.
Lu: And they also pointed out some limitations, like the Weibull gap-time framework of Leriou and Ntzouas hasn't been applied to in-play yet, which is a clear area for future research.
Meng: So, the summary shows they’ve done a lot of work but still have those gaps they need to fill before it becomes a completely complete system.
Tom: And that’s where our discussion moves next: what are the specific technical improvements they suggest to make this model even better than what we already have.
The paper's improvements: Lu: Let’s talk about the improvements they suggest for "A market-initialised accelerated failure time model for in-play football forecasting."
Jane: They are suggesting several technical enhancements to make it even better, starting with time decay, where team strengths are re-estimated after each gameweek using exponential time decay based on previous matches.
Tom: That’s about capturing the slow evolution of a team’s strength across different periods, making sure the baseline isn't stuck in yesterday's data.
Lu: They also have half-specific shape parameters to better capture the rise in scoring rate as matches progress, which suggests a single shape parameter isn't enough to describe how quickly things ramp up.
Meng: That’s interesting because if you only use one shape parameter, you might miss the nuance of how fast a match is actually heating up.
Tom: And they suggest conditional Weibull sampling based on the elapsed time since the last goal, which means conditioning on that elapsed time to sample remaining survival time.
Jane: That directly addresses that non-memoryless issue, making sure our predictions are accurate even when goals are infrequent or happen in bursts.
Lu: They also propose including covariates like red card difference, where positive values indicate a player advantage, which substantially affects match dynamics by reducing expected goal arrival times.
Meng: So if you’re looking at player advantages, this covariate could actually reduce the expected time until the next goal happens because it changes the underlying dynamic.
Tom: And they also include PSxG deviation to capture shot quality more granularly than just total cumulative PSxG, which is a nice way to refine that attacking output measure.
Jane: So in short, they’re suggesting we make sure we model not just the timing of goals but also the quality of those goals and external factors like cards.
Lu: And they also have this calibration approach where you jointly fit team-strength parameters to oneXtwo and over/under betting markets via squared error minimization, which is what makes their in-play forecasting stronger <ref:2605.16066#pg1,jointly fit team-strength parameters to 1X2 and over/under betting markets>.
Meng: That joint fitting sounds like a lot of work, but if it leads to better in-play forecasts, the effort might be justified for production use.
Tom: It’s definitely about pushing the model beyond just fitting the final score and into modeling the whole process of goal arrival moment by moment.
Conclusion: Jane: So, to wrap up our discussion on "A market-initialised accelerated failure time model for in-play football forecasting," we’ve covered how this paper uses market calibration and dynamic covariates to improve in-play prediction accuracy.
Tom: The main thing is that it demonstrates that this combined approach gives us a model that performs very well against exchange prices on a large test set of matches.
Lu: And the main implication for the field is that you need to account for pre-match market information and continuous updates during the game to get reliable in-play results.
Meng: For deployment, it means we have to build systems that can handle those joint optimization tasks and complex parameter fitting without introducing too much latency.
Lalam: I think the paper confirms that calibration to market prices is a dominant driver of predictive accuracy, which is a strong signal for any future work in this area.
Tom: It’s about getting better at modeling the process rather than just trying to predict an outcome, right?
Jane: Right, and this paper does exactly that by focusing on the evolution of goal arrival over time during the match.
Lu: The authors also highlighted their limitations, specifically that they haven't fully applied their proposed gap-time framework to in-play yet as a clear area for future research.
Meng: That means we can’t assume this is the final version of the system until it’s tested under those tougher conditions.
Lalam: And that’s fair; methodical replication will be needed to truly validate those profitability claims outside of simulation environments.
Tom: So, to wrap up, "A market-initialised accelerated failure time model for in-play football forecasting" is a solid piece of work that shows how leveraging market data and dynamic features can significantly improve in-play forecasting capabilities.
School of Engineering Mathematics and Technology, University of Bristol
stat.AP
Submitted: 2026-05-15
Updated: 2026-11-11
Comments: 25 pages, 6 figures, 6 tables. Substantially revised: doubled evaluation sample, revised model specification, additional benchmarks, and updated results and conclusions
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 81/100
The gist: In-play football forecasting models have struggled to match the accuracy of betting exchange prices, which aggregate information from many market participants The calibrated Weibull model exhibits a…
Key concepts
- Weibull Accelerated Failure Time Model
- This is a statistical model used to predict how long until an event occurs, such as a goal in football. It's an 'accelerated failure time' model because it focuses on the rate at which goals happen over time, allowing the researchers to estimate scoring rates based on match dynamics.
- Market Calibration
- This is the process of adjusting the model's parameters (like team strength) using real-time betting market prices from Betfair Exchange at kick-off. This step injects pre-match information directly into the model, making its predictions more relevant to what the market expects.
- Post-Shot Expected Goals (PSxG)
- PSxG measures the expected goals a shot is likely to result in. By including this as a time-varying covariate, the model can dynamically adjust its prediction of scoring rates based on how effective shots are becoming during the match, capturing in-play momentum.
- Time Decay
- This technique adjusts team strength estimates after each gameweek by using an exponential function. This accounts for the fact that a team's strength changes over time as they play more matches, ensuring the model reflects current form rather than just static pre-match data.
Terminology
Summary
In-play football forecasting models have struggled to match the accuracy of betting exchange prices, which aggregate information from many market participants The calibrated Weibull model exhibits a similar classification accuracy to Betfair across 140 evaluation matches, while retaining interpretable team-level parameters and covariate effects that the market does not provide
Model Extensions and Calibration
The authors close the gap by combining two extensions to a Weibull accelerated failure time model: calibrating team strength parameters to Betfair Exchange prices at kick-off to capture pre-match market information, and including post-shot expected goals as a time-varying covariate to capture in-play information The calibration approach involves jointly fitting team-strength parameters to 1X2 and over/under betting markets via squared-error minimisation which applies to any intensity-based goal arrival model and enables stronger in-play forecasting They evaluate all models against Betfair Exchange in-play prices across 140 matches from the second half of the 2024–25 season, using log-loss, Ranked Probability Score, and classification accuracy at minute intervals
How the training signal is built
The model incorporates several adaptations for in-play prediction to evolve beyond static pre-match estimates These adaptations include:
-
Time decay, where team strengths are re-estimated after each gameweek using exponential time decay based on previous matches
-
Half-specific shape parameters, which show that the hazard increases substantially faster in the second half
-
Conditional Weibull sampling, which accounts for the fact that the distribution is not memoryless by conditioning on elapsed time since the last goal
-
In-play covariates, such as post-shot expected goals (PSxG) and red cards, which update scoring rates as events occur
Where the gains come from
The comparison with two alternative continuous-time scoring models confirms that market calibration is the dominant driver of predictive accuracy The calibrated Weibull and Maia models perform similarly, and both outperform the Zou model, confirming that calibration to market prices is the dominant driver of predictive accuracy Furthermore, betting simulation with the calibrated Weibull model against Betfair in-play odds yields a 4.5% return on investment (Sharpe ratio 5. p. 094) over 17,458 bets, suggesting an inefficiency within in-play football markets
Predictive Performance and Market Advantage
The calibrated Weibull model achieves the highest classification accuracy (70.2%) among all models evaluated at minute 0 However, Betfair retains the best RPS and log-loss across all models, suggesting that while the model captures predictive information not fully reflected in exchange prices, Betfair still holds general dominance throughout the match as the market incorporates information beyond goals and shots that our covariates do not capture The profitability results show that Kelly staking with the calibrated model achieves an ROI of 4. p. 25% (Sharpe 5. p. 0 p. 094) over 17,458 bets, confirming that calibrated team strengths substantially improve in-play trading performance The authors conclude that while accuracy and profitability are distinct objectives and the small sample size is a limitation for betting evaluation
Future Research Directions
Future research could extend the model in several directions, such as incorporating expected threat from passing sequences to capture sustained momentum more effectively than shot-level statistics The half-time gap could also be narrowed by incorporating half-time market prices as a second calibration point, updating team strengths mid-match in the same way we calibrate to pre-match odds Extending the evaluation to multiple leagues and seasons would test the generalisability of both the underlying Weibull model and the calibration approach The access to Betfair’s Pro tier historical data, which records limit order book depth at 50ms intervals, would further enable executable rather than last-traded prices
Betting Simulation Insights
The betting simulation results show that the model's edge is derived from an information timing advantage: the model observes a goal and updates immediately, whereas exchange prices may lag by one or two minutes The similar profitability across both settings indicates that the model’s edge does not arise from exploiting stale post-goal prices Kelly staking avoids this by sizing bets proportional to edge, reducing exposure when the perceived advantage is small The authors consider the predictive accuracy results more robust than the betting returns and believe that methodical replication is needed to validate profitability claims in sports forecasting
Model Comparison Summary
The comparison with the Bayesian birth process of Zou et al. (2020) and the Cox process of Maia et al. p. 2 (2026), both calibrated to the same pre-match odds, confirms that calibration to market prices is the dominant driver of predictive accuracy The similarity between two structurally different models indicates the importance of pre-match calibration for in-play forecasting accuracy The predominant positive direction suggests the uncalibrated model overestimates scoring rates, consistent with the over-prediction of high goal counts observed in Figure 2 The calibration successfully incorporates market information into team-strength parameters across structurally different model classes, while retaining each model’s structure for in-play updating The most accurate pre-match forecaster on the evaluation set is the uncalibrated Weibull achieving 56. p. 04% pre-match accuracy versus 61. p. 04% for Betfair
Key Findings on Covariates
The inclusion of PSxG deviation and red cards significantly improves the model fit, with including PSxG deviation achieving the best BIC (ΔBIC = −53.4) The positive coefficients for these covariates confirm that sustained attacking pressure shortens expected goal arrival times The model’s edge is derived from an information timing advantage, as the Kelly bets fall into a “goal window” where a goal occurred in the preceding five-minute interval The model's performance against Betfair improves at minute 45 (the second-half restart), as the half-time interval allows the market to fully absorb first-half information The model’s edge does not arise from exploiting stale post-goal prices, as profitability and forecast accuracy are not equivalent objectives The model’s edge is derived from an information timing advantage, as the Kelly bets fall into a “goal window” where a goal occurred in the preceding five-minute interval
Betting Simulation Results
The betting simulation shows that calibrated Weibull staking yields an ROI of 4.5% (Sharpe 5.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Market calibrated accelerated failure time model for in-play football forecasting: The improved AI system can capture pre-match market information by calibrating team strength parameters to Betfair Exchange prices at kick-off via squared-error minimisation, leading to stronger in-play forecasting.
-
Incorporation of post shot expected goals as a time varying covariate: The system can incorporate post shot expected goals (PSxG) as a time varying covariate to capture shot volume and quality beyond what the score alone reflects, improving predictive accuracy.
-
Use of half specific shape parameters for the Weibull distribution: The model can utilize half specific shape parameters to better capture the rise in scoring rate as matches progress, leading to improved fit compared to models using a single shape parameter or score state dependent ones.
-
Time decay implementation for team strengths: The system can implement exponential time decay where previous matches are weighted by an exponential function of elapsed days, capturing the evolution of team strengths across gameweeks.
-
Conditional Weibull sampling based on elapsed time since last goal: The system can use a conditional Weibull inverse CDF to sample remaining survival time based on the elapsed time since the last goal, conditioning on the current score and elapsed time.
-
Inclusion of red card difference as a covariate: The system can include red card difference (own minus opponent) as a covariate, where positive values indicate a player advantage, which substantially affects match dynamics by reducing expected goal arrival times.
-
Integration of shot quality via PSxG deviation: The system can use the PSxG deviation covariate to improve fit by capturing more granular measures of attacking output compared to just the total cumulative PSxG process.
-
Calibration of scoring rate parameters using over/under markets: The AI system can calibrate scoring rate parameters to Betfair Exchange prices at kick-off using over/under markets, which provide direct estimates of the goals distribution, fixing the total scoring rate and yielding a unique solution.
-
Comparison with alternative continuous time scoring models: The system can compare its calibrated model against two alternative continuous time scoring models, both calibrated to the same pre-match odds, confirming that market calibration is the dominant driver of predictive accuracy.
Sources
Related papers
- Unraveling time-varying causal effects of multiple exposures: integrating Functional Data Analysis with Multivariable Mendelian Randomization
- Strategic Causal Policy Learning with Manipulable Eligibility
- BayesJudge: Uncertainty-Aware Bayesian Meta-Evaluation of Human and LLM Judgments
- Automating structural reliability analysis with a multi-agent large language model framework
- Improving Minority Population Sampling with BISG Probabilities: Evidence from a Survey of Jewish Americans
- When Do Sports Leagues Stop Being Competitive? Bayesian Detection of Competitive-Balance Eras from Match-Level Data