Forecast Sports Outcomes under Efficient Market Hypothesis: Theoretical and Experimental Analysis of Odds-Only and Generalised Linear Models
New AI models outperform existing methods by aligning with bookmakers' profit confidence, tested on 90,014 football matches.
A team of researchers including Kaito Goto and Naoya Takeishi has published a groundbreaking paper introducing two novel methods for converting sports betting odds into accurate outcome probabilities. Their work addresses fundamental challenges in using betting odds as benchmarks for sports forecasting and market efficiency analysis under the Efficient Market Hypothesis. The researchers analyzed an extensive dataset of 90,014 football matches across five different bookmakers, revealing limitations in existing conversion methods like Multiplicative, Shin, and Power approaches.
Their first innovation, the Odds-Only-Equal-Profitability-Confidence (OO-EPC) method, represents a significant advancement by operating without historical data for model fitting. Unlike traditional approaches, OO-EPC aligns with bookmakers' actual pricing objectives of maintaining equal confidence in profitability for each possible outcome. The researchers provided empirical evidence showing their OO-EPC method outperforms existing odds-only methods for most bookmakers in their dataset. Beyond controlled experiments, they successfully applied OO-EPC in real-world conditions across six iterations of an annual basketball outcome forecasting competition.
The second breakthrough comes in the form of the Favourite-Longshot-Bias-Adjusted Generalised Linear Model (FL-GLM), which utilizes historical data while addressing a key market inefficiency. Traditional generalised linear models often attempt to capture relationships already accounted for by the Efficient Market Hypothesis, but FL-GLM takes a more elegant approach by fitting just one parameter to specifically capture the favourite-longshot bias. This creates a more interpretable alternative to existing multinomial and logistic models, with the research demonstrating FL-GLM's superior performance across all bookmakers in their historical football match analysis.
- OO-EPC method converts betting odds without historical data, outperforming existing methods for most bookmakers in 90,014-match dataset
- FL-GLM model uses single parameter to capture favourite-longshot bias, beating multinomial/logistic GLMs across all bookmakers
- Methods tested in real-world basketball forecasting competitions and provide interpretable alternatives for market efficiency analysis
Why It Matters
Provides more accurate sports forecasting tools and insights into market efficiency, with applications in finance and betting analytics.