Every time someone asks how EdgeXI works, the honest answer starts with what the models don't do: they don't predict who will win. They estimate which side has a statistical advantage, by how much, and whether that advantage is large enough to be worth acting on.
That distinction is not a technicality. It is the operating principle behind everything we build.
What the Models Take In
EdgeXI's machine learning models are trained on structured historical match data across five T20 leagues: IPL, BBL, WBBL, CPL, and Super Smash. The inputs include match-level performance data, venue characteristics, squad composition, bowling and batting matchups, and seasonal form trajectories.
For head-to-head predictions, the core question is: given the specific conditions of this match — venue, pitch, both squads' current form, and the statistical matchup between bowling attack and batting order — what is the probability distribution across possible outcomes?
The models do not read injury news on the morning of a match. They do not process social media sentiment or press conference remarks. They work with structured data, and their outputs are only as current as the data fed into them. We address this limitation partly through our coverage window (matches 11 through 47 of the IPL), which excludes the period of highest uncertainty around squad formation at the start of a season.
How the Output Is Generated
The models produce a probability estimate for each head-to-head outcome. This is not a binary call. It is a distribution — Team A has a 58% probability of winning, Team B has a 42% probability — with an associated confidence interval that reflects the uncertainty in the estimate.
A recommendation is only issued when the probability distribution is sufficiently skewed and the confidence interval is sufficiently tight. In many matches, neither condition is met. The models flag those matches as outside recommendation range, and we publish nothing for them. Selectivity is a feature, not a limitation.
When a recommendation is issued, the output includes a suggested bankroll allocation based on the Kelly Criterion, scaled to the magnitude of the estimated edge. Higher-confidence, larger-edge recommendations receive higher allocations. Smaller-edge recommendations receive smaller ones, even if the direction is the same.
What the Models Don't Do
They don't account for individual match variance. A batter having an exceptional game, a bowling attack finding unexpected swing, a dropped catch in the 18th over — none of these are predictable from historical data. They are part of the variance that sits within the probability distribution.
They don't guarantee outcomes across individual matches. The models produce estimates that, when applied consistently over a large sample of matches, should deliver positive expected value. Past seasons bear this out. But any individual match can go against the statistical edge, and approximately one in three of our recommendations results in a loss. That is expected, not exceptional.
They don't claim to be the only valid view. Two analysts using different data sources and different model architectures can reach different probability estimates for the same match. Our estimates reflect the specific methodology we have developed and refined over three seasons. They are not the definitive truth about any match.
Why Machine Learning Specifically
Machine learning models are well-suited to this problem for a specific reason: the variables that drive T20 outcomes are numerous, interact in non-linear ways, and shift across different contexts. A fast bowler's effectiveness is not just a function of their economy rate. It is a function of that rate against left-handed openers, in the first six overs, at venues with shorter square boundaries, in matches where the pitch is dry. Capturing those interactions in a conventional statistical model requires a large number of manually specified variables. Machine learning approaches can identify those interactions from the data directly.
The tradeoff is interpretability. Neural network layers and ensemble methods are less transparent than a regression coefficient. We address this by building interpretability checks into the model validation process: each model version is tested not just for accuracy, but for whether its feature weights are consistent with cricket domain knowledge. If a model assigns high predictive weight to a variable that has no plausible causal relationship with match outcomes, that is a signal of overfitting, and the model is retrained.
Three Seasons of Data
Our published ROI figures cover three full seasons across five T20 leagues (2023, 2024, 2025). These are the outcomes produced by applying the same framework — consistent data sources, consistent recommendation thresholds, consistent bankroll sizing — across a defined set of markets on Bet365, William Hill, and Ladbrokes (retail) and Betfair Exchange.
The retail figures are higher than the exchange figures. Retail bookmakers offer better headline odds, while exchanges offer tighter margins across a larger number of markets. The difference is documented in detail in our Understanding Our Returns post.
These historical figures are the best available evidence of model performance. They are not a forecast. The 2026 IPL season begins with the same methodology applied to new data. Whether the outcomes match, exceed, or fall below the historical average will be determined by how well the model's probability estimates reflect the actual distribution of results — and that is something we will only know at the end of the season.
Past performance is not a guarantee of future results. EdgeXI recommendations are provided for research purposes only. Please gamble responsibly.