DEVELOPMENT JOURNAL / #002 · 2026-08-11
Three core probability models, verified against the full test suite
Model layer | Elo · Dixon–Coles · LightGBM
3
MODELS VERIFIED
179
TESTS PASSING
100%
PASS RATE
As of 11 August 2026, the three probability models at the core of the analytics engine—Elo, Dixon–Coles and LightGBM—are confirmed to be producing correct, test-passing output. This entry records how that was verified and what it means for the platform.
What was verified
The analytics engine estimates outcome probabilities from several independent models before any fair-odds or value calculation takes place. Three of those models form the current core: Elo for team-strength rating, Dixon–Coles for low-scoring football adjustments, and LightGBM for nonlinear relationships across many features.
Each of the three was audited and run against its full verification suite. All of them produced correct, expected output—179 tests in total across the three models, passing at 100%. This is the first time the model layer has been confirmed end-to-end in this way, rather than assumed to be working from partial checks.
| Model | Role in the engine | Verification status |
|---|---|---|
| Elo | Team-strength rating and baseline win probability | Verified |
| Dixon–Coles | Low-scoring football and score-dependency adjustments | Verified |
| LightGBM | Nonlinear analysis across many features | Verified |
The compatibility issue behind it
In plain terms: a compatibility gap between the codebase and the version of Python now in use had been quietly preventing two of the three models from completing their verification suite. The models could not report a clean result, which meant their output could not be trusted without further work.
That gap has been resolved. With the code and the runtime aligned again, all three models run their suites through to completion and return the expected results. The fix was about restoring a reliable, repeatable check—not about changing how the models reason about a match.
How the models stay verified
A one-time pass is not the goal. The verification suite now runs automatically as part of the development process, so any future change that would disturb these results is caught early rather than discovered later.
This is consistent with how the platform is being built overall: reliability and transparency first, with the model layer held to a standard that can be re-checked on demand rather than taken on trust.
What is still ahead for the model roster
The three verified models are the current core, not the finished roster. Several additions are still upcoming and are not claimed as complete: a standalone Poisson model, Negative Binomial, Bayesian shrinkage, and Monte Carlo simulation.
Each of those will be held to the same standard before it is treated as part of the working engine: audited, run against a full suite, and only then relied upon. Progress on them will be recorded here as it happens.
The value of a model is only as good as the evidence that it runs correctly. Verifying that evidence—and keeping it verifiable—is part of the work, not an afterthought.
METHODOLOGY
How EV Bet Engine analyses markets
DEVELOPMENT JOURNAL
View all development entries
EV Bet Engine is the source of record for this article. Earlier versions may exist on other platforms; the latest text and update history are maintained here.