Module VIII: Using the platform
41. Models and AI analysis
Seven models, cards on the table
Seven predictive models run on the platform: five textbook base models (Elo, Poisson, Monte Carlo, logistic regression, gradient boosting) and two in-house combinations (Veto, Balance). Each has its own page with six tabs, and we publish all their numbers, the bad ones too.
The "About the model" tab explains the methodology with formulas, strengths and weaknesses — a live illustration of the chapters of Module VI (28–34). The "Results" tab is the back-measurement: hit rate, ROI, average odds, average CLV and the closing-line-beat rate (chapter 12), broken down by odds band. The "League breakdown" and "Bookmakers" tabs show where and on which bookmaker’s prices the model performs better or worse, and the "Strategies" tab runs a bankroll simulation with five staking strategies (flat, martingale, fibonacci, value, Kelly), with a drawdown and bust-rate column.
The calibration panel: this is how to choose a model
The most important block of the Results tab is the calibration panel: Brier score, log loss, ECE and a reliability diagram, on which it is visible whether, when the model says 60%, the outcome really occurs in 60% of cases (chapter 32). The panel also shows the models’ calibration ranking.
The lesson, which the research cited in chapter 14 (Walsh and Joshi, 2024) also supports: a model should be chosen by calibration, not by hit rate. A badly calibrated model leaks money directly at value- and Kelly-based stake sizing, because it systematically overestimates its own rightness.
The daily AI analysis and the AI ledger
On the AI page a separate language model’s daily, per-match analyses can be read: its own 1X2 estimate, a goal-count, BTTS, corner and card opinion, three scenarios of increasing risk and a multiple example. The AI deliberately sees neither the platform models’ prediction nor the odds, so as to give reasoning independent of them; on the card a separate band shows the AI’s raw estimate and the calibrated probability adjusted toward the market consensus.
The AI ledger page settles every analysis type after the fact: hit rate, profit in units, ROI, and a 95% confidence interval next to every ROI. If the interval overlaps zero, the page states it: the result is not statistically distinguishable from zero — this is the reasoning of chapter 27 on live data.
What is all this good for if it does not beat the market anyway?
In chapter 14 we showed: according to our own research the market’s closing prices are more accurate than any of our models, and we do not hide this but measure and publish it. The models and the AI are not tip buttons but reference points: on them you learn to read matches as probability, you see the working of variance live (Module IV), and you can check at any time what a method actually knows, before anyone sells you a "working system".
If a model or AI analysis goes well for a month, that is still noise. The question is always the same: what does the calibration, the CLV and the confidence interval say. These are exactly the three instruments you find on every one of our pages.