Diego Garcia

TESTING

Five specialised prediction models. Each covering a distinct betting market. Each independent. Together they map the full surface of a football match.

Cronos·Hyperion·Coeus·Apollo·Themis
SCROLL

THE SYSTEM

Five specialists. One output.

Most prediction systems treat football as a single problem. Kairox Tempo treats it as five. Each model was designed independently for its market's statistical structure. They do not share logic, only data. The edge comes from comparing calibrated model probabilities against bookmaker odds, not from beating the bookmaker at their own game.

INDEPENDENCE

No model feeds another. Each produces its own probability estimate for its own market. This keeps errors contained: a bad Cronos prediction does not corrupt Hyperion.

CALIBRATION

Raw model output is probability-calibrated using isotonic regression. The goal is not accuracy in the classification sense but accurate probability estimates that reflect real-world frequencies.

THE EDGE

Edge is computed post-inference: model probability against scraped bookmaker odds. Negative edge bets are never surfaced. No odds features in training: the model is blind to bookmaker pricing.

ARCHITECTURE

How the system is built

Each model draws from the same historical match database and produces its own probability output. Those outputs converge into a single prediction payload, enriched with scraped odds and computed edge per qualifying bet line.

FIVE INDEPENDENT MODELS · ONE PREDICTIONPREDICTIONprobability+ edgeCronos1X2 outcomeHyperionGoals O/UCoeusCorners O/UApolloShots / SoTThemisCards O/UDATA SOURCESMatch historyLeague positionsxG / possessionHead-to-headReferee statsScraper odds

THE PANTHEON

The five models

Each architecture was chosen for its market's statistical structure, not for consistency across the system. A Poisson model for goals, Negative Binomial for corners and shots, XGBoost for outcomes, LightGBM for cards. The right tool for each problem.

01
THE ARBITER

Cronos

XGBoost + Isotonic Calibration

Match outcomes · 1X2

30 league-specific engineered features, tuned per league to capture what matters most in each competition: cumulative points-per-game by venue, head-to-head history, league position, xG performance, referee tendencies. The feature set is not shared across leagues. Cumulative statistics beat rolling windows in noisy domains. Simplicity learned the hard way after watching accuracy drop from 60.8% to 53.2% when switching to rolling features.

OUTPUT

home_win · draw · away_win with probability and edge per line

02
THE ACCOUNTANT

Hyperion

Dixon-Coles Poisson + Isotonic Calibration

Goals over/under · 0.5 to 5.5

Dixon-Coles corrects the standard Poisson model's known underestimation of low-scoring matches through a tau correlation parameter. Time-decay weighting ensures recent form dominates. Home advantage modelled explicitly per league: La Liga 0.2786, Premier League 0.2107, Bundesliga 0.2726.

OUTPUT

over/under at six lines with probability and edge per line

03
THE SURVEYOR

Coeus

Negative Binomial + Isotonic Calibration

Corners over/under · 6.5 to 13.5

Corners are overdispersed count data; variance exceeds the mean. Standard Poisson underestimates this. Separate home and away NB models with independent overdispersion parameters, combined for total corners. The dual model structure captures asymmetric home/away corner distributions.

OUTPUT

over/under at eight lines per team and total with probability and edge

04
THE ANALYST

Apollo

Negative Binomial + Isotonic Calibration

Shots and SoT over/under

Same overdispersion reasoning as Coeus. Shot volume and shots-on-target have different variance structures so each gets its own model. Window lengths chosen by validation: eight matches for shots, twelve for SoT. Separate models for home shots and total shots rather than deriving one from the other.

OUTPUT

home shots at multiple thresholds · total SoT at multiple thresholds

05
THE WARDEN

Themis

LightGBM Poisson

Cards over/under · 1.5 to 8.5

Cards correlate with match intensity, referee tendencies, and rivalry context rather than raw team strength. LightGBM handles the feature interactions more flexibly than a pure count model. Encoding: red = 2, yellow = 1. No data leakage: card counts are not derived from any match outcome feature.

OUTPUT

total cards over/under at eight lines with probability and edge

WHERE IT IS GOING

The models work. The architecture is in place. What is being built now is the layer that sits on top: the staking system, the bankroll management, the interface that makes all of this usable by someone other than me.

More leagues are being added. The betting strategy that sits on top of the raw model output is the part that actually determines profitability, and that is what the testing phase is measuring.

TESTING · 5 MODELS · LEAGUES EXPANDING