OddsBeacon ingests match outcomes, scores, and player performance data across table tennis, tennis, and soccer — then runs a three-stage ensemble calibrated on forward-chained backtests. No look-ahead. No revision after publish. The pipeline is documented here in full.
PIPELINE · END TO END
From raw match feed to tonight's pick
01
INGEST
—
Table tennis, tennis, soccer — 24 active leagues. 1.5M+ match records across 141K athletes.
02
CLEAN
~9.4M pts
Disambiguate player IDs, normalize set scoring, drop walkovers and retirements.
03
MODEL
3-ensemble
Anchor (baseline) · Edge · Momentum. Final output is probability-weighted Consensus.
04
VALIDATE
—
Backtest forward only. No look-ahead. Calibrated with isotonic regression.
05
PUBLISH
14 min
Median latency from match start to prediction publish. Locked-state recorded at publish time.
BACKTEST · ROLLING MONTHLY · FORWARD-CHAINED
Accuracy curve
"Each month of training sharpens the edge — the models improve continuously."
NO DATA
BY CONFIDENCE TIER
LOADING…
FEATURE IMPORTANCE · ENSEMBLE · SHAP
Composite performance index (4-system)
34.2%
H2H recency-weighted
18.6%
Form streak (rolling 10)
12.4%
Venue × playstyle interaction
8.9%
Rest hours between matches
6.7%
Match format & scoring system
5.4%
Tournament stage / round
4.1%
Surface, environment & conditions
3.3%
Home crowd / venue effect
2.8%
Other (24 features)
3.6%
EDITORIAL STANDARDS
Every prediction is timestamped at publication and never re-edited. Misses appear in the public record exactly as the hits do. Confidence tiers are calibrated; "HIGH" should hit roughly two-thirds of the time, and the live data on the left says it does.