How We Built the fault.bet Model That Finds Mispriced Tennis Markets

April 2026 · 8 min read · View live results

Tennis betting markets are efficient — but not perfectly efficient. Bookmakers set prices based on rankings, recent form, and public perception. They don't account for pressure performance under specific conditions, weather interactions with serve mechanics, or fatigue accumulation across multi-week tour swings.

We built a fault.bet modeling system that captures these edges. Here's how.

The Data

• 1.6 million historical matches — ATP + WTA, 2005 to present
• 562,000 point-level records — for serve modelling and pressure analysis
• Complete weather history — temperature, wind speed, humidity for every outdoor match via Open-Meteo
• Live exchange prices — real-time market data from multiple sources

Architecture

Data Collection (16GB dedicated server)
├── WTA API → matches + serve stats
├── ESPN API → ATP results
├── FlashScore → serve stats (90%+ coverage)
├── Open-Meteo → weather per match
└── Historical archives → match database

Feature Engineering (49 features per player)
├── Surface-matched Elo/WElo
├── 20-match rolling serve/return windows
├── EWMA (exponential weighted moving average)
├── Pressure metrics (choke index, closing rate)
└── Weather-adjusted serve projections

Model Ensemble (4 models)
├── LightGBM
├── Random Forest
├── Logistic Regression
└── XGBoost
    → Optimised ensemble weights

Key Features That Drive Predictions

1. Weighted Elo (WElo) — standard Elo with a recency decay. A player's last 10 matches count more than matches from 6 months ago. We maintain separate ratings per surface.

2. Choke Index — a proprietary metric measuring how often a player loses games from a winning position. Some players crumble at 5-4 serving for the set. The market prices them the same as clutch closers.

3. Weather-Adjusted Serve Projections — wind above 20 km/h reduces first-serve accuracy by 5-8% on average. But some players are barely affected while others fall apart. We interpolate between calm/windy hold rates based on actual match-day forecasts.

4. Rolling Serve/Return Windows — 20-match rolling hold %, break %, ace rate, and double fault rate. Captures current form without overfitting to a single match.

Model Performance

• ATP V1 Ensemble: AUC 0.75 (out-of-sample), Brier 0.20
• WTA V3 Ensemble: AUC 0.71 (out-of-sample), Brier 0.22

We only fire signals when the model disagrees with the market by 6% or more AND the market price is confirmed from a live exchange. No estimated prices, no noise.

Live Results

Since March 2026, every signal is publicly tracked:

• 65 wins, 57 losses (53% strike rate)
• +14.5% bank growth using proportional staking
• +8.5% ROI on flat stakes

Full verified record: fault.bet/results