How the BreakPoint ML Model Works

The Challenge of Tennis Prediction

Tennis is uniquely difficult to predict. Unlike team sports, individual matchups matter enormously. A player who dominates on clay might struggle on grass. Recent form, surface expertise, and head-to-head history all play critical roles.

At BreakPoint, we built an ensemble ML model that combines the strengths of multiple prediction approaches to tackle this complexity.

Our Two-Layer Approach

Layer 1: Glicko-2 Rating System

Every player in our database has a surface-specific Glicko-2 rating — separate ratings for clay, grass, hard indoor, and hard outdoor. The Glicko-2 system improves on traditional Elo in two important ways:

• Rating Deviation (RD): Measures how certain we are about a player's rating. A player who hasn't played in months has a higher RD, meaning their rating is less reliable.
• Volatility: Captures how consistently a player performs. A player with erratic results has higher volatility.

Layer 2: ML Ensemble (Logistic Regression + XGBoost)

Our ML model takes a 10-feature vector for each matchup:

| Feature | Description | |---------|-------------| | Rating difference | Glicko-2 gap between players | | Player 1 RD | Rating confidence for player 1 | | Player 2 RD | Rating confidence for player 2 | | Form difference | Recent win rate gap (last 10 matches) | | Surface win rate (P1) | 2-year surface-specific win percentage | | Surface win rate (P2) | 2-year surface-specific win percentage | | Surface WR difference | Gap in surface expertise | | Months inactive (P1) | Time since last match | | Months inactive (P2) | Time since last match | | Tournament round | Round encoding (R32 through Final) |

The ensemble combines Logistic Regression (good at linear relationships) with calibrated XGBoost (captures non-linear patterns like surface specialization).

Why Calibration Matters

Raw ML probabilities are often overconfident. Our XGBoost model originally pushed predictions to extremes — predicting 85% win probability when the true rate was closer to 70%.

We solved this with isotonic calibration (CalibratedClassifierCV), which maps raw probabilities to their true observed frequencies. Our Expected Calibration Error (ECE) dropped to just 0.025, meaning when we say "70% probability," the player actually wins about 70% of the time.

The Market-Model Blend

Here's our most important insight: the betting market is efficient, but not perfect. We found that our model systematically overvalues underdogs — a common bias in ML prediction systems.

Our solution is a two-tier blending strategy:

• Normal matches: 60% model probability + 40% market-implied probability
• Heavy favorites (>80% implied): 40% model + 60% market

This correction significantly improved our real-world accuracy while preserving our edge in closer matchups where the model adds the most value.

Finding +EV Opportunities

An opportunity exists when our blended probability exceeds the vig-normalized implied probability from bookmaker odds. We normalize for vig (the bookmaker's margin, typically 2-3%) to avoid counting phantom edges.

Our thresholds:

• 4%+ edge: Recommended bet
• 3-4%: Borderline (worth watching)
• >10%: Suspicious (likely stale odds or model limitation)

What's Next

We're continuously improving our model with:

• Expanded training data (more historical matches)
• Weather and court speed factors
• Player fatigue modeling (tournament scheduling)
• Real-time odds movement tracking

Want to see the model in action? Check out today's opportunities or run your own prediction.