Chapter 25: Game Outcome Prediction - Key Takeaways

Executive Summary

Game outcome prediction synthesizes player evaluation, team analysis, and situational factors into probabilistic forecasts. While prediction models can achieve reasonable accuracy (65-68%), betting markets represent highly efficient aggregations of information that are difficult to beat consistently. This chapter provided frameworks for building prediction models, evaluating their performance, and understanding market efficiency.

Core Concepts

1. Prediction Targets

Relationship:

Win Prob = P(Spread > 0) = CDF_normal(Spread / StdDev)

2. Baseline Models

Always compare against baselines:

3. Key Prediction Factors

Primary Factors: - Team offensive/defensive efficiency - Home court advantage (~3-4 points) - Recent performance trend

Situational Factors: - Rest days differential (~1-2 points) - Travel distance (minor) - Altitude (Denver: +1-2 points) - Schedule density

Information Factors: - Injuries (varies by player impact) - Lineup changes - Back-to-back games

4. Evaluation Metrics

For Binary Predictions: - Accuracy: % correct predictions - Brier Score: Mean squared probability error (lower = better) - Log Loss: -log(predicted probability of actual outcome)

For Spread Predictions: - MAE: Mean absolute margin error - RMSE: Root mean squared error (~11-12 points typical) - ATS %: Against-the-spread accuracy

Formulas:

Brier Score = (1/n) × sum((p_i - o_i)^2)
Log Loss = -(1/n) × sum(o_i×log(p_i) + (1-o_i)×log(1-p_i))

5. Probability Calibration

Well-calibrated model: When predicting X%, outcomes should occur X% of the time.

Testing Calibration: 1. Group predictions by probability range 2. Compare predicted vs. actual win rates 3. Plot calibration curve 4. Calculate Expected Calibration Error (ECE)

Practical Application Checklist

Building a Prediction Model

• [ ] Define prediction target (spread, win prob, total)
• [ ] Collect historical game data (3+ seasons)
• [ ] Calculate team strength metrics (Elo, efficiency, etc.)
• [ ] Engineer situational features (rest, travel, home)
• [ ] Handle missing data (injuries, early season)
• [ ] Train model with time-based validation
• [ ] Evaluate against baselines
• [ ] Check calibration
• [ ] Document uncertainty

Pre-Game Prediction Process

• [ ] Update team ratings with most recent results
• [ ] Apply home court adjustment
• [ ] Incorporate known injuries/absences
• [ ] Check for situational factors (B2B, travel)
• [ ] Generate spread prediction
• [ ] Convert to win probability
• [ ] Apply uncertainty bounds
• [ ] Compare to market line (if available)

Model Evaluation

• [ ] Calculate accuracy on holdout data
• [ ] Compare to relevant baselines
• [ ] Check calibration across probability ranges
• [ ] Test for systematic biases
• [ ] Measure ATS performance (if applicable)
• [ ] Calculate statistical significance
• [ ] Report confidence intervals

Key Formulas Quick Reference

Win Probability from Spread

Win_Prob = norm.cdf(Spread / Std_Dev)
Example: 6-point favorite, 12-pt std dev: norm.cdf(0.5) = 69%

Elo Expected Score

E_A = 1 / (1 + 10^((R_B - R_A) / 400))

Elo Update

R_new = R_old + K × (Actual - Expected)

Pythagorean Expectation

Win% = PF^n / (PF^n + PA^n)
Where n ≈ 14 for NBA

Kelly Criterion

Kelly% = (bp - q) / b
Where: b = odds, p = win prob, q = 1-p

Break-Even Win Rate

At -110 odds: 110/210 = 52.4%
At -105 odds: 105/205 = 51.2%

Common Mistakes to Avoid

Mistake 1: Ignoring Sample Size

Problem: Drawing conclusions from small samples Solution: Calculate statistical significance; need 500+ games for reliable conclusions

Mistake 2: Overfitting

Problem: Model performs well on training data, poorly on new data Solution: Use time-based validation, regularization, simpler models

Mistake 3: Look-Ahead Bias

Problem: Using information not available at prediction time Solution: Strict time-based data separation, walk-forward validation

Mistake 4: Ignoring Calibration

Problem: Focusing only on accuracy, ignoring probability quality Solution: Always check and report calibration metrics

Mistake 5: Underestimating Markets

Problem: Assuming your model beats the market Solution: Use closing lines as the primary benchmark

Mistake 6: Ignoring Vig

Problem: Reporting gross win rates without accounting for betting costs Solution: Always calculate net ROI including transaction costs

Summary: Model Performance Expectations

Accuracy Benchmarks

ATS Expectations

Required Sample Sizes

Market Efficiency Summary

Why Markets Are Efficient

1. Competition: Many sophisticated participants
2. Speed: Information incorporated in minutes
3. Arbitrage: Price differences eliminated quickly
4. Data: Same public data available to all

Where Edges Might Exist

1. Speed: Acting on news faster than markets adjust
2. Private info: Injury, lineup information
3. Micro-markets: Less liquid, less efficient
4. Live betting: More noise, more opportunity
5. Promos: Bonuses can create positive EV

What Doesn't Work

1. Simple systems (home dogs, fading public)
2. Historical patterns (markets adapt)
3. Public models (sharps build similar models)
4. Past ATS performance (no persistence)

Further Study Recommendations

1. Statistical foundations: Study probability, regression, time series
2. Elo systems: Implement and calibrate your own rating system
3. Market analysis: Track line movements and closing line value
4. Simulation: Build Monte Carlo game and season simulators
5. Machine learning: Explore advanced prediction techniques
6. Evaluation: Master proper scoring rules and calibration analysis