How Prediction Engine Works

Our MLB pipeline uses three separate team models — one each for moneyline (win/loss classification), run line (margin regression), and totals (over/under regression). Each model is trained on 42-44 engineered features including pitching matchups, bullpen usage, park factors, umpire tendencies, and rolling performance windows.

Player prop models cover five batter statistics (hits, runs, RBIs, home runs, combined HRR) and pitcher strikeouts. The pitcher K model identifies starters who are likely to hit their strikeout over/under, with 86.6% accuracy at the 80%+ confidence tier.

The Live K Tracker runs four XGBoost checkpoint models (innings 2 through 5) that predict a pitcher's final strikeout total from in-game data — pace, pitch count, K rate, and game context. By inning 5, the model achieves 90.7% over/under accuracy. The Live Hits Tracker uses 59 features including park factors, platoon splits, umpire tendencies, and weather to predict remaining hits for batters mid-game.

NBA team models use Random Forest for moneyline and spread prediction, and XGBoost for totals. The moneyline model reaches 78.4% accuracy at the 70%+ confidence filter. Spread predictions use regression to estimate winning margin, and the totals model trains daily on rolling data from three seasons.

Player props cover points, assists, rebounds, double-doubles, and triple-doubles. The points and assists models use XGBoost regression combined with CDF (cumulative distribution function) probability to generate over/under confidence for each prop line.

The Live Player Stats Tracker runs nine XGBoost models — three stats (points, rebounds, assists) across three quarterly checkpoints (Q1, Q2, Q3). By Q3, the assists model hits 94% over/under accuracy. These models compare real-time projections against sportsbook prop lines to generate live over/under signals during games.

Models improve over time through a self-training feedback loop. Each day, the pipeline grades yesterday's predictions against actual results, collects new game data, and feeds it back into the training set. Monthly retrains incorporate the latest data to keep models calibrated against the current season. This isn't a static system — it learns and adapts.

Every set of odds implies a probability. When a sportsbook offers +150 on a team, they're implying that team has about a 40% chance of winning. If your model calculates a 52% chance, that 12-point gap is your edge — and that bet is +EV.

Here's the math: if you bet $100 at +150 odds with a true 52% win probability, your expected value per bet is:

You won't win every bet. But over hundreds of bets, the math plays out. This is the same principle that makes casinos profitable — except here, the edge is on your side.

Prediction Engine shows both confidence (model's probability a pick wins) and edge (the gap between model probability and Vegas implied probability). A pick can be high-confidence but low-edge if the sportsbook already agrees. The most valuable picks have both high confidence AND significant edge — meaning the model sees something the market hasn't priced in.

Sportsbooks build a margin (the “vig” or “juice”) into every line. Standard -110/-110 odds imply both sides have a 52.4% chance of winning — totaling 104.8%, not 100%. That extra 4.8% is the house edge. To beat the vig, you need to be right more than 52.4% of the time on standard bets. Casual bettors who pick based on fandom or narratives rarely clear that bar consistently.

The simplest and most popular approach. You bet the same fixed amount — one “unit” — on every play regardless of confidence. A unit is typically 1-3% of your total bankroll.

A step up from flat betting. You assign 1-3 units based on the model's confidence level. Prediction Engine labels picks by tier — a high-confidence play with large edge gets 3 units, a standard play gets 1.

The Kelly Criterion is a mathematical formula that calculates the optimal bet size to maximize long-term bankroll growth. It accounts for both the probability of winning and the odds being offered.

Example: Model says 60% win probability at +120 odds (decimal 2.2). Kelly = (1.2 × 0.60 − 0.40) / 1.2 = 0.267, or 26.7% of bankroll. In practice, full Kelly is extremely aggressive.

Most professionals use half Kelly or quarter Kelly — you calculate the full Kelly percentage then bet 25-50% of it. This dramatically reduces variance while capturing most of the long-term growth.