The Hydralytic Engine
Select a sport to explore the model architecture behind each prediction.
Hydra is not a database of historical results. It is a physics-first simulation engine — three deep data pipelines feed eleven specialized heads, each one trained to master a single outcome of a plate appearance.
Most betting tools ask: "what happened before?" Hydra asks: "what will the pitcher throw, what will the batter do with it, and what does the park do to the result?"
Pitcher Physics Engine
We measure what the batter sees at the decision point — not what crosses the plate.
Most models track pitch movement at home plate. This head tracks it at 30 feet — the moment a batter must commit. The difference between what the pitch looks like mid-flight and where it actually ends up is the deception. We quantify it across 25 physical features: vertical approach angle, horizontal break, late movement in both planes, force vectors, and location relative to the batter's personal zone. The head then converts that physics profile into a pitch-level Expected Run Value — the probability it induces a take, whiff, foul, or ball in play, and if in play, the distribution of what happens next.
Engine Insight: The Third-Time-Through Myth
Managers pull pitchers to avoid the 'third time through the order' — a real league-wide trend. But this head has identified pitchers whose expected run value improves in the 6th and 7th inning. They're being pulled at their statistical peak because managers are applying a population average to an individual. We don't.
Batter DNA
We separate what a batter intended from what actually happened.
Every swing produces three residuals: exit velocity above or below what the pitch quality warranted, launch angle above or below the batter's own intent profile, and spray angle deviation. Together they form a skill delta — the gap between what actually happened and what should have. A positive delta means the batter won the interaction. This head averages those deltas across thousands of plate appearances to separate true contact ability from luck, then layers in spray tendencies, hard-hit rates, sweet spot frequency, and zone aggression by pitch type — building a personalized offensive profile for every hitter in the league.
Engine Insight: What Statcast Misses
Standard xwOBA assigns the same expectation to every 95 mph ground ball. Our model knows a 95 mph grounder from a pull-heavy hitter into a defensive shift is fundamentally different from the same number off a hitter with elite grounder spray distribution. The distinction is worth 40+ points of batting average — and most books don't price it.
Environmental Topology
Parks aren't just dimensions — they're defense, wind, and air density combined.
This head does something most park models don't — it strips out defense first. Fielding quality is isolated and removed before calculating what the walls, dimensions, and foul territory actually contribute. Wind sensitivity is then mapped per stadium and per direction, because a 10 mph wind blowing out to left in Wrigley has a different run value than the same wind in Yankee Stadium. Air density anomalies from elevation, temperature, and humidity shift expected ball flight on every batted ball. The output is effective park factors broken down by outcome, batter handedness, and spray direction.
Engine Insight: The Fenway Fallacy
The book on Fenway is 'right-handed pull hitters dominate.' Our defense-adjusted topology disagrees. The highest-EV batter profile in Fenway is a left-handed spray hitter — by going to all fields, they weaponize Fenway's geometry and become dangerous no matter where the ball is hit.
11 Specialist Heads
One head per outcome. Each sees only what matters for that specific result.
The outputs of the three data heads flow into eleven specialist heads — one per plate appearance outcome. Each head has five input branches, and critically, each receives a different feature set tailored to what it's predicting. The home run head gets batter wind interaction and launch intent. The double play head gets sprint speed, infield alignment, and situational leverage. A single head trying to predict all eleven outcomes would learn compromises across all of them. These don't. They specialize.
Each head produces its own probability independently. Those eleven signals are then converged into a complete plate appearance outcome distribution. That distribution is handed to a Monte Carlo simulation engine that runs thousands of game iterations — producing win probabilities and scoring distributions rather than a single number to bet against.
See what the engine produces today.
View Today's MLB Picks →