marianopsa/pitch-aleph-1.2

PitchAleph 1.2: Quantitative In-Game Tactical Probability Engine

PitchAleph 1.2 is an open-weights, calibrated machine learning engine designed for real-time association football simulation and convex capital allocation.

The architecture pairs an XGBoost multi-class gradient booster with Isotonic Regression calibration (CalibratedClassifierCV via FrozenEstimator), eliminating probability compression and preventing retail favorite bias.

Quickstart (Load in Python / Colab)

import joblib
import pandas as pd
from huggingface_hub import hf_hub_download
from sklearn.frozen import FrozenEstimator

# 1. Download and deserialize weights
model_path = hf_hub_download(repo_id="marianopsa/pitch-aleph-1.2", filename="model.joblib")
model = joblib.load(model_path)

# 2. Score match state (Home trailing 0-1, Min 58, Elo +99)
features = pd.DataFrame([{
    'match_time_ratio': 58.0 / 90.0,
    'shot_differential': 6,
    'team_A_possession_share': 0.59,
    'goal_differential': -1,
    'team_A_prematch_elo': 1849.0,
    'team_B_prematch_elo': 1750.0,
    'prematch_elo_diff': 99.0,
    'event_seconds': 3480.0,
    'event_duel': 0, 'event_foul': 0, 'event_pass': 0, 'event_save': 0,
    'event_set_piece': 0, 'event_shot': 0, 'event_stoppage': 0, 'event_touch': 1
}])

probs = model.predict_proba(features)[0]
print(f"Home Win: {probs[0]:.4f} | Draw: {probs[1]:.4f} | Away Win: {probs[2]:.4f}")