Prediction Engine

How predictions are evaluated

• Every hour, build-prediction-features reads flow_snapshots and bias_index_history, writes a feature vector to prediction_features, and runs the prediction model

• The model outputs probability_up, probability_down, and direction for 4h / 24h / 72h horizons

• evaluate-predictions runs at :30 each hour. When a horizon elapses, it fetches SOL price from Binance at feature_time and feature_time+horizon, computes pct_change and actual_direction, then marks the prediction as HIT or MISS

• No lookahead bias: predictions are locked at feature_time and never updated after the model runs