We built a probabilistic forecasting system for the House Edge challenge. It predicts the probabilities of 0×, 1×, 2×, 5×, and 20× payouts using only observations before each query’s cutoff.
Our approach combines learned lag relationships, payout-size models, shared signals from neighboring machines, and flexible fallback models. We evaluated with whole episodes held out and stress-tested frozen submissions across 15 simulated scenarios, including unfamiliar rules, sparse histories, and changed payout distributions.
The final predictor runs offline on CPU using Python’s standard library and a saved checkpoint. Our main challenges were handling missing ticks, detecting changing behavior, and improving generalization without over-tuning public validation.
