Hidden Edge

Inspiration

We wanted to explore how much hidden structure could be recovered from a system that appears random. House Edge challenged us to forecast simulated slot-machine payouts, then adapt when the underlying patterns changed.

What it does

Part 1 — Main Round: Hidden Edge predicts probabilities for five payouts: 0×, 1×, 2×, 5×, and 20×, using machine histories and evidence from related machines.

Part 2 — Adaptation Round: It forecasts the same outcomes on fresh data after updating its learned rules and models to reflect changed patterns.

How we built it

Part 1: We combined Bayesian rule tracking, gradient-boosted models for payout occurrence and size, and a joint probabilistic model. Cursor and Grok agents helped us explore approaches, develop code, and troubleshoot. We exported the models for offline inference using Python’s standard library.

Part 2: We reused that architecture but changed the primary lag pairs from (2,5), (4,7), and (1,3) to (3,8), (1,4), and (2,6). We re-estimated model settings and retrained the boosted models on the adaptation data.

Challenges we ran into

Part 1: Separating patterns from noise, handling missing ticks, and ensuring every prediction used only information available before its cutoff.

Part 2: Identifying which patterns had changed and updating the model within the shorter development window.

Accomplishments that we're proud of

Part 1: Building a forecasting pipeline that combines learned patterns with explicit uncertainty.

Part 2: Adapting that pipeline to new rules and producing a separate submission. Both rounds contain 1,536 valid probability forecasts with corresponding code packages.

What we learned

Part 1: Understanding hidden structure and representing uncertainty helped us build more targeted models.

Part 2: Reusable architecture made adaptation faster, but the assumptions still needed to be checked against new data. Across both rounds, AI-assisted development reinforced the need for careful validation.

What's next for Hidden Edge

We want to automate discovery of new lag rules, improve change detection, and evaluate on more independent datasets. We also envision a dashboard showing forecasts and how the inferred patterns evolve.