Inspiration
Slot machines appear random, but the challenge’s simulated machines have payout patterns that change over time. We wanted to use gameplay logs to estimate each machine’s next-spin probabilities while accounting for hidden state and incomplete histories.
What it does
Our model predicts the probabilities of 0×, 1×, 2×, 5×, and 20× payouts for every main-round test query. It produces a CSV and runs offline from the code and saved configuration in our submission ZIP.
How we built it
We analyzed the released training histories and validation probabilities to identify candidate lag patterns and two payout-size templates for each machine type. We combined temporal filters for machine behavior with models of changes shared across an episode, then calibrated and blended their forecasts using validation episodes. Every prediction uses only observations before the requested spin.
Challenges and what we learned
Missing spins, short histories, and changing machine behavior made the hidden state difficult to estimate. We tested several model families, lag sets, and blends, and learned that estimating each machine’s current win probability was harder than identifying its payout-size templates. Our best six-fold, episode-grouped public validation estimate was 58.564/60. The private test score is unknown.
