Separating Participation from Puzzle Difficulty

For the Mathematical Contest in Modeling, I helped build a pipeline that treated Wordle participation, attempt distributions, and puzzle difficulty as distinct modeling tasks. ARIMA described participation over time, a neural network mapped word features to attempt distributions, and K-means grouped words by difficulty.

Feature engineering and time-series diagnostics connected each model to a specific question, producing forecasts and difficulty profiles from the same small dataset.

Reading the Results Together

The participation fit reached R² = 0.982. Attempt-distribution prediction was less consistent, with reported RMSE of 3.57 and MAPE of 45.28%, so the results were interpreted alongside difficulty clusters rather than reduced to one headline accuracy.

The submitted report combined participation forecasts, predicted attempt distributions, and three difficulty groups. The entry received Successful Participant recognition.