This version supplements the legacy Homework 1 PDF/DOCX. Use it for future offerings after the instructor supplies dates, submission logistics, and a private evaluation set.
Build and assess models for Razorback football win margin while distinguishing two different tasks:
- Pregame forecast: use only features known before kickoff.
- Postgame explanation: use game statistics to explain the observed margin, but do not call this a forecast.
Your report must include a feature-availability table. For every candidate feature, state when it is available and whether it is permissible for each task.
- Inspect the data, units/definitions, missing values, and seasons represented.
- Create a chronological train/validation/test split. Do not randomly mix future games into the training period for a forecasting claim.
- Establish a simple baseline before complex models.
- Train at least two model families, including one interpretable baseline. Document hyperparameters, seed, and compute time.
- Report MAE and one additional justified regression metric on the held-out set; include uncertainty or bootstrap intervals if feasible.
- Plot predictions, residuals, and the three largest-error cases. Explain likely causes without claiming causality from correlation alone.
- Complete an instructor-held evaluation after submitting the initial pipeline.
- Identify at least three leakage-prone features or transformations.
- Defend one included feature and one excluded feature.
- Explain whether the goal is prediction, explanation, or both.
- Use the AI-use record if generative AI is permitted.
- Test one AI-proposed improvement against held-out error and computational cost; do not credit it based only on training loss.
lastname_firstname_regression/
README.md
environment.yml or requirements.txt
analysis.py or analysis.ipynb
AI_USE.md
results/
metrics.json
figures/
Grade with the shared rubric. A short individual check may ask you to classify a feature as pregame/postgame, explain a residual, or modify the split.