Final Report

A leakage‑safe, time‑aware pipeline that forecasts daily sales per store.

Executive Summary

  • Model: XGBoost tuned with Optuna on a chronological split.
  • Test results (log scale): RMSE 0.169, MAE 0.130, R² 0.833.
  • Data coverage: 2013‑01‑01 to 2015‑07‑31; 844,338 open‑store rows; 1,115 stores.
  • Main drivers: Promotions, competition distance, weekday and store effects.

1. Problem & Approach

  • Business need: reliable daily forecasts for staffing and inventory at the store level.
  • Predict log1p(Sales) per store to stabilize variance.
  • Use a time‑aware split with the last six weeks as test to mirror real deployment and prevent leakage.

2. Data & Features

  • Built from train.csv joined with store.csv. Keep open days only. Remove Customers at training and inference time.
  • Key transforms: log1p(Sales), log1p(CompetitionDistance), one‑hot for low‑cardinality categoricals, calendar parts (year, month, week, weekday, weekend).
  • Useful engineered signals: a leakage‑safe Store_AvgSales (computed only on the training window), a small set of promo × store/assortment interactions, and ActiveStoreCount for daily coverage.

3. Key EDA Findings

  • Promotions work. Promo days shift the sales distribution higher and lift means.
  • Promo2 underperforms. Stores flagged with Promo2 trail non‑Promo2 stores on average.
  • Competition proximity often helps. Shorter distances correlate with higher sales, consistent with dense retail zones.
  • Store heterogeneity matters. Type b stores are high‑performing but more volatile; others are steadier.
  • School holidays show a measurable effect; state holidays mostly imply closures and were dropped.

alt text alt text alt text alt text

4. Validation & Modeling

  • Split: final six weeks (2015‑06‑19 → 2015‑07‑31) as test; earlier period for train/validation; no date overlap.
  • Retrained on train+val and evaluated once on the holdout.
  • Result: RMSE 0.169, MAE 0.130, R² 0.833 on the log scale.

alt text alt text

5. What Drives the Forecast

  • Promotions are the strongest, most consistent signal across importance, permutation, and SHAP.
  • Competition distance ranks among the top drivers; closer stores often see higher predicted sales.
  • Calendar/store effects (weekday and store baselines) provide steady lift.

alt text alt text alt text

6. Implications

  • Promo planning: plan staffing and inventory around proven promo uplift; measure by store and weekday. Treat Promo2 cautiously.
  • Operations: accuracy is strong enough for schedule and inventory planning; add a buffer for peak days.
  • Location signal: near‑competitor stores can thrive; proximity often proxies for foot traffic.

7. Next Steps

  1. Add prediction intervals so planners see a range, not just a point.
  2. Automate a nightly batch run that writes store‑day forecasts to CSV for the dashboard.
  3. Set a simple retraining cadence (for example, quarterly) and monitor a few basic checks like promo mix and competition distance.
  4. Create a lightweight dashboard to compare forecasts vs. actuals by store and day.

Justin Castillo · Email: jcastillo.hotels@gmail.com | GitHub: justin-castillo | LinkedIn: linkedin.com/in/justin-castillo-69351198