Retail Demand Forecasting

Preview mode. Browse milestones and edit starter code here. Sign in to run and submit.

0/8EDA

Python

advanced~8h8 milestones

The problem

A retailer plans purchasing a quarter ahead. Buyers need a daily unit forecast for every store-item pair so they can commit to orders before the season starts. Forecast low and shelves go empty. Forecast high and cash sits in a warehouse. The forecast has to be produced for the 90 days after the last day of history, with no access to sales in that window.

Your task · Read the Panel

Establish the shape of the panel and the spread between series.

Implement explore_demand(train_df) returning a dict with:

row_count

series_count: number of distinct store_id and item_id pairs

first_date and last_date: as strings

mean_units: mean units_sold, rounded to 5 places

promo_share: fraction of rows on promotion, rounded to 5 places

quietest_series_mean and busiest_series_mean: the smallest and largest per-series mean units_sold, rounded to 5 places

The gap between the last two numbers is why RMSE alone will not tell you whether this forecast is any good.

Constraints

▪The test window is the 90 days immediately after training history, so a random split leaks the future
▪Series levels range from about 1 unit per day to over 100, so a single scale-dependent metric hides most of the error
▪Promotions and prices for the forecast window are known in advance and are given as features. Sales are not.
▪No lagged sales are available inside the forecast window
▪Hidden labels are scored server-side

Graded on

Seasonal naive baseline: RMSE <= 18.5
Model: RMSE <= 15.0 and sMAPE <= 0.52
Calibrated model: absolute bias <= 0.45 units with RMSE <= 14.5
Full 90-day forecast produced within 10 seconds

Evaluated server-side against a hidden test set.

Skills you'll demonstrate

Time SeriesBacktestingLeakage PreventionFeature EngineeringGradient BoostingForecast CalibrationError Metrics