Built an end-to-end forecasting pipeline for 30,490
product store combinations from a 58M row Walmart (M5)
dataset, staging raw data in SQL and generating LightGBM demand forecasts across
5 business levels: company, category, department, store, and item.
Diagnosed a crossed (non-nested) hierarchy structure, where product and geography rollups
intersect rather than nest. That structure is the reason a naive bottom-up rollup doesn't
reconcile cleanly at every level.
12.5%
Improvement in company-wide forecast accuracy
0%
Mismatch across all hierarchy levels (exact coherence)
vs. up to 2.5% unreconciled
30,490
Product store combinations forecast
5
Business levels reconciled
company → item
Applied MinT reconciliation to enforce forecast consistency across all levels, against up to
2.5% inconsistency in the
unreconciled forecasts.