Forecasts that survive a shock
Which forecasting approaches stay accurate when holidays, promotions and the store environment change?
The chart
01 / Symptom
Retailers plan pricing, inventory and staffing on forecasts, but a model trained on the past can fail once conditions change.
02 / Tests
Walmart's public data: 421,570 weekly records across 3,331 store-department series. Seven models, from a seasonal baseline to a neural network, each tested under four market conditions.
03 / Diagnosis
Still running. Data issues are already fixed: 1,358 negative or zero sales, markdowns recorded only from November 2011, and a holiday flag that misses the peak week.
04 / Prescription
Due once the models are scored: the most reliable approach for inventory and staffing decisions.
05 / Outcome
The research plan was presented in October 2026. Results aren't in yet, so none are shown.
How the tests ran
- Merge Walmart's public sales, store and feature files: 421,570 weekly records across 3,331 store-department series, February 2010 to October 2012.
- Fix data issues before modeling: 1,358 negative or zero sales, markdown fields recorded only from November 2011, and a holiday flag that misses the peak shopping week.
- Compare seven models from a seasonal naive baseline to exponential smoothing, ARIMA, regression, SVR, random forest and a neural network, plus logistic regression for high-demand weeks.
- Test each model under four conditions: a stable period, a fuel price spike, a holiday shock with the first markdowns, and a new store segment. Robustness is the rise in WMAE from the stable period.
Tools
- SAS Studio
- SAS OnDemand
- Python
- WMAE / RMSE / MAPE
Findings
421,570
weekly sales records
143
weeks, Feb 2010 to Oct 2012
4
market conditions planned
The midterm research plan was presented in October 2026. Model results are not in yet, so none are shown here.