← SKAY · Portfolio
Open Live App ↗GitHub
Time-Series BI · ML · Live

Walmart Sales
Forecasting

End-to-end ML pipeline across 45 stores — five models benchmarked, Optuna-tuned, a 4-page Power BI report, and a 5-page Streamlit app. Best result: 2.17% MAPE with Prophet.

2.17%
Best MAPE
45
Stores
5
Models benchmarked
4
Power BI pages

What I built

01
End-to-end pipeline
Merged 3 raw CSVs (sales, features, stores) into a weekly time series across 45 Walmart stores — no leakage, temporal split.
02
Model benchmark
Prophet, XGBoost, and SARIMAX benchmarked head-to-head. Optuna with TimeSeriesSplit CV squeezed Prophet down to 2.17% MAPE.
03
Feature engineering
Lag features, rolling statistics, calendar encodings, interaction terms — 30+ features ranked by XGBoost gain.
04
Power BI report
A 4-page report — Executive Overview · Model Comparison · Forecast Outlook & Risk · Forecast Trust & Operational Readiness — with a data model, interactive slicers, a DAX query and Python visuals.

Model benchmark

The benchmark. One winner.

BESTProphet (Optuna)MAE $1.02M · 2.17% MAPE
XGBoost (Baseline)MAE $1.22M · 2.63% MAPE
XGBoost (Optuna+FE)MAE $1.38M
SARIMAXMAE $3.30M · 7.09% MAPE

Evaluation window: May – Oct 2012 · 20% holdout · temporal split · no data leakage · Prophet + regressors (5th model) in the repo

Feature importance

Seasonality dominates. Always.

weekofyear
92%
month
81%
lag_1
68%
roll_std_4
54%
IsHoliday
38%
Temperature
27%

Forecast

Actual vs Prophet vs future

Illustrative shape — the real forecast plots are in the live app.

Actual sales
Prophet (2.17% MAPE)
Future forecast
ProphetXGBoostSARIMAXOptunaPandasscikit-learnStreamlitPower BIMatplotlibSeabornPython

See it running live.

Streamlit app — full interactive forecasting dashboard, model comparison, and future projections.

Open Streamlit App ↗View Source

Built by SKAY · Oussama Skia — AI / ML Engineer · Casablanca