Project Overview

Retail demand forecasting system built using XGBoost, LightGBM and Random Forest on 135,000+ aggregated retail sales records. Features include lag variables, rolling statistics, calendar effects, event indicators and pricing signals. Forecasts are converted into inventory recommendations and explained using SHAP feature attribution.

Total Forecast Demand
—
units across all stores
Average Daily Forecast
—
per store-dept combo
High Risk Inventory
—
combos need review
Best Model RMSE
89.40
XGBoost · lowest error

Business Insights

4 KEY FINDINGS
Demand Forecast
Forecasted demand exceeds 1.3M units for the next planning cycle across all stores.
Inventory Risk
1,204 store-department combinations flagged for immediate inventory review.
Best Model
XGBoost achieved RMSE 89.40 and MAPE 13.95% — best among all tested models.
Key Drivers
Rolling 7-day average + lag features account for 54% combined importance.

ML Pipeline

6 STAGES
Raw Sales Data
Feature Engineering
XGBoost Training
Forecast Generation
SHAP Explainability
Inventory Optimization
Python
XGBoost
LightGBM
SHAP
Pandas
Chart.js

135K+

Records

31

Features

70

Series

1941

Days

Forecast & Risk Analysis

DEMAND TRENDS
28-Day Forecast Trend
Aggregated daily demand across all stores
Inventory Risk Distribution
Store-department breakdown

Model Performance

BENCHMARKS
Model Comparison
RMSE · MAE · MAPE across all models
Feature Importance
XGBoost gain-based · top 10

Project Details

MODEL SUMMARY
Best Model
XGBoost
RMSE
89.40
MAPE
13.95%
Forecast Horizon
28D

SHAP Explainability

FEATURE ATTRIBUTION
SHAP Summary Plot
Generated from XGBoost predictions
SHAP Summary Plot
Mean |SHAP| Values
Top 10 features by average impact

Recent sales history (rolling means + lags) contributes most strongly to predictions, confirming strong autocorrelation in retail demand.

Inventory Recommendations

ALL RISK LEVELS
Store Department Date Forecast Sales Safety Stock Rec. Inventory Risk Level