📈 Walmart Store #1 Weekly Sales Forecasting

Table of Contents

• Project Overview
• Data & Sample
• Installation
• Feature Engineering
• Modeling
• Results

---

Project Overview

This repository demonstrates an end-to-end feature engineering pipeline and XGBoost modeling workflow focused on forecasting weekly sales for Store #1 only.

All work including previously done feature engineering in this project was performed by Tez1s.

It includes:

• Loading pre-processed data for Store #1
• A suite of engineered features capturing seasonality, holidays, rolling trends, and interactions
• Training an XGBoost regressor with time‑aware cross‑validation
• Evaluation of model performance in terms of RMSE and MAE

Note: All feature engineering was performed by me prior to this project; here you see only the downstream pipeline (features → model).

---

Data & Sample

Data used in this repo is already cleaned and filtered to Store #1. After feature engineering, each row represents one week’s record:

Date	Weekly_Sales	Holiday_Flag	Temperature	Fuel_Price	CPI	Unemployment	Month	is_holiday	pre_holiday	post_holiday	Weekly_Sales_MA30	holiday_sales_impact	temp_fuel_interaction	Day	Lag1_Day
2010-02-26	1,409,727.59	0	46.63	2.561	211.319643	8.106	2	0	0	0	1,576,836.00	0.0	119.4194	26	NaN
2010-03-05	1,554,806.68	0	46.50	2.625	211.350143	8.106	3	0	0	0	1,554,615.00	0.0	122.0625	5	26.0
2010-03-12	1,439,541.59	0	57.79	2.667	211.380643	8.106	3	0	0	0	1,504,011.00	0.0	154.1259	12	5.0
2010-03-19	1,472,515.79	0	54.58	2.720	211.215635	8.106	3	0	0	0	1,469,148.00	0.0	148.4576	19	12.0
2010-03-26	1,404,429.92	0	51.45	2.732	211.018042	8.106	3	0	0	0	1,467,823.00	0.0	140.5614	26	19.0

---

Installation

1. Clone the repository

   git clone https://github.com/Tez1s/walmart_xgb_cv.git
   cd walmart_xgb_cv
   

Feature Engineering

All features have been pre‑computed and bundled into final_walm.csv:

• Rolling Mean (30‑day)

  • Weekly_Sales_MA30: 30‑day moving average of Weekly_Sales.

• Holiday Effects

  • is_holiday: 1 if the week contains a holiday, else 0
  • pre_holiday / post_holiday: Flags for the weeks immediately before and after a holiday
  • holiday_sales_impact: Difference between actual sales and the rolling mean during holiday weeks

• Interaction Terms

  • temp_fuel_interaction: Product of Temperature and Fuel_Price to capture joint effects

• Lag Features

  • Lag1_Day: Sales from the previous week (weekly lag)

• Date Components & Economic Indicators

  • Month
  • Day
  • Holiday_Flag (original Kaggle flag)
  • Temperature, Fuel_Price, CPI, Unemployment

---

Modeling

Algorithm XGBoost Regressor (xgboost.XGBRegressor)

• base_score=0.5
• booster='gbtree'
• n_estimators=1000
• early_stopping_rounds=50
• max_depth=3
• learning_rate=0.01

Cross‑Validation

• TimeSeriesSplit with n_splits=5, ensuring each validation fold only trains on past data
• Each fold trains the model on an increasing subset of data, maintaining temporal order

Evaluation Metrics

• RMSE (Root Mean Squared Error)
• MAE (Mean Absolute Error)

---

Results

The XGBoost model significantly outperformed traditional time series methods in both RMSE and MAPE metrics, showcasing its strength in capturing complex patterns in Store #1’s weekly sales data.

Model	RMSE	MAPE
SARIMAX	38499.72	2.07%
Prophet	64,206.78	3.30%
XGBoost	19,274.35	1.13%
Model Performance:	19,274.35	1.13%

• 📉 RMSE: 19,274.35
• 📈 MAPE: 1.13%

These results indicate strong predictive performance, with the model capturing weekly sales trends with high accuracy and minimal percentage error.

Show Your Support

If you found this project helpful or interesting, please consider giving it a ⭐️!
Your support encourages further development and helps others discover this work.