weekly-sales-prediction-et-model

Weekly Sales Prediction – Regression & Classification Project

🎥 Project Presentation Video

<iframe width="560" height="315" src="https://www.youtube.com/embed/j-wl1eWXPtA?si=YqVcC9vzCuyrWLRW" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen></iframe>

1. Dataset Overview

This project analyzes weekly store-level sales using the Walmart Retail Dataset.
The dataset includes sales history, store metadata, and economic indicators such as CPI, Fuel Price, Temperature, and Unemployment.
All files (train.csv, features.csv, stores.csv) were merged into a unified analytical dataset.

---

2. Exploratory Data Analysis (EDA)

A comprehensive EDA was conducted to understand structure, patterns, seasonality, and correlations.
Analyses performed:

• Weekly sales distribution
• Store-level comparison
• Department-level trends
• Correlation heatmap
• Holiday vs non-holiday
• Seasonality by Month, Week, DayOfYear
• Outliers & missing values

Key Insights:
Sales show strong seasonal behavior, high variability between stores, and significant holiday effects.

---

3. Feature Engineering

Advanced feature engineering was applied to improve predictive power.

Added Features:

• Date components: Month, Week, DayOfYear
• Interaction features: Size × Dept
• Log transformation: Log_Weekly_Sales
• KMeans clustering: Cluster_ID + cluster distances
• One-hot encoding
• Holiday flags and calendar features

Insight:
Engineered features improved the model's ability to capture store structure, seasonality, and complex interactions.

---

4. Baseline Model – Linear Regression

A simple Linear Regression model trained on the raw dataset.

Result:
High error and poor performance — appropriate as a baseline but insufficient for real prediction needs.

---

5. Improved Regression Models

With engineered features, three models were trained:

Insight:
Extra Trees delivered the best performance, effectively capturing nonlinear patterns and complex interactions.

---

6. Regression Model Export

The winning regression model (Extra Trees Regressor) was exported as:

extra_trees_winner.pkl

and uploaded to this repository.

---

7. Regression → Classification Conversion

Weekly Sales were converted into binary classes using a median split:

• Class 0 → Below median
• Class 1 → Above median

The resulting classes were balanced, enabling stable classification training.

The class distribution after converting Weekly Sales into a binary target is shown below:

---

8. Classification Models

Three models were trained:

Business Insight (FN > FP):
False Negatives (predicting low sales when actual sales are high) are far more costly due to stockouts and lost revenue.

---

Challenges & Fixes

• Extra Trees initially caused long training times, so sampling was applied during tuning.
• Duplicate holiday fields (IsHoliday_x / IsHoliday_y) required cleanup.
• Date and categorical fields caused dtype mismatches and were corrected before modeling.

Final Notes

• All work performed on Google Colab (CPU).
• Reproducibility ensured using fixed random seeds.
• Sampling was used where needed to improve training time.
• Both final models (regression + classification) are included in the repository.