🛒 E-Commerce Weekly Sales Forecasting

This project analyzes the weekly sales data of a UK-based online retailer. With rising inventory costs and customer demand for fast, reliable service, it's crucial to anticipate revenue fluctuations in advance.

By forecasting weekly total sales, this model helps the business:

• 📉 Minimize overstock and understock risks
• 🚚 Improve inventory and fulfillment efficiency
• 📊 Support strategic planning with data-driven insights

---

📌 Problem Statement

Can we accurately forecast weekly sales revenue based on historical sales patterns, calendar seasonality, and recent performance trends?

---

🧠 Modeling Approach

We treat this as a supervised regression problem, using XGBoost to forecast weekly revenue. The model incorporates both short-term memory (lags, rolling means) and long-term seasonal indicators (month, week of year, spike-week flags).

---

🛠️ Workflow Overview

1. 🧹 Data Cleaning

• Filtered for valid UK transactions (removed cancellations and returns)
• Ensured accurate parsing of timestamps and numeric fields

2. 📊 Exploratory Analysis

• Ranked top products by quantity sold
• Analyzed category-level revenue and sales
• Identified and investigated a strong December spike in 2011

3. 📆 Feature Engineering

• Time-based fields: Month, Quarter, WeekOfYear
• Lag and rolling statistics: Sales_Lag_1, Sales_Lag_4, RollingMean_4, RollingSTD_4
• Seasonal flags: IsDecember, IsQ4, IsSpikeWeek
• Cyclical encodings: Month_sin, Month_cos

4. ⚙️ Modeling

• Trained XGBoostRegressor on weekly sales data
• Held out last 6 weeks for evaluation (time-aware split)

5. 📉 Evaluation

• MAE: £41,024
• RMSE: £64,712
• Captured the year-end spike trend, though slightly underestimated its magnitude

6. 🔍 Interpretation

• WeekOfYear was the most important feature
• Seasonal indicators provided useful signal
• Spike underprediction was attributed to unmodeled holiday events (e.g., promotions)

---

🔍 Visual Example

📉 The model captured trend and seasonality well, though it slightly underpredicted the final December spike.
Actual vs Predicted Weekly Sales (Final 6 Weeks)

---

✅ Why XGBoost?

• Handles non-linear relationships and feature interactions
• Supports calendar-based features, lags, and rolling statistics
• Interpretable via feature importance

---

📂 Repository Contents

File	Description
E-Commerce Sales Forecast.ipynb	Full modeling notebook
train_df_ready.csv / test_df_ready.csv	Cleaned datasets (linked externally if large)
forecast_plot.png	Visualization of model predictions
README.md	This file

---

📦 Future Improvements

• Add holiday and promotion flags (e.g., Black Friday, Cyber Monday)
• Test alternative models like Prophet or hybrid LSTM
• Extend forecasting horizon beyond 6 weeks

---

👨‍💻 Author

Shawn Waringu
Data Scientist & Analyst

LinkedIn GitHub

---