Retail Sales Analysis Project

Overview

This project analyzes and forecasts retail store sales using historical sales data and store-level metadata.
The dataset combines daily sales records with store attributes such as promotions, assortment type, competition, and more.

This project demonstrates the full ML lifecycle: data cleaning, feature engineering, exploratory data analysis, modeling, evaluation, and deployment via a Streamlit dashboard.

⚠️ Note: Project was developed as a 7-day and continues to be updated for enhancements.

---

Project Progress

Day 1 – Data Loading & Initial Inspection

• Loaded train (daily sales) and store datasets using Python (pandas).
• Merged datasets on Store ID.
• Converted Date column to datetime and inspected data types.
• Checked for missing values, duplicates, and basic statistics.
• Saved merged and cleaned dataset for further processing.

---

Day 2 – Data Cleaning & Feature Engineering

• Handled missing values for:
  CompetitionDistance, CompetitionOpenSinceMonth, CompetitionOpenSinceYear, Promo2SinceWeek, Promo2SinceYear, PromoInterval
• Standardized StateHoliday (categorical → numeric)
• Extracted time-based features: Year, Month, Day, WeekOfYear, DayOfYear
• Created competition features: CompetitionOpenDate, DaysSinceCompetitionOpen
• Created promotion feature: IsPromoMonth
• One-hot encoded StoreType and Assortment
• Final dataset: 29 ML-ready features

Exported as → cleaned_dataset.csv

---

Day 3 – Exploratory Data Analysis (EDA)

• Focused on open stores (Open == 1)
• Generated 7 visualizations highlighting:
  • Promotions increase sales by ~39%
  • December is the peak month
  • Store type "b" outperforms others
  • Weekday trends: Monday strongest, Sunday weakest
  • Sales ↔ Customers correlation: 0.82
• Saved visualizations in /images

Added → day3_explanations.txt
Notebook → 3_eda.ipynb

---

Day 4 – Modeling

• Removed closed-store rows and fixed PromoInterval logic

• Dropped irrelevant columns (Date, CompetitionOpenDate)

• Training dataset: 844,392 rows × 26 features

• Models trained:

  • Linear Regression → RMSE: 1295, MAE: 940, Score: 82.6%
  • HistGradientBoostingRegressor → RMSE: 929, Score: 91.0%
  • XGBoost (tuned) → RMSE: 870, Score: 93.6%

• Visualizations produced: feature importance, residuals, learning curves, actual vs predicted

Notebook → 4_modeling.ipynb

---

Day 5 – Evaluation & Diagnostics

• Performed error analysis and residual checks
• Fixed feature mismatches between train and test sets
• Visualized actual vs predicted sales for key stores
• Configured .gitignore and serialized XGBoost model using joblib

Notebook → 5_evaluation.ipynb

---

Day 6 – Production Pipeline

• Built a robust prediction pipeline:
  • preprocess() → prepares input data
  • predict_sales() → runs XGBoost predictions
• Generated output artifacts:
  • results_predictions.csv → actual vs predicted sales
  • metrics.csv → RMSE, MAPE (~12%), mean error
  • top_best_predictions.csv / top_worst_predictions.csv
• Pipeline ensures scalability, modularity, and real-time readiness

Pipeline code → day6_pipeline.py

---

Day 7 – Streamlit Dashboard

• Created a fully interactive dashboard for retail sales forecasting:
  • Upload your own CSV or use the sample dataset (100 rows)
  • Run predictions with one click
  • Visualize actual vs predicted sales, error distributions, best/worst predictions
  • Download predictions CSV for further analysis
• Dashboard uses Plotly and Matplotlib/Seaborn for interactive and static visualizations

App code → 7_app.py
Demo & repo → GitHub Link

---

Tech Stack

• Python: pandas, numpy, matplotlib, seaborn, scikit-learn, XGBoost, joblib
• Jupyter Notebook
• Streamlit for interactive dashboard
• Git & GitHub

---

Dataset Credit

• Rossmann Store Sales (Kaggle)
• Used for learning and analysis purposes only.

---

Next Steps

• Feature engineering v2 (lags, rolling windows)
• Multi-step forecasting
• Advanced models: LSTM, Prophet, CatBoost
• Deployment improvements and cloud integration

---

Quick Start

1. Clone the repository:

git clone https://github.com/y-india/retail-sales-analysis-project.git

2. Navigate to the project folder:

cd retail-sales-analysis-project

3. Install dependencies:

pip install -r requirements.txt

4. Run the Streamlit app:

streamlit run 7_app.py

5. Use the dashboard to upload your CSV or try the included sample dataset (test_dataset_100.csv)