Retail Store Sales Time Series

Level: Medium → Hard
Goal: Daily sales forecasting for multiple stores/items (panel + time series).

Dataset

• Source: Store Sales – Time Series Forecasting (Favorita) – Kaggle
• Link: https://www.kaggle.com/competitions/store-sales-time-series-forecasting/data

Download Instructions

1. Open the competition link above.
2. Go to "Data" tab.
3. Accept competition rules (if needed).
4. Download train.csv and save to data/store_sales/.

Data Loading

import pandas as pd

df = pd.read_csv("data/store_sales/train.csv")
df["date"] = pd.to_datetime(df["date"])
df = df.set_index("date").sort_index()

print(df.head())
# Columns usually include: date, store_nbr, family, sales, onpromotion, oil_price, etc.

Implementation Steps

1. Data Exploration

• Load training data
• Understand data structure:
  • Multiple stores
  • Multiple product families
  • Time series panel data
• Inspect available features (promotions, oil prices, holidays, etc.)
• Examine data quality and missing values

2. Exploratory Data Analysis (EDA)

• Aggregate sales by store, family, and time
• Plot overall sales trends
• Analyze seasonality patterns (daily, weekly, monthly, yearly)
• Identify holiday effects
• Analyze promotion effects
• Examine relationships between features
• Calculate sales statistics by store/family

3. Data Preprocessing

• Handle missing values
• Create time-based features (day of week, month, year, holidays)
• Encode categorical variables (store, family)
• Handle promotions and external variables
• Prepare panel data structure

4. Feature Engineering

• Time Features:
  • Day of week, month, year
  • Holiday indicators
  • Lag features (previous day, week, month sales)
• Rolling Statistics:
  • Rolling mean, std (7-day, 30-day windows)
  • Rolling max, min
• External Features:
  • Oil prices
  • Promotion indicators
  • Store/family characteristics

5. Model Building

• Univariate Approaches (per store-family):
  • ARIMA/SARIMA for individual series
  • Exponential smoothing
• Panel Data Approaches:
  • Hierarchical models
  • Store/family-specific models
• Machine Learning:
  • Feature-based models (Random Forest, XGBoost, LightGBM)
  • Handle panel structure appropriately
• Hybrid Approaches:
  • Combine classical and ML methods
  • Ensemble forecasts

6. Model Evaluation

• Use time series cross-validation (walk-forward)
• Split data temporally (respect time ordering)
• Calculate metrics (MAE, RMSE, MAPE)
• Evaluate at different aggregation levels (store, family, overall)
• Compare multiple approaches

7. Forecasting

• Generate forecasts for test period
• Aggregate forecasts appropriately
• Handle multiple stores/families
• Include uncertainty estimates
• Visualize forecasts

8. Advanced Analysis (Optional)

• Analyze store/family-specific patterns
• Identify important features
• Analyze promotion effectiveness
• Compare store performance
• Seasonal pattern analysis

Expected Deliverables

1. EDA Report:

   • Sales trends and patterns
   • Seasonality analysis
   • Feature analysis
   • Store/family comparisons

2. Model Results:

   • Selected model(s) with parameters
   • Performance metrics (by store/family and overall)
   • Feature importance (if using ML)
   • Forecast plots

3. Code:

   • Complete Python notebook
   • Functions for data processing
   • Feature engineering utilities
   • Modeling pipeline

Tips

• This is panel data - handle multiple stores/families appropriately
• Strong seasonality expected (weekly, monthly, yearly patterns)
• Holidays and promotions have significant effects
• Use appropriate aggregation levels for analysis
• Feature engineering is crucial for good performance
• Consider hierarchical models (store → family → item)
• Use time series cross-validation (don't shuffle time series data)
• Handle missing values carefully (forward fill, interpolation)
• External features (oil prices, promotions) can improve forecasts
• Consider store/family-specific models vs global models
• This is competition data - aim for good performance
• Document all feature engineering decisions
• Compare simple models (ARIMA) with complex models (XGBoost)