Forecasting Book Sales to Guide Investment Decisions
Publishers face the challenge of anticipating future sales to optimize inventory, procurement, and marketing investments. This project developed and compared multiple forecasting models on historical book sales data to support those decisions.
The workflow began with basic time-series analysis such as seasonal decomposition and autocorrelation analysis. I then implemented a range of models — from Auto-ARIMA to XGBoost and Long Short-Term Memory (LSTM) neural networks — to fit the data and generate future predictions.
Results showed that advanced models outperformed classical methods, with XGBoost reducing forecast error significantly compared to Auto-ARIMA, especially in capturing complex patterns.
Tech Stack
- Models: Auto-ARIMA, XGBoost, LSTM
- Libraries: Python, Pandas, Statsmodels, Scikit-learn, TensorFlow/Keras
- Hyperparameter Tuning: Keras Tuner, Grid Search