Sales & Demand Forecasting for Book Titles

Technical report on hybrid time-series forecasting for book sales.

📊Project Overview

This project forecasts weekly demand for selected best-selling books over a 32-week horizon. It combines classical time-series modelling (SARIMA) with gradient-boosted trees (XGBoost) to capture both linear seasonal patterns and nonlinear effects.

📂Data Sources

• Historical Sales Data (2008–2014): Weekly transaction records (~200,000 rows)
• Books Metadata: ISBN, title, author, imprint, publisher group, product class, and category (~500 unique titles)
• Variables: Sales volume, value, average selling price (ASP), recommended retail price (RRP), binding type, and time interval
• Frequency: Weekly time series per title (aggregated from raw transactions)
• Titles: Including The Alchemist and The Very Hungry Caterpillar.

💼Business Problem

Accurate demand forecasting is critical for publishers and retailers to balance inventory, pricing, and marketing decisions.
Traditional forecasting models often underperform due to seasonal volatility, promotions, and title-specific behavior.

This project aimed to build a robust forecasting framework that predicts weekly sales 32 weeks ahead for each book title, helping the publisher:

• Reduce stockouts and overstock
• Optimize reprint timing
• Improve pricing and promotion planning

⚙️Approach & Methodology

We developed a hybrid time-series forecasting pipeline that integrates statistical and machine learning models to improve prediction accuracy, interpretability, and strategic usability.
The workflow followed five structured phases:

1️⃣ Data Preparation

• Cleaned and aligned weekly time series for each ISBN.
• Handled missing weeks using interpolation and smoothing techniques.
• Removed outliers caused by bulk promotions or anomalous spikes.
• Resampled data to a consistent weekly frequency, filling gaps with zeros.
• Book Selection: Focused modeling on two representative titles:
  • The Alchemist — steady, long-term sales pattern.
  • The Very Hungry Caterpillar — strong seasonal variation.

2️⃣ Exploratory Analysis

• Conducted time series decomposition to separate trend, seasonality, and residual components.
• Performed ACF/PACF analysis and stationarity testing to identify autocorrelation structures.
• Applied Auto ARIMA to establish an initial statistical baseline and optimal parameters for trend modeling.

3️⃣ Feature Engineering

• Generated lag features (4, 8, 12 weeks) and rolling averages to capture recent sales dynamics.
• Created time-based features (month, quarter, week number, year) to model cyclical patterns.
• Incorporated categorical metadata (category, binding, imprint) to represent product characteristics.
• Engineered trend and residual features for downstream hybrid modeling.
  

4️⃣ Modeling Framework

• SARIMA Model: Captured trend and seasonal components.
• XGBoost Regressor: Modeled nonlinear residuals left by SARIMA, improving fine-grained accuracy.
• LSTM Model: Captured temporal dependencies for sequential forecasting.
• Hybrid Models: Combined statistical and machine learning forecasts via two approaches:
  • Sequential Hybrid (SARIMA → XGBoost): XGBoost learns residual errors from SARIMA.
  • Parallel Hybrid (SARIMA + LSTM): Weighted ensemble integrating both models’ strengths.
• Tuned hyperparameters using KerasTuner and time-series cross-validation to ensure robust performance.

5️⃣ Evaluation & Aggregation

• Evaluated model performance using:
  • Mean Absolute Error (MAE)
  • Mean Absolute Percentage Error (MAPE)
  • Root Mean Squared Error (RMSE)
• Aggregated weekly forecasts to monthly summaries for executive-level decision-making and inventory management.

👩‍💻 My Role

• Led data engineering and feature pipeline design
• Implemented SARIMA and XGBoost hybrid model for 32-week forecast horizon
• Designed forecast comparison visualizations (actual vs predicted)
• Evaluated model stability and generalization across multiple titles
• Authored final forecast interpretation and business recommendations

📈 Key Findings

Best Model: SARIMA + XGBoost hybrid

Average Forecast Accuracy: 91% (MAPE < 9%)

Improvement: +18% accuracy over standalone SARIMA baseline

Forecast Horizon: 32 weeks ahead (8 months)

Performance Insights:

• Sales followed strong seasonal peaks (Q4 holiday season)
• Children’s and Fiction categories had highest volatility
• Non-fiction titles showed stable, predictable demand patterns

🔍 Analytical Insights

• Hybrid modeling effectively combined the interpretability of SARIMA with the flexibility of XGBoost
• Lag features and rolling averages improved short-term responsiveness to recent trends
• Performance metrics confirmed residual learning reduced forecast errors significantly
• Visual trend analysis revealed predictable post-holiday sales dips ideal for restocking decisions

✅Recommendations

• Automate weekly forecast updates to align with sales and marketing cycles
• Use forecasts to optimize print and distribution schedules
• Expand feature set to include promotion and weather data for richer context
• Integrate the forecasting pipeline into Power BI dashboards for continuous tracking

💡Business & Regulatory Impact

• Enabled the publisher to anticipate demand 8 months ahead
• Reduced stock imbalances and lost sales opportunities
• Provided an explainable, reproducible forecasting framework
• Improved planning collaboration between sales, printing, and supply chain teams

🔧Future Work

• Improve speaker identification for more granular sentiment.
• Strengthen topic coherence through improved embeddings.
• Validate early-warning indicators using historical backtesting.
• Develop interpretability metrics for LLM-driven summaries.

📘Conclusion

This project showcased how hybrid statistical–machine learning forecasting can transform sales planning for the publishing industry.

By leveraging historical patterns and feature-based learning, the model provided more accurate, explainable, and scalable forecasts, empowering data-driven decision-making across the organization.