Supply Chain Forecasting with Deep Learning

Overview

This repository demonstrates a comparative study of Machine Learning and Deep Learning (GRU, LSTM) models applied to supply chain demand forecasting. The goal is to:

1. Forecast future sales by understanding historical sales trends.
2. Optimize inventory to match store stock with actual demand and reduce storage costs.
3. Optimize replenishment by determining ideal order quantities, lowering warehousing and transportation expenses.
4. Identify key sales drivers (e.g., promotions, seasonality) to enhance business strategies.

Problem Statement

• Sales Trend Forecasting: Learn from past data to predict future demand.
• Inventory Optimization: Balance stock levels to minimize excess and shortages.
• Replenishment Optimization: Compute optimal order sizes and frequencies.
• Sales Driver Analysis: Uncover features that impact sales volume.

Models Compared

Machine Learning

• Linear Regression: Fits a straight line to data; fast and interpretable but limited to linear patterns.
• XGBoost: Gradient-boosted decision trees with built-in regularization, parallel split finding, and native handling of missing values.

Deep Learning

• CNN (Convolutional Neural Network): Extracts local temporal features in time-series via 1D convolutions.
• LSTM (Long Short-Term Memory): Captures long-term dependencies in sequences with memory cells.
• CNN + LSTM: Hybrid model combining CNN’s feature extraction with LSTM’s sequence modeling.
• GRU (Gated Recurrent Unit): Simplified RNN unit addressing vanishing gradients with fewer parameters.
• Transformers: Self-attention architecture for modeling long-range dependencies without recurrence.

Key Findings

• XGBoost achieved the best performance among classical Machine Learning models with minimal hyperparameter tuning.
• The CNN + LSTM hybrid model yielded the highest forecast accuracy among Deep Learning approaches.

Advantages

XGBoost

1. Out-of-the-box performance with default settings.
2. Handles sparse and missing data natively.
3. Regularized objective to reduce overfitting.
4. Parallel training speeds up model building.
5. Feature importance for interpretability.

CNN + LSTM

1. Convolutional layers learn local temporal patterns automatically.
2. Pooling layers reduce dimensionality while retaining key information.
3. Flatten + RepeatVector transforms convolutional outputs for LSTM input.
4. Dense output layer integrates learned features into forecasts.

Getting Started

Prerequisites

• Python 3.7+
• Install required packages:

  pip install pandas numpy scikit-learn xgboost tensorflow matplotlib seaborn