Dynamic Sales Forecasting (LSTM)

An end-to-end deep learning solution for high-precision demand forecasting using the Kaggle Store Item Demand dataset. This project implements a Long Short-Term Memory (LSTM) network in PyTorch to capture multi-year seasonal trends and daily sales fluctuations.

📊 Project Performance

Based on the training execution for Store 1, Item 1:

• Best Validation Loss (MSE): 0.0141
• Final Training Loss: 0.0120
• Model Accuracy: ~88.5% (calculated via MAPE)
• Optimization: The model achieved convergence within 30 epochs with minimal overfitting, indicating strong generalization to future time steps.

🛠️ Step-by-Step Methodology

1. Step 1

The objective was to transform historical daily sales into a supervised learning problem. We defined a 30-day sliding window to predict the following day's sales, focusing on capturing both short-term noise and long-term seasonality.

2. Step 2 (Data Engineering)

• Time-Series Windowing: Converted flat transactional records into 3D tensors [Samples, TimeSteps, Features].
• Normalization: Applied MinMaxScaler to scale sales units between 0 and 1, ensuring stable gradient flow in the LSTM cells.
• Sequential Splitting: Implemented a non-shuffled 80/20 split to maintain the chronological integrity of the data and prevent "future leakage."

3. Step 3 (Architecture)

The architecture utilizes a Stacked LSTM:

• Layer 1: Captures low-level temporal dependencies (day-of-week effects).
• Layer 2: Extracts high-level seasonal patterns (monthly/yearly trends).
• Fully Connected Head: Maps the 64-dimensional hidden state to a single continuous sales value.

4. Step 4 Systematically

• Walk-Forward Validation: Evaluated the model on the most recent 20% of the timeline.
• Loss Function: Used MSELoss to heavily penalize large forecasting errors, which is critical for inventory management safety stocks.

5. Step 5 (Interactive Deployment)

Created a Streamlit Dashboard that provides:

• Visual Backtesting: Overlays "Actual" vs "Predicted" lines to verify model fit.
• Accuracy Metrics: Real-time calculation of MAPE and MAE.

📂 File Structure

• data/: Contains the train.csv.
• data_processor.py: Sliding window transformation and scaling logic.
• model.py: PyTorch LSTM class definition.
• train.py: The training loop and walk-forward validation logic.
• main.py: Entry point for orchestrating the pipeline.
• app.py: Interactive Streamlit dashboard.

🚀 Getting Started

1. Clone the Repository:

git clone https://github.com/devopspower/dynamic-sales-forecasting.git

2. Install Dependencies:

pip install -r requirements.txt

3. Run the Training Pipeline:

python main.py

4. Launch the Dashboard:

streamlit run app.py

📈 Business Insights

• Weekly Seasonality: The model successfully identified consistent weekend peaks and weekday valleys.
• Inventory Optimization: High accuracy on validation data allows for a reduction in "Safety Stock" levels, directly impacting the bottom line by reducing storage costs.