README.md

metadata

title: Store Sales Time Series Forecasting
emoji: 📈
colorFrom: blue
colorTo: green
sdk: docker
sdk_version: 1
app_file: app.py
pinned: false
short_description: Forecast store sales over 13 weeks using LSTM
license: mit
tags:
  - time-series
  - forecasting
  - lstm
  - retail
  - pytorch
library_name: pytorch
model-index:
  - name: ImprovedCashFlowLSTM
    results:
      - task:
          type: time-series-forecasting
        dataset:
          name: Kaggle Store Sales
          type: store-sales-time-series-forecasting
        metrics:
          - name: MAE
            type: mae
            value: 710.75
          - name: RMSE
            type: rmse
            value: 1108.36
          - name: MAPE
            type: mape
            value: 7.16
          - name: R2
            type: r2
            value: 0.8633

Store Sales Time Series Forecasting

Overview

This Hugging Face Space hosts an LSTM-based model (ImprovedCashFlowLSTM) for forecasting store sales over a 13-week horizon, trained on the Kaggle Store Sales Time Series Forecasting dataset.

Model Architecture

• Type: LSTM
• Input Size: 20 features
• Hidden Size: 128
• Number of Layers: 2
• Dropout: 0.2
• Forecast Horizon: 13 weeks
• Parameters: 227,441
• Features: sales, onpromotion, trend, dayofweek_sin, dayofweek_cos, month_sin, month_cos, is_weekend, quarter, sales_lag_1, sales_lag_7, sales_lag_14, sales_ma_7, sales_ma_14, sales_ratio_7, sales_ratio_14, promo_lag_1, promo_running_sum, dcoilwtico, is_holiday

Performance Metrics

• MAE: $710.75
• RMSE: $1108.36
• MAPE: 7.16%
• R² Score: 0.8633

Graphs

Loss Curve

Residual Graph

Sample Prediction

Usage

1. Synthetic Data: Click "Generate Synthetic Data" to create a 21-timestep sample dataset (sales ~$3,000–19,000) and view a 13-week forecast.
2. Custom Data: Upload a CSV with 21 timesteps and 20 features matching the above feature names.
3. Output: View forecasts in a table and Plotly graph with uncertainty bands.

Example

from inference import load_model_and_artifacts, predict
model, scaler, feature_names, config = load_model_and_artifacts()
sequences = scaler.transform(your_data[feature_names].values).reshape(1, 21, 20)
predictions, uncertainties = predict(model, scaler, sequences)
print("Predictions:", predictions)

Training Details

• Dataset: Kaggle Store Sales Time Series Forecasting
• Data Shape: (8250, 20)
• Sequences: 8217, length 21
• Preprocessing: StandardScaler, conservative outlier removal
• Epochs: 70 (early stopping)
• Optimizer: Adam with ReduceLROnPlateau
• Loss: MSE

Limitations

• Predictions are sensitive to input data quality and scaling.
• Performance may degrade on low-sales stores (<$3,000) not included in training.
• Assumes consistent feature engineering as in training.

License

MIT License