SaleFore AI 📈 – Ultra-Accurate, GPU-Optimized Sales Forecasting Suite

"Turning raw retail data into crystal-clear sales foresight."
🎯 Powered by: XGBoost, LightGBM, CatBoost, Prophet, Optuna, and CUDA
💼 Engineered by: Vishnupriyan P R

---

🚀 Overview

SaleFore AI is a modular, production-ready, GPU-accelerated machine learning system that delivers highly accurate sales forecasting for retail and e-commerce businesses. With automated preprocessing, feature engineering, model training, and Power BI-ready outputs — this suite was built to replace traditional forecasting pipelines with intelligent automation and accuracy.

---

🧠 Key Features

• 🔁 Automated Feature Engineering – Temporal, cyclical, rolling, lag, and store-level features
• 🧠 Ensemble Models – Uses XGBoost, CatBoost, LGBM, RF, ElasticNet, Prophet
• ⚙️ Hyperparameter Tuning – Smart optimization using Optuna
• ⚡ GPU Support – Full GPU-acceleration (RTX tested)
• 🧪 Time Series Validation – Non-leaky splits using TimeSeriesSplit
• 📦 Modular Pipeline – Train once, reuse with ease; full model persistence
• 📊 Power BI Compatible Output – Exported forecasts ready to be visualized
• 🔒 Failsafe CPU fallback – No GPU? No problem.

---

📁 Project Structure

SaleFore-AI/
├── data/
│   ├── raw/               # Unprocessed datasets
│   ├── processed/         # Cleaned and feature-rich data
│   └── forecasts/         # Final forecast results
├── models/                # Trained models and artifacts
├── outputs/               # Metrics, logs, plots, evals
├── src/
│   ├── config.py          # Constants, file paths, parameters
│   ├── preprocess.py      # All cleaning + feature engineering logic
│   ├── train_model.py     # Model training and tuning
│   ├── forecast.py        # Inference pipeline for future dates
│   └── utils.py           # Common utilities
└── requirements.txt       # Dependencies list

---

🛠️ Installation

git clone https://github.com/vishnupriyanpr183207/SaleFore-AI
cd SaleFore-AI
pip install -r requirements.txt

---

🔧 Configuration (in config.py)

Set paths for:

• RAW_DATA_PATH, PROCESSED_DATA_PATH
• FORECAST_OUTPUT_PATH
• MODEL_SAVE_DIR, etc.

Also configure:

• GPU flags
• Forecasting horizon
• Train/val/test split logic
• Feature flags (e.g., enable/disable Prophet)

---

🔄 ML Pipeline

1. Data Preprocessing (preprocess.py)

• Handles missing data, removes outliers
• Generates rolling mean, lag, day-of-week, month, holiday, promo, event, etc.

2. Model Training (train_model.py)

• Trains multiple models in parallel (LightGBM, CatBoost, Prophet, XGBoost, ElasticNet, RF)
• Optimizes using Optuna with early stopping
• Exports: model.pkl, metrics.json, importance.png

3. Forecasting (forecast.py)

• Loads trained models
• Predicts N future time steps
• Saves output in Power BI-compatible CSV

---

📈 Output

outputs/
├── forecast_results.csv         # Final predictions
├── model_metrics.json           # Accuracy, RMSE, MAPE, etc.
├── plots/
│   ├── feature_importance.png
│   └── prediction_vs_actual.png

---

📊 Dashboard Integration

1. Import forecast_results.csv in Power BI or Tableau
2. Connect model metrics
3. Build dynamic trend visualizations & what-if simulators

---

🗂️ Project ER Diagram

erDiagram
    RAW_DATA ||--o{ PROCESSED_DATA : cleaned_from
    PROCESSED_DATA ||--o{ FEATURES : generates
    FEATURES ||--|| CONFIG : uses
    FEATURES ||--o{ MODEL_TRAINING : input_for
    MODEL_TRAINING ||--|{ MODELS : generates
    MODEL_TRAINING ||--o{ VALIDATION_RESULTS : outputs
    MODELS ||--o{ FORECASTS : used_for
    CONFIG ||--|| FORECASTS : controls
    FORECASTS ||--o{ PLOTS : visualized_as
    FORECASTS ||--o{ METRICS : evaluated_by
    METRICS ||--|| PLOTS : illustrated_with

    RAW_DATA {
        string store_id
        string date
        int sales
        int promo_flag
    }

    PROCESSED_DATA {
        string store_id
        string date
        float clean_sales
        int is_holiday
        string weekday
    }

    FEATURES {
        float rolling_mean
        float lag_7
        int day_of_week
        bool is_event
    }

    CONFIG {
        string model_type
        int forecast_horizon
        string gpu_flag
    }

    MODEL_TRAINING {
        datetime start_time
        datetime end_time
        string model_name
    }

    MODELS {
        string file_path
        string format
        datetime trained_on
    }

    FORECASTS {
        date future_date
        float predicted_sales
        string model_used
    }

    METRICS {
        float rmse
        float mape
        float r2_score
    }

    PLOTS {
        string file_path
        string type
    }

Loading

---

Made with 💓 Vishnupriyan P R
📧 priyanv783@gmail.com
🔗 GitHub: vishnupriyanpr

“Tools should disappear into the background and let you build.” — Vishnupriyan P R, caffeinated coder ☕

📜 License

MIT License – use, share, and scale freely.