Time Series Forecasting Dashboard

Multi-model time series forecasting with LightGBM, AutoTS, XGBoost, and Prophet

Production-grade forecasting dashboard demonstrating state-of-the-art time series models with automatic preprocessing, feature engineering, and model comparison.

Features

• Multiple Forecasting Models:

  • LightGBM - Gradient boosting with time-based features
  • AutoTS - Automated ensemble model selection
  • XGBoost - Alternative gradient boosting implementation
  • Prophet - Facebook's production forecasting library

• Automatic Preprocessing:

  • Missing value imputation
  • Outlier detection
  • Frequency normalization
  • Train/test splitting

• Feature Engineering:

  • Time-based features (month, day, quarter, etc.)
  • Lag features (1, 7, 14, 30 periods)
  • Rolling statistics (mean, std)
  • Difference features (rate of change)

• Model Comparison:

  • RMSE, MAE, MAPE metrics
  • Training time tracking
  • Automatic best model selection
  • Confidence intervals

• Production Ready:

  • FastAPI backend with OpenAPI docs
  • Docker containerization
  • Health checks and monitoring
  • Rate limiting
  • CORS configuration

Quick Start

Local Development

1. Install dependencies:

cd backend
pip install -r requirements.txt

2. Run the server:

uvicorn main:app --reload --port 8005

3. Access API docs:

• Swagger UI: http://localhost:8005/docs
• ReDoc: http://localhost:8005/redoc

Docker

cd backend
docker build -t timeseries-forecast .
docker run -p 8005:8005 timeseries-forecast

API Usage

1. Upload Dataset

curl -X POST "http://localhost:8005/api/upload" \
  -H "Content-Type: multipart/form-data" \
  -F "file=@your_timeseries.csv"

Expected CSV format:

date,value
2023-01-01,100.5
2023-01-02,102.3
2023-01-03,98.7
...

Response:

{
  "filename": "your_timeseries.csv",
  "rows": 365,
  "columns": ["date", "value"],
  "start_date": "2023-01-01",
  "end_date": "2023-12-31",
  "frequency": "D",
  "missing_values": 0
}

2. Generate Forecast

curl -X POST "http://localhost:8005/api/forecast?dataset_id=your_timeseries.csv" \
  -H "Content-Type: application/json" \
  -d '{
    "model_type": "all",
    "horizon": 30,
    "date_column": "date",
    "value_column": "value",
    "frequency": "D"
  }'

Model Types:

• lightgbm - LightGBM only
• autots - AutoTS only
• xgboost - XGBoost only
• prophet - Prophet only
• all - Run all models and compare (default)

Response:

{
  "results": [
    {
      "model_name": "lightgbm",
      "forecast": [
        {
          "date": "2024-01-01",
          "forecast": 105.2,
          "lower_bound": 95.3,
          "upper_bound": 115.1
        },
        ...
      ],
      "metrics": {
        "model_name": "lightgbm",
        "rmse": 3.45,
        "mae": 2.67,
        "mape": 2.3,
        "training_time_seconds": 1.23
      },
      "historical_data": [...]
    },
    ...
  ],
  "best_model": "lightgbm"
}

Model Details

LightGBM

Strengths:

• Fast training and inference
• Handles large datasets well
• Good with noisy data
• Feature importance analysis

Use Cases:

• High-frequency data (daily, hourly)
• Large datasets (>10K points)
• When speed is critical

Parameters:

• Objective: regression
• Boosting: GBDT
• Learning rate: 0.05
• Num leaves: 31

AutoTS

Strengths:

• Fully automated model selection
• Ensemble of multiple methods
• No hyperparameter tuning needed
• Built-in cross-validation

Use Cases:

• When you want "best" model automatically
• Exploring multiple approaches
• Production systems with varying data

Methods Evaluated:

• ARIMA, ETS, Prophet, LSTM
• Regression models
• Statistical models

XGBoost

Strengths:

• Robust gradient boosting
• Regularization prevents overfitting
• Handles missing values
• Parallel processing

Use Cases:

• Alternative to LightGBM
• When regularization is needed
• Baseline comparison

Parameters:

• Objective: regression
• Max depth: 6
• Learning rate: 0.05
• Subsample: 0.8

Prophet

Strengths:

• Handles seasonality automatically
• Robust to missing data
• Interpretable components
• Industry-proven (Facebook)

Use Cases:

• Business metrics (sales, traffic)
• Daily/weekly seasonality
• Holiday effects
• Trend changes

Features:

• Automatic trend detection
• Multiple seasonality
• Holiday effects
• Outlier handling

Technical Stack

• Backend: FastAPI, Python 3.11
• Models: LightGBM, AutoTS, XGBoost, Prophet
• Data: pandas, numpy, scikit-learn
• Deployment: Docker, Render.com
• Testing: pytest, httpx

Performance

Typical Training Times (1000-point dataset):

• LightGBM: ~1-2 seconds
• XGBoost: ~1-2 seconds
• Prophet: ~3-5 seconds
• AutoTS: ~10-30 seconds (evaluates multiple models)

Inference: <100ms per forecast

Deployment

Render.com

1. Fork this repository
2. Create new Web Service on Render
3. Connect repository
4. Render will auto-detect render.yaml
5. Deploy!

Environment Variables

See .env.example for all configuration options.

Required:

• PORT - API port (default: 8005)
• ENVIRONMENT - development/production

Optional:

• MAX_FILE_SIZE_MB - Max CSV upload size (default: 50)
• MAX_FORECAST_HORIZON - Max forecast periods (default: 365)
• CORS_ORIGINS - Allowed origins

Development

Project Structure

timeseries-forecast-demo/
├── backend/
│   ├── api/
│   │   ├── __init__.py
│   │   └── models.py           # Pydantic models
│   ├── forecasting/
│   │   ├── __init__.py
│   │   ├── preprocessing.py    # Data cleaning, features
│   │   └── models/
│   │       ├── lightgbm_forecaster.py
│   │       ├── autots_forecaster.py
│   │       ├── xgboost_forecaster.py
│   │       └── prophet_forecaster.py
│   ├── tests/
│   ├── main.py                 # FastAPI app
│   ├── config.py               # Settings
│   ├── requirements.txt
│   └── Dockerfile
├── render.yaml
├── .env.example
└── README.md

Testing

cd backend
pytest tests/ -v --cov=. --cov-report=html

Adding New Models

1. Create new forecaster in forecasting/models/
2. Implement: train(), forecast(), evaluate()
3. Add to main.py model selection
4. Update API docs

Use Cases

• Financial: Stock price, revenue forecasting
• Operations: Demand planning, inventory optimization
• Marketing: Traffic prediction, campaign ROI
• IoT: Sensor data, anomaly detection
• Energy: Load forecasting, consumption prediction

Limitations

• Data Requirements: Minimum 60-90 points recommended
• Frequency: Best for daily/weekly data
• Stationarity: Some models assume stationarity
• Exogenous Variables: Not supported in current version

Future Enhancements

• Multi-variate forecasting
• Exogenous variables support
• LSTM/GRU deep learning models
• Anomaly detection
• Interactive frontend dashboard
• Model persistence/caching
• Hyperparameter optimization

License

MIT

Author

Built to demonstrate production-grade ML engineering with modern forecasting techniques.

Technologies: LightGBM, AutoTS, XGBoost, Prophet, FastAPI, Docker