Build Multi-Step Ahead Forecasts with PyTorch Forecasting and statsmodels

                      
                     
from typing import Any, Dict
import os
import pandas as pd
import torch
from pytorch_forecasting import TimeSeriesDataSet, TemporalFusionTransformer, Trainer
from statsmodels.tsa.holtwinters import ExponentialSmoothing

# Configuration
DATA_PATH = os.getenv('DATA_PATH', 'data.csv')  # CSV file path

# Load dataset
try:
    data = pd.read_csv(DATA_PATH)
    data['date'] = pd.to_datetime(data['date'])  # Ensure date is in datetime format
except Exception as e:
    raise RuntimeError(f"Failed to load data: {e}")

# Prepare the dataset for TimeSeriesDataSet
max_encoder_length = 30  # days
max_prediction_length = 15  # days

try:
    training_data = TimeSeriesDataSet(
        data,
        time_idx='date',
        target='value',
        group_ids=['id'],
        max_encoder_length=max_encoder_length,
        max_prediction_length=max_prediction_length,
    )
except Exception as e:
    raise RuntimeError(f"Error in dataset preparation: {e}")

# Initialize and train Temporal Fusion Transformer
try:
    model = TemporalFusionTransformer.from_dataset(training_data)
    trainer = Trainer(max_epochs=10)
    trainer.fit(model, train_dataloader=training_data)
except Exception as e:
    raise RuntimeError(f"Model training failed: {e}")

# Forecasting
def forecast_multi_step(model: Any, data: Any) -> Dict[str, Any]:
    try:
        predictions = model.predict(data)
        return {'predictions': predictions}
    except Exception as e:
        raise RuntimeError(f"Forecasting failed: {e}")

if __name__ == '__main__':
    result = forecast_multi_step(model, training_data)
    print(result)  # Print predictions