​
Overview

This section provides tutorials about training TimeGPT under specific conditions. Learn how to extend predictions across multiple time series and over long horizons with ease.

TimeGPT is designed to handle time series forecasting tasks of varying complexities. The tutorials below will guide you through key strategies for effective training and deployment.

​
Quick Start: General Training Steps

Below is a concise overview of how to start training with TimeGPT.

1

Prepare Your Data

Ensure your time series data is clean, properly formatted, and includes all necessary features (e.g., timestamps, values, external variables).

2

Select Your Forecasting Approach

Decide whether you need a single series or multi-series approach, and whether you need short or long horizons.

3

Configure TimeGPT

Set up the relevant hyperparameters for your forecasting needs (window size, horizon, seasonalities, etc.).

4

Train and Evaluate

Train the model on your dataset and evaluate performance with appropriate error metrics (MAPE, RMSE, etc.).

5

Refine and Deploy

Use performance insights to refine your model, then deploy it in a production environment.

​
Tutorials

​
Example Training Code

Below is a simplified example of how you might train TimeGPT in Python. Adjust hyperparameters as needed for your specific use case.

# Example: Training TimeGPT for a multi-series scenario
import timegpt

# Load your dataset
data = timegpt.load_data("my_dataset.csv")

# Initialize the model
model = timegpt.TimeGPT(
    horizon=30,           # Forecast horizon
    max_epochs=50,        # Number of training epochs
    seasonality=24,       # Seasonal period
    learning_rate=1e-3
)

# Train the model
model.fit(data)

# Generate predictions
predictions = model.predict(data)

# Evaluate results
error = model.evaluate(predictions, data)

print(f"Forecast error: {error:.2f}")

Congratulations! You now have an overview of how to set up and train TimeGPT for both single and multiple series forecasting as well as for long-horizon use cases.