Timer-XL Revisited: A Hands-On Guide to Zero-Shot Forecasting

Two practical use cases of this pretrained model in action

Image created with Dalle*3

Part 1 discussed Timer-XL in-depth, and what makes this model unique.

In short, Timer-XL is a milestone because it provides:

1. Unified forecasting: One model handles varying context and prediction lengths.

2. Time Attention: A novel attention mechanism that avoids the flaw in standard attention for time series (which assumes permutation invariance of observations) and distinguishes the different covariates— while preserving permutation-equivalence.

Fortunately, the univariate pretrained version of Timer-XL is open-source!

This article walks through 2 tutorials on using Timer-XL for:

• Long-context forecasting on the ETTh2 dataset — with rolling forecast for a more rigorous evaluation.

• Fun case: Forecasting financial time series (S&P 500).

Let’s get started!

✅ Find the 2 Timer-XL notebooks in the AI Projects folder (Project 18 and Project 19)

ETTh2 Forecasting

Timer-XL is a pretrained model, so there are only a few public datasets available that avoid data leakage.