Interpretable Forecasting of Realized Volatility with Temporal Fusion Transformer

With Google's Transformer forecasting model

Past variable importance over time. The Lookback window has a length of 35 datapoints. Each bar shows the normalized contribution of all historical variables for each individual time-step to generate the forecasts

Our previous article discussed Temporal Fusion Transformer (TFT) – a highly accurate and interpretable model.

Personally, I have use TFT for 3 reasons:

• Heterogeneous Inputs: TFT integrates past observed covariates, future known inputs, and exogenous static covariates.

• Probabilistic Forecasting: The original TFT included probabilistic forecasting through quantiles – but it can be further refined (e.g., conformalized quantile regression).

• Interpretability: One of the few DL models explicitly designed for interpretable forecasts.

This article uses TFT to forecast realized volatility – a key metric for assessing a stock's future risk. Additionally, we’ll leverage TFT’s advanced interpretability features to provide explainable forecasts.

Let’s get started!

✅ Find the hands-on project for this article in the AI Projects folder (Project 13), along with other cool projects!

About Realized Volatility in Forecasting

What is realized volatility and why it’s useful?