Cross-Frequency Time Series Meta-Forecasting

TL;DR

This paper introduces the Continuous Frequency Adapter (CFA), a novel meta-forecasting model that learns frequency-invariant representations, enabling improved generalization across different sampling frequencies in time series forecasting.

Contribution

The paper proposes CFA, a new model that overcomes the limitation of previous meta-forecasting methods by effectively handling multiple sampling frequencies.

Findings

CFA significantly improves performance on unseen frequencies.

CFA enables forecasting over larger multi-frequency datasets.

Meta-forecasting can be extended to multiple frequencies with the new model.

Abstract

Meta-forecasting is a newly emerging field which combines meta-learning and time series forecasting. The goal of meta-forecasting is to train over a collection of source time series and generalize to new time series one-at-a-time. Previous approaches in meta-forecasting achieve competitive performance, but with the restriction of training a separate model for each sampling frequency. In this work, we investigate meta-forecasting over different sampling frequencies, and introduce a new model, the Continuous Frequency Adapter (CFA), specifically designed to learn frequency-invariant representations. We find that CFA greatly improves performance when generalizing to unseen frequencies, providing a first step towards forecasting over larger multi-frequency datasets.