Performance of long short-term memory artificial neural networks in nowcasting during the COVID-19 crisis

TL;DR

This study evaluates the effectiveness of LSTM neural networks in macroeconomic nowcasting during COVID-19, showing they often outperform traditional models and introducing interpretability methods with accessible software tools.

Contribution

The paper compares LSTM performance with the dynamic factor model during COVID-19 and introduces an interpretability methodology with multi-language software support.

Findings

LSTM outperformed DFM in two-thirds of cases based on error metrics.

LSTM provided more stable and consistent forecast revisions.

A new interpretability method for LSTMs is developed and shared in multiple programming languages.

Abstract

The COVID-19 pandemic has demonstrated the increasing need of policymakers for timely estimates of macroeconomic variables. A prior UNCTAD research paper examined the suitability of long short-term memory artificial neural networks (LSTM) for performing economic nowcasting of this nature. Here, the LSTM's performance during the COVID-19 pandemic is compared and contrasted with that of the dynamic factor model (DFM), a commonly used methodology in the field. Three separate variables, global merchandise export values and volumes and global services exports, were nowcast with actual data vintages and performance evaluated for the second, third, and fourth quarters of 2020 and the first and second quarters of 2021. In terms of both mean absolute error and root mean square error, the LSTM obtained better performance in two-thirds of variable/quarter combinations, as well as displayed more gradual forecast evolutions with more consistent narratives and smaller revisions. Additionally, a methodology to introduce interpretability to LSTMs is introduced and made available in the accompanying nowcast_lstm Python library, which is now also available in R, MATLAB, and Julia.