GenGNAR: genetic graph search for GNAR models with applications to inflation forecasting
File(s)
Author(s)
Estan-Ruiz, Sergio
Nason, Guy
Type
Working Paper
Abstract
Forecasting inflation accurately is central to monetary policy, yet statistical forecasting in economics often relies either on simple univariate models or on heavily
parametrised multivariate models such as VARs. Recent work on Random Generalised Network Autoregressive (RaGNAR) models showed that graph-based time series
models can provide accurate and timely inflation forecasts, but networks that they are based on are random. This paper introduces GenGNAR, a genetic algorithm framework for graph selection in GNAR models that replaces random exploration with a directed search based on forecasting performance validation. We apply the method to UK CPI inflation forecasting and show that GenGNAR yields systematic improvements over RaGNAR, with gains that are particularly pronounced at medium and longer forecasting horizons. We also study several variants of the framework,
including changing the graph generation distribution to a Stochastic Block Model to favour important nodes, and compare the best-performing specification against
standard statistical, machine learning, and official Bank of England forecasts. The results suggest that directed graph search can substantially improve the empirical
performance of GNAR models in macroeconomic forecasting.
parametrised multivariate models such as VARs. Recent work on Random Generalised Network Autoregressive (RaGNAR) models showed that graph-based time series
models can provide accurate and timely inflation forecasts, but networks that they are based on are random. This paper introduces GenGNAR, a genetic algorithm framework for graph selection in GNAR models that replaces random exploration with a directed search based on forecasting performance validation. We apply the method to UK CPI inflation forecasting and show that GenGNAR yields systematic improvements over RaGNAR, with gains that are particularly pronounced at medium and longer forecasting horizons. We also study several variants of the framework,
including changing the graph generation distribution to a Stochastic Block Model to favour important nodes, and compare the best-performing specification against
standard statistical, machine learning, and official Bank of England forecasts. The results suggest that directed graph search can substantially improve the empirical
performance of GNAR models in macroeconomic forecasting.
Date Issued
2026-06-15
Citation
2026
Copyright Statement
© 2026 The Author(s).