Forecasting NZ GDP With LSTM Modelling

Abstract

This thesis investigates traditional macroeconomic modelling technique’s forecasting ability of GDP growth rates against a Long Short Term Memory Neural Networks forecasting ability of GDP growth rates for the New Zealand economy. The main forecast time frame is from March 2012 to December 2019 just before Covid-19. The models use an 8-step quarterly time horizon which is two years ahead. The models use pseudo-real time data to train the models in this thesis. I use a Root Mean Squared Error (RMSE) to measure the performance of the models and compare the forecasts with a Diebold-Mariano test (DM test). After the pre-Covid-19 period I use the models to forecast the Covid-19 period from March 2020 to September 2022. Finally, I compare the LSTM models to ANZ forecasts. The research in this thesis shows that the traditional model Vector Autoregression (VAR) has the best performance across all 8-time horizons and the LSTM out performances the VARMAX model for most of the 8-time horizons for the pre-Covid-19 period. The LSTM model performed best over the Covid-19 period but the RMSE results were so high that no meaningful insights could be used from the models. Finally, the LSTM outperforms ANZ forecasting over the Covid-19 period and slightly under performs before Covid-19. This thesis contributes to the research that is just beginning in this area of testing traditional models to newly developed models and helps Economists inspire more confidence in their forecasting abilities.