Real-time GDP nowcasting in New Zealand : an ensemble machine learning approach : a thesis presented for the degree of Master of Philosophy, School of Natural and Computational Sciences, Massey University, New Zealand

Abstract

Gross Domestic Product (GDP) measures the monetary value of all final goods and services that are produced in a region during a period of time. For most countries, GDP is released a limited number of times a year and often with a lag. Understanding the current economic situation, instead of figures quarters ago, is of vital importance for both policy and private entrepreneurs. It is crucial to create a live GDP predictor that could Nowcast current GDP growth rate in the period of government statistics release delay. The Econometric approach for GDP Nowcasting has dominated the forecasting area for many years. However, most of the traditional econometric models could only incorporate a small handful of variables with a linear model structure, which could not meet the requirement of the “big data” era for a better model prediction ability with a large amount of unbalanced variables. With the improvement of computation ability and the increment of high frequency variables, data-driven approaches like Machine Learning Methods have been applied into Nowcasting area. It does not only show a stronger forecasting ability in handling large number of predictors but also present a superior robustness for non-linear data structure. In this research, an Ensemble Method constructed by several Machine Learning Methods have been generated to provide more timely available GDP figures in the period of government statistics release delay. Having integrated an input dataset with data from multiple data sources such as public statistical websites, Reserve Bank of NZ and Stats NZ, our cooperators New Zealand Transport Agency (NZTA) and PayMark, this study is conducted by first applying different Machine Learning methods such as Lightgbm, Xgboost, Support Vector Machine, K- Nearest Neighbors, Ridge Regression, Lasso, Adaboost models. Then these algorithms are combined to generate an Ensemble Model with the assistance of an averaging method, which weights each model individually based on its historical prediction accuracy. The result of the final Ensemble Model is compared with the most commonly used benchmark ARIMA model and Random Walk model in terms of Mean Square Error (MSE) and Median Absolute Error(MAE) value. Statistical tests, such as Friedman Test and Wilcoxon Signed-Rank Test, are employed to check the significance of model superiority. The results indicate that the Ensemble Model significantly outperforms individual Machine Learning algorithm and Random Walk model in forecasting accuracy. When compared with the ARIMA model, it shows slightly better prediction ability with more fore-sights especially in a fluctuating environment.