Why is economic forecasting important? Reliable economic forecasts, or accurate predictions about the direction of the economy, serve to help individuals, households, policymakers and firms make sound decisions that could lead to growth, employment and inflation. However, as Niels Bohr, a Nobel laureate in Physics once said, “Making predictions is very difficult, especially about the future.” Today, reliable economic forecasts can be made with a combination of techniques including econometric methods and machine learning.
Tackling this topic at an inaugural virtual public lecture titled “Machine learning in forecast combinations”, Professor Jun Yu from SMU School of Economics spoke to an audience of more than 70 via Zoom on Tuesday, 8 December 2020.
Prof Yu, who is Lee Kong Chian (LKC) Professor of Economics and Finance, asked: “Why do nonlinear econometric and nonparametric models largely fail in generating reliable economic forecasts?” Quoting statistician George Box, who said “all models are wrong, but some are useful”, Prof Yu elaborated that economic activities typically involve many economic agents, making it both an art and a science to build a good econometric model. By testing the validity of economic theories, such models could generate more reliable economic forecasts.
Prof Yu, who joined SMU in January 2004, has published more than 70 papers in leading journals, delving into the areas of economics, finance and econometrics. His articles with Professors Peter Phillips, Yanru Wu and Shuping Shi, for detecting the presence of bubbles in financial assets and real estate, gained the attention of researchers and practitioners. Many central banks (including the Federal Reserve Bank of Dallas) have used these techniques for early warning signals of economic bubbles. Prof Yu currently holds editorial positions in three leading journals in econometrics.
Prof Yu has been awarded various external research grants from the Singapore Ministry of Education Academic Research Fund Tier-2 and Tier-3 grants. He was also the lead Principal-Investigator of the Tier-3 programme on the economics of ageing. He was awarded the LKC Professorship, which is funded through the Lee Kong Chian Fund for Excellence from the Lee Foundation. The award acknowledges and rewards distinguished scholars who are internationally recognised and highly regarded by peers.
Factors such as model uncertainty, and the selection of one particular model possibly leading to overconfident inferences and riskier forecasts, affect the reliability of forecasting.
When structural instabilities and nonstationarities are present in data, a common approach in the econometrics literature is to combine forecasts from different econometric models. This technique is known as forecast combinations. In economics, structural instabilities and nonstationarities are widely expected as the economic environment changes constantly.
Prof Yu explained that forecast combinations, which basically use a set of models, each generating a forecast and then combining forecasts from all models, typically generate a good performance. Those who use the forecast combinations technique firstly need to address the choice of candidate models, followed by the choice of weights to be used.
Prof Yu elaborated on machine learning (ML) as the scientific study of algorithms and statistical models used by computer systems to effectively perform a specific task without using explicit instructions, relying on patterns and inference instead. In recent years, ML methods have found successful applications in predicting economic activities, especially when the underlying relationship linking response and explanatory variables is complicated. However, ML methods typically assume stability in the underlying relationship. Hence, existing ML methods may not be suitable for making economic forecasting when data involve structural instabilities and nonstationarities.
Prof Yu proposed a novel framework that first allows machine learning methods to simultaneously consider multiple model specification, and then averages the outputs by the weights that are either pre-specified or estimated. In empirical applications to forecast key macroeconomic variables (such as GDP growth, the inflation rate, and the unemployment rate) and financial variables (such as the interest rate), Prof Yu’s research found that combining machine learning methods can produce more accurate forecasts than individual machine learning techniques and traditional econometric methods.
“Most of the machine learning methods do not account for model specification uncertainty. It is possible that the best machine learning strategy in one period may not be the best machine learning strategy in another period,” commented Prof Yu. “The idea of forecast combinations can be applied to machine learning strategies and what we propose can be regarded as a new ensemble learning algorithm.”
The virtual lecture concluded with a Q&A session that delved into the possibilities of using machine learning methods to analyse high frequency financial data.