Essays on macroeconomic forecasting with machine learning and leading indicators

Abstract

This dissertation consists of three essays on forecasting macroeconomic time series with machine learning techniques, evaluating the forecast performance of the Organization for Economic Co-operation and Development (OECD) Composite Leading Indicators (CLIs) and proposing a new index based on the latest Artificial Intelligence algorithms. The first essay investigates if the CLI improves forecasting performance under different conditions and model assumptions. We use a variety of time series and machine learning models to evaluate if incorporating the CLI into the forecast reduces the errors for inflation, unemployment rate and industrial production. We hypothesize that the CLIs of high income, highly developed nations will provide more accurate forecasts than those of developing countries, due to the more complete data those countries have from being able to pour more resources into data collection, analysis and storage. In the second essay we analyze how each country’s CLI is constructed and what components are used for its calculation. We use machine learning algorithms with feature selection to evaluate if the given weights in the construction of a country’s CLI is optimal. If discovered that it is not, we use the weights found to test if the newly weighted index provides better forecasts than the current OECD CLI. The third essay expands on the second. We propose a new index, similar to the CLI, but constructed based on using machine learning algorithms to find the best combination of economics variables available in the OECD database that explain variations in future values, thus providing the most accurate forecasts possible. Additionally, we can determine the components that most influence the different target variables- inflation rates, unemployment rates and industrial production for each country in our analysis. We can then use forecast encompassing tests and compare forecast errors from this new composite index to the CLIs to show relative improvement in the predictions.