Using Mixed Frequency Data to Forecast Recessions and GDP

Abstract

This dissertation studies how and if mixed frequency time series data should be used to forecast recessions in the US and Canada, as well as GDP in a selection of OECD countries. The first chapter combines daily and weekly financial data with monthly macroeconomic indicators in a mixed frequency probit (MFP) regression to forecast and nowcast US and Canadian recessions. This chapter adds to the existing literature in multiple ways, including; developing a mixed frequency binary model that could be helpful for topics outside of recession forecasting, examining how higher frequency data should be weighted in the context of recession prediction, as well as a methodology that can nowcast current economic conditions. Overall I find significant improvements in the forecasting and nowcasting accuracy of recessions when using mixed frequency data, compared to a benchmark model that aggregates data into the same frequency. The second chapter extends from the first to apply machine learning techniques to the same problem. I add to the existing literature by incorporating mixed frequency data directly into a classification artificial neural network (MF-ANN) as well as using novel cross validation methods to tune hyperparameters and carry out feature selection with time series data. Overall when comparing US recession forecasting results to the reduced form methodology of Chapter 1, I find mixed results. While some metrics indicate similar performance between the two methods, the ANN makes less extreme forecasting errors on average. The third chapter uses a seemingly unrelated regressions (SUR) approach with a mixed frequency framework to forecast GDP of 10 OECD countries. This chapter adds to the literature as a way to efficiently include cross country information in GDP forecasting equations, as well as being an effective methodology when the researcher is constrained by small sample sizes. Overall we find that SUR outperforms OLS for the majority of countries and forecasting horizons, however as sample sizes are extended this benefit is reduced.