Evaluating macro-finance interactions using mixed frequency methods

This dissertation examines how macroeconomic variables influence financial market volatility and correlations using mixed frequency time series methods. The modelling framework allows combining high-frequency and low-frequency data within the same model and thus allows directly relating the economic data to the low-frequency component of volatility or correlations. The dissertation sheds light on which economic variables influence the low-frequency component of volatilities and correlations, as well as examines various methods to improve long horizon forecasts for stock market volatility by utilising the information in macroeconomic variables.

The first essay considers the relative and combined importance of macroeconomic fundamentals and survey-based sentiment data for modelling US equity market volatility in a GARCH-MIDAS framework. It uses a data set which accurately takes into account real-time data revisions to lags of the macroeconomic data and extends the analysis to include several new variables. Forward-looking macroeconomic data is important for forecasting volatility, even after the information in sentiment indicators is controlled for. On the other hand, for example, consumer confidence indicators contain information complementary to forward-looking macroeconomic variables. Overall, models combining macroeconomic and sentiment data tend to improve in-sample fit and in some cases also out-of-sample forecast accuracy compared to models only driven by one type of data. The improvements in forecasting performance are, however, not statistically significant, and therefore the results do not strongly advocate using several explanatory variables in the MIDAS polynomial.

In the second essay I assess the time-variation in predictive ability arising from the inclusion of macroeconomic and financial data in a GARCH-MIDAS model for US stock market volatility. I compare forecasts from a GARCH-MIDAS model to forecasts from a nested GARCH model, and therefore the differences in forecasting performance directly reflect the impact of economic data. While forecasting performance between the two models is similar when considered over the full out-of-sample period, there is clear time-variation in relative forecasting performance over sub-samples. I suggest the variation could arise from the phase of the business cycle or the volatility environment and find particularly strong evidence in favour of economic variables being important for volatility forecasting during low-volatility periods. Forecast combination methods and a decision rule based on conditional predictive ability produce consistently better forecasts than the GARCH model, although statistical significance of the improvements depend on the loss function considered.

The third essay considers the time-variation in the co-movement of equity returns and exchange rate returns in several markets using the DCC-MIDAS model. Determining the economic drivers of the low-frequency correlation aids in differentiating between the various theoretical explanations for the correlation, which predict both a positive and a negative relationship. The essay concentrates on the portfolio rebalancing channel and on two hypotheses suggested in the earlier literature, namely flight-to-quality and quantitative easing (QE) related search-for-yield, in addition to examining the sensitivity of the correlation to other economic variables related to portfolio rebalancing motives, such as the business cycle. Although there are common elements driving the return correlation in the different markets, for instance, interest rate differentials and quantitative easing measures, their impact on the correlation varies, suggesting the underlying theoretical explanation differs across markets. While the onset of US QE1 had a clear impact on the correlations, overall the results suggest that being in a QE regime is more important than announcement effects for the long-term correlation.