Abstract
The current thesis attempts to highlight and offer some insight on the issues of regime shifts, contagion and predictability in financial markets. The first chapter explores an important variable in finance and economic analysis, the state of the equity market. I apply a univariate Markov Switching model and try to characterise the non-linear dynamics of the stock markets in G7 countries. Using this particular class of model, I am able to capture the behaviour of the time series in different regimes and consequently to detect bull and bear regimes. The empirical findings demonstrate statistically significant evidence of more than one regime in each of these stock markets. The in-sample analysis suggests that a simple two-state model with regimes characterised as a high volatility bear regime with negative mean returns and a bull regime with positive mean returns and low volatility is able to capture the time-varying volatility in stock returns. The second chapter examines the correlation between stock and bond returns, considering a range of univariate, bivariate and trivariate Markov Switching models for the U.S. and the U.K. using data for the sample period 1986-2010. The objective of this chapter is two-fold: first to observe how the correlation between returns on stocks and bonds changes as the economy moves from a bull to bear regime, and second to explore the effect of monetary policy, as expressed by interest rates, on the correlation of stock and bond returns. A specification test reports evidence that the joint distribution of stock and bond returns may only be described by a two-regime MS-VAR(1) model. I have found evidence of flight-to-quality phenomena, as during bear regimes, the prices of the two asset classes tend to co-move less compared to a bull regime. This result appears robust to the inclusion of an economic predictor variable, the three-month T-bill rate, in which case not only the mean, the variance and the correlation between stock and bond returns become time-varying as driven by the hidden Markov regime, but also the ability of current interest rate to forecast subsequent stock and bond returns becomes strongly time-varying. The third chapter tests for contagion firstly, within the Euro Area (EA hereafter), and secondly from the U.S. to the E.A. Using "co-exceedances" - the joint occurrences of extreme negative and positive returns in different countries in a given day - I define contagion within regions as the fraction of the co-exceedances that cannot be explained by fundamentals (covariates). On the other hand, contagion across regions can be defined as the fraction of the co-exceedance events in the E.A. that is left unexplained by its own covariates, but that is explained by the exceedances from the U.S. Having applied a multinomial logistic regression model to daily returns on 14 European stock markets for the period 2004-2012,1 can provide the following summary of the results. Firstly, I found evidence of contagion within the E.A. Especially, the E.A. ten-year government bond yield and the EUR/USD exchange rate fail to adequately explain the probability of co-exceedances in Europe. Therefore, these variables are important determinants of regional crashes. Secondly, I have observed that negative movements in stock prices follow continuation patterns - co-exceedances cluster across time. Thirdly, there is no statistically significant evidence of contagion from the U.S. to the E.A., in the sense that U.S. exceedances fail to explain high probabilities of co-exceedances in the E.A. This result holds under a large battery of robustness checks. The fourth chapter investigates the out-of-sample predictability of stock returns, and addresses the issue of whether combinations of individual model forecasts are able to provide significant out-of-sample gains relative to the historical average. Empirical analysis for the German stock returns over the period from 1973 to 2012 implies that, firstly, term spread has the in-sample ability to predict stock returns, secondly, and most importantly, this variable successfully delivers consistent out-of-sample forecast gains relative to the historical average, and thirdly, combination forecasts do not appear to offer significant evidence of consistently beating the historical average forecasts of the stock returns. Results are robust using both statistical and economic criteria, and hold across different out-of-sample forecast evaluation periods.