This thesis investigates the relationship between the business cycle and risk premia-based investing strategies in the US stock market since 1963. A business cycle score is developed using a four-phase classification scheme to categorize the business cycle into either an expansion, slowdown, recession, or recovery. Based on the classification scheme, two predictive classification models – the multinomial logistic regression and the support vector machines model – are developed using 25 leading macroeconomic, financial, and credit variables on five horizons of 0- to 12-months ahead. Out-of-sample forecasting tests reveal that the two models correctly predict 66% of the regimes on average with the multinomial logistic regression showing the best performance, particularly at short horizons with 77% and 74% accuracy at 1- and 3-months ahead. The regression results of the multinomial logistic regression also confirm prior literature on the significance of key indicators. The 6-month lagged yield spread, the S&P 500 yearly return, the number of new non-farm jobs added, and the ISM purchasing mangers’ index are significant at four out of five horizons. In general, many macroeconomic indicators have predictive power at short horizons, while financial and credit variables are highly predictive at longer horizons. The predictions of the two models are then applied in market-timing asset allocation strategies with the objective of timing the business cycle to maximize risk-adjusted return. The assets of analysis are the market portfolio and four long-short factor portfolios which includes the size, value, momentum, and volatility factor. Factor investing – also known as risk premia-based investing – entails targeting specific drivers of asset returns such as the ability of small-cap portfolios to outperform large-cap portfolios. Factor risk-premia can be targeted in single-factor models to generate higher returns, but combining several factors in multifactor models can both increase returns while lowering risk exposures due to high diversification benefits between factors. In order to find optimal portfolios over the business cycle, mean-variance and Black-Litterman optimization has been applied to the five assets for each regime. The out-of-sample results of market-timing strategies applying mean-variance optimization are not particularly favourable for investors timing the business cycle. Only at three horizons are the dynamic strategies able to outperform the market portfolio in terms of Sharpe ratio and Jensen’s alpha, however, the information ratio is negative across the board. When compared to a naïve 1/N portfolio, none of the dynamic strategies are able to outperform this benchmark. In addition, neither of the two predictive models dominate one or the other across the horizons, but the average out-of-sample performance of portfolios using the support vector machine are marginally better. When applying the Black-Litterman approach, we are able to generate less extreme weights that more accurately reflects the estimated risk premia over the business cycle. As a result, the out-of-sample results of all dynamic market-timing strategies are able to outperform the benchmark in terms of Sharpe ratio, significant Jensen’s alpha and positive information ratios. However, once again, the 1/N portfolio outperforms all the dynamic portfolios and active investors are best off resorting to naïve diversification rather than attempting to time asset allocation over the business cycle.
|Educations||MSc in Applied Economics and Finance, (Graduate Programme) Final Thesis|
|Number of pages||107|