The present thesis analyzes whether shifts in the direction of the stock market can be predicted by supervised learning algorithms and aims to contribute to the emerging literature on the use of machine learning in finance. Previous findings on the feasibility of market timing strategies vary. Many proposed strategies fail to predict market movements consistently enough to beat the returns achieved by buy-and-hold investors. While most market timing strategies are based on standard econometric models, research on the market timing ability of machine learning techniques remains in its infancy.In this thesis, logistic regressions, artificial neural networks, and support vector machines were used to predict whether the excess return of the S&P 500 is positive or negative over time horizons of one week, two weeks, and one month. Based on the predictions, market timing strategies that allocate capital between stocks and a risk-free asset were developed and subsequently backtested over a 23-year period.The results indicate that the proposed strategies outperform various buy-and-hold portfolios, even after transaction costs are deducted. Forecasts over the two-week and one-month horizons were significantly more accurate than predictions over the one-week horizon. A timing strategy based on a hybrid model that combines different algorithms outperformed all other models in terms of risk-adjusted returns. The findings suggest that, contrary to the prevailing view, shifts in the stock market are in fact predictable.
|Date of Award||1 Feb 2021|
- Universidade Católica Portuguesa
|Supervisor||Dan Tran (Supervisor)|
- Market timing
- Machine learning
- Quantitative finance
- Mestrado em Gestão e Administração de Empresas