I investigate whether an expert system can be used for profitable long-term asset management. The trading strategy of the expert system needs to be based on market predictions. To this end, I generate binary predictions of the market returns by using statistical and machine-learning algorithms. The methods used include logistic regressions, regularized logistic regressions and similarity-based classification. I test the methods in a contemporary data set involving data from eleven developed markets. Both statistical and economic significance of the results are considered. As an ensemble, the results seem to indicate that there is some degree of mild predictability in the stock markets. Some of the results obtained are highly significant in the economic sense, featuring annualized excess returns of 3.1% (France), 2.9% (Netherlands) and 0.8% (United States). However, statistically significant results are seldom found. Consequently, the results do not completely invalidate the efficient-market hypothesis.

Introduction

This paper examines whether it is possible to use expert systems for long-term asset management in the stock markets. The decision rule of such expert systems is trivially simple: Invest in stocks if the stock market is likely to rise and invest in the money market if the stock market is likely to decline. However, to this end, one needs predictions of the market movements. According to mainstream opinion in economics, it is impossible to predict the stock markets, as that would generate an arbitrage opportunity. This view is known as the efficient-market hypothesis (EMH; e.g. Fama, 1991). However, there are other schools of thought. For example, the adaptive-markets hypothesis of Lo (2004) states that individuals use simple heuristics to trade in the stock markets, and consequently, they are not completely rational. This seems to contradict EMH. Moreover, there are theoretical constructions within the discipline of neoclassical economics (e.g. Singleton, 2006, Chapter 9) which show that there can be some degree of predictability in the stock markets, even if the assumptions of EMH are in force. Thus, it is a question of obvious empirical interest if the markets can be predicted or not. The empirical evidence regarding stock market predictability is mixed. In an influential paper, Welch and Goyal (2008) refuted previous reports of market predictability. The argument was that most authors hitherto had investigated in-sample correlations and the models had no out-of-sample predictive power. Even the in-sample correlations were often lost when the models were updated by new data. Thus, the results could be refuted as statistical artefacts. However, others have challenged the findings of Welch and Goyal (2008). For example, Chevapatrakul (2013) has produced significant out-of-sample predictions for the UK stock market. Similarly, Skabar (2013) and Fiévet and Sornette (2018) have published significant results regarding daily data from the US market. Thus, the debate is ongoing. In this paper, I use contemporary statistical and machinelearning methods to generate out-of-sample predictions in 11 developed stock markets. The methods considered involve ordinary least squares, logistic regressions, regularized regressions (e.g. Tibshirani, 1996; Zou, 2006) and similarity-based classification (Skabar, 2013). Some authors have reported it to be easier to give a binary prediction of profit or loss than to give an estimate of the expected return (e.g. Leung, Daouk, & Chen, 2000; Nyberg, 2011; Nyberg & Pönkä, ۲۰۱۶). At any rate, it is such sign predictions that the expert system ultimately needs to manage the investment. Thus, I have chosen sign prediction as the objective of this study. I use a combination of statistical tests and trading simulations to assess the potential of the expert system to perform profitable asset management. The rest of this paper is organized as follows. Chapter 2 surveys related work. (The lessons learned from previous work to a large degree guide the modelling choices made in this paper.) Chapter 3 introduces the material and methods. The results are presented in Chapter 4. These are divided in two main categories: Main results (Chapter 4.1) and results obtained from sensitivity analyses (Chapter 4.2). Chapter 5 concludes and presents directions for future work.