5. Conclusions
This paper demonstrates how algorithmic methods and macroeconomic covariates can be used to construct an expert system to perform independent trading in the stock markets. Some of the results are qualitatively appealing, but statistically significant results are seldom found. Moreover, for a large number of statistical tests, positive results are bound to occur randomly. In Table 1, there are four significant associations, about as many as would be expected at random (4.95). In the trading test (Table 3), there are 39 mean-variance dominating portfolios, significantly less would be expected if the portfolios were equally likely to be mean-variance dominating or not (P = 0.02). However, if the trading costs are removed (Appendix), there are 61 dominating portfolios, more than would be expected at random (P = 0.008). Interesting results are found by using Lasso and Elastic Net in Netherlands. These results are both statistically and economically significant (Tables 1 and 3, respectively) and they are robust both towards the choice of statistical test and increasing the trading costs (Appendix). Moreover, very interesting results are found if the target of the analysis is shifted, i.e. if the algorithms are used to predict large downturns (more than 2% monthly) in place of sign return. In that case, statistically significant predictions can be generated in most markets and economically significant results can be obtained in all markets.
