Evaluation of Machine Learning impact on Asset Risk Premia measurement

Numerous literature has tried to empirically demonstrate which method between the Support Vector Machines, the Decision Trees and Artificial Neural Networks performs best.
The free lunch theorem of Machine learning (Wolpert 1997) states that every possible algorithm on average has the same error when applied in the classification of new unseen data. The implication and meaning given to this "theorem" is the fact that there is no a priori machine learning algorithm that is always better than the others. Some methods have low performance on some issues, while others have a stronger performance in other tasks.

It is important to underline that this theory only holds when referring on average to all possible types of data distributions. When navigating to real problems analysis, it is advisable to choose and test algorithms that best approximate a given problem. As a consequence, it can be asserted that the objective of the researcher is not to find the perfect algorithm that can be applied and solves all kinds of questions properly, but the one that can work better with the nature of the problem.

Moving on to the precise field of study, which mainly refers to the area of time series analysis, it is appropriate to draw attention to an interesting comparison of interesting comparison between the Support Vector Machines and the Artificial Neural Networks in the field of Time series analysis, conducted by Samsudin, Shabri and Saad in 2010. The analysis made is interesting because often these two methods are considered very similar in "predictive performances".

Both the performances of SVM and ANN are tested on five different time series analysis topics: the chemical process concentration, the IBM common stock closing price, the chemical process temperature, the sunspot’s data and the international airline passengers. The main feature of these topics is their different statistical characteristics, i.e. seasonality, trend, nonlinearity, with different period of observations (1 hour, 2 hours, Daily, Monthly, Yearly).

Even if both SVM and ANN had good generalization performance, SVM outperformed ANN in predicting chemical process concentration, the IBM common stock closing price, and the Chemical process viscosity in terms of RMSE and MAE, while ANN outperformed SVM for sunspot’s data and international airline passengers. Therefore, SVM worked better in series homogeneous non-stationary (series 1 and 3), and in a series without evident trend and seasonality (series 2). Instead, in a series where data has a cyclical pattern (series 4), or with seasonality (series 5), ANN demonstrated to perform better than SVM.
In any case, the major drawback of this analysis is the use of very few data, which could have affected the result.