High-Performance Computing Evaluation of Supervised Learning Models and Human Expert Criteria for Stock Screening Using Fundamental Factors

Abstract

We have designed a High Performance Computation Fundamental Analysis Stock Screener and Ranker system (HPC.FASSR), build on PyCOMPSs to compare the performance of various supervised learning algorithms like Neural Networks, Random Forests, Support Vectors Machines, or AdaBoost, and well-known human expert trader's criteria for selecting stocks based on fundamental factors. The parallelization of HPC.FASSR with PyCOMPSs allows us to explore a huge number of configurations in a short time. We have run an extensive collection of experiments in MareNostrum 4, the main supercomputer in the Barcelona Supercomputing Center. Results show that, although the human expert beats the market, almost all well-tuned machine learning models beat the expert. This comes at the expense of resources taken from supercomputing, and hence, in a certain sense, our results give general support to reject a computational cost form of market efficiency, besides disputing a semi-strong form of the Efficient Market Hypothesis.

Keywords: High performance computation, supercomputer, supervised learning algorithms, stock screening, fundamental analysis

JEL Classification: C63, C45, C52, C53, C55