ABSTRACT — This paper focuses on the training procedure of artificial neural networks for portfolio management. More specifically, it investigates a neural network -with weights derived from a genetic algorithm embedded in the training process, that automates a buy/sell/hold procedure by learning from the direction of a stock within a “diversified portfolio” (in this case a portfolio of the Dow Jones Industrial Index; hence the loose usage of the term). The binary indicators are generated using a function that observes the daily percentage changes of the index and assigns the corresponding operator if the change is a more than 5% gain, more than a 5% loss, or anything in between. Further development of the project can incorporate a diversified portfolio of securities and reallocate the portfolio weights and/or target beta and then execute the appropriate transactions or complete liquidation of the portfolio depending on its performance. While genetic algorithms are not the end-all of metaheuristic methods, the healthy skepticism that machine learning in a panacea for quantitative finance warrants a rethinking of neural networks. Embedding more traditional techniques could provide better alternatives to the sometimes unnecessarily -and even fatally, complex methods developed in the field.