Deep Neural Trading: Comparative Study With Feed Forward, Recurrent and Autoencoder Networks

Algorithmic trading approaches based on news or social network posts claim to outperform classical methods that use only price time series and other economics values. However combining financial time series with news or posts, requires daily huge amount of relevant text which are impracticable to gather in real time, even because the online sources of news and social networks no longer allow unconditional massive download of data. These difficulties have renewed the interest in simpler methods based on financial time series. This work presents a wide experimental comparisons of the performance of 7 trading protocols applied to 27 component stocks of the Dow Jones Industrial Average (DJIA). The buy/sell trading actions are driven by the stock value predictions performed with 3 types of neural network architectures: feed forward, recurrent and autoencoder. Each architecture types in turn has been experimented with different sizes and hyperparameters over all the multivariate time series. The combinations of trading protocols with variants of the 3 neural network types have been in turn applied to time series, varying the input variables from 4 to 17 and the training period from 8 to 16 years while the test period from 1 to 2 years.