Deep Learning constructs feature using only raw data. The leaned representation of the data outperforms expert features for many modalities including Radio Frequency (Convolutional Radio Modulation Recognition Networks), computer vision (Convolutional Deep Belief Networks for Scalable Unsupervised Learning of Hierarchical Representations) and audio classification (Unsupervised feature learning for audio classification using convolutional deep belief networks). In the case of Convolutional Neural Networks (CNN), the data representation is leaned in a supervised fashion with respect to a task such as classification. For a typical CNN to generalize to unseen data in requires very large quantities of data. The amount of data available is often not sufficient to train a CNN. GANs allow features to be learned unsupervised. This reduces that potential for features being overfitted to the training data and in turn means that a classification algorithm trained on the features will generalize on a smaller amount of data. In fact, GANs promote generalization beyond the training data, as will be seen.

For a full review of a GAN: Generative Adversarial Nets The Generator is trained to generate data that looks like historical price data of the target stocks over a gaussian distribution. The Discriminator is trained to tell the difference between the data from the Generator and the real data. The error from the Discriminator is used to train the Generator to defeat the Discriminator. The competition between the Generator and the Discriminator forces the Discriminator to distinguish random from real variability.

Data Historical prices of stocks are likely not very predictive of the future price of the stock, but it is free data. Technical indicators are calculated using the historical prices of stocks. Not being a trader I don't know the validity of technical indicators, but if a sufficient number investors use technical indicators to invest such that they move the market, then the historical prices data of stocks should suffice to predict the direction of the market correctly more then 50% of the time. Training The GAN is trained on 96 stocks off the Nasdaq. Each stock is normalized using a 20-day rolling window (data-mean)/(max-min). The last 356 days (1.4 years) of trading are held out as a test set. Time series of 20 day periods are constructed and used as input to the GAN. Once the GAN is finished training, the activated weighs from the last layer of convolutional lays is used as the new representation of the data. XGBoost is trained to classify whether the stock will go up or down over some period of time.

Testing The data the was held out in the training phase is run through the Discriminator portion of the GAN and the activated weights of the last convolutional layer are extracted. The extracted features are then classified using XGBoost.

Results The confusion matrix shows the results of the model's classification. The perfect confusion matrix would only have predictions on the main diagonal. Each number off the main diagonal is a misclassification.

Predictions of Up or Down movement over 10 Days

The predictions over a 10 day period are quite good.

Predictions of Up or Down movement over 1 Day

Predicting over a short time interval seems to be harder. Results loss significant accuracy when trying to predict the next day movement of the stock.

Predictions 10% Gain Over 10 Days

Just knowing that the stock will go up or down is of limited use. A lot of stocks will go up in a day but an investor will want to only buy the stocks that will go up the most, maximizing returns. This time the XGBoost model was trained to predict stocks that would go up by 10% or more over the following 10 days.