Dark-Pool

Dark pools are private venues that the investor can exchange large amount of stock there without impacting market price overly. It is ubiquitous in modern trading. However, unlike traditional stock exchange, there is little transparency of trade execution and the information flow is asymmetric in dark pools. Also unlike traditional stock exchange, current finance literatures show limited understanding about the inner working of dark pools. Based on these motivations, we developed machine learning methods by exploiting asymmetric data to predict liquidity in dark pools trading. The model building can be categorized into three steps. The first step is to apply multiple methods to do data cleaning and feature selection. Then, a battery of machine learning algorithms with different techniques are applied to the clean data. Finally, all methods are ranked and the best model is used on predictions. Based on our analytical results, we arrived at several conclusions: First, our results show an significant improvement on classification result comparing with vanilla out-of-box machine learning algorithms. Adaboost classifier shows the best classification result. Second, due to high-imbalanced asymmetric data, our analysis gives a strong indication that we are reaching limit within the selected framework. By studying the multivariate mutual information between features and output variable, it is clear that the existing features share limited information with output variable. Hence, generating useful features or adding more features into the model are reasonable direction of future work.