Smart Index Replication with Deep Learning

Stock market indices are composed of many different stocks. For example, the S&P 500 consists of 500 stocks of large American companies. Tracking and predicting its performance is complex due to the high number of stocks involved. Identifying a low number of stocks that capture the majority of the index moves is highly valuable - this reduced replication index can then be used to predict the full index performance in a cost-effective way.

This white paper describes a method for smart index replication using a deep learning auto-encoder, which can effectively predict the performance of a larger index. The neural network is trained on historical data, fast and efficiently using both CPUs and GPUs. Backtesting results for the example of S&P 500 show that a set of 10 stocks is enough to accurately predict the index's performance with less than 8% error.