A demo of how our stock market predictor works.
At Hack The 6ix in 2019, a group of friends and I worked on Prophecy, a tool designed to help you make better investment decisions. By creating a data stream from the stock market to our program, we trained a LSTM-based neural network to predict future stock prices using historical stock performance as a ground truth. Long Short-Term Memory networks are good at finding patterns in time-series data. Depending on how much we overfit to the very data the network had to predict on, future performance within a short time horizon sometimes was predicted correctly. The idea here was if you let the network "cheat" a bit by looking at the performance of the stock whose future it has to predict, it may gain contextual clues to help with the forecasting. However, it was uncertain whether predictions were generalizable or just happening to correctly map on to some of the regular fluctuations of stock prices. While I learned a lot about LSTMs throughout this project, clearly this problem is incredibly difficult otherwise we'd have used this to get wealthy a long time ago.