Predicting Stock Prices with Multi-Layer Perceptron Raamis Hussain
Outline Try to predict daily closing price of Apple stock using various features Date, daily open, daily high, daily low, price/earnings ratio, market capitalization, volume Predict trends as well as actual price using 1200 days of data
Configuring MLPRegressor 1-3 hidden layers Rectified Linear Unit Function as activation function: f(x) = max(0, x) Stop when validation score stops improving
Predicting Price Trends Try to predict general price movement over some time period Success = predicting positive/negative change Use date as only feature
Predicting Price Trends
Success! We are able to accurately predict the general trend of the price using our MLP Next step: Train MLP with features and try to predict actual price Train features for closing price 15 days in the future
15-Day Prediction Test
Doesn’t Work! The Neural network is making predictions of the past instead of the future Apparently the features are too closely related to same-day price MLP predicting closing price of day on same day as given features Try with less features
15-Day Prediction Test #2 Features: Date, p/e ratio, market cap, volume
Conclusions Even with features that are not related to the price, the prediction isn’t very accurate MLP predicts same-day price instead of future price We need more general and sophisticated features to train on Earnings per share Debt to equity ratio And others Our simplistic method is not sophisticated enough to predict exact stock prices