LSTM stock prediction This simple script predicts future stock prices of AMAZON based off data from TESLA, MS and APPLE in two ways There are two LSTM classes for each method with a bidirectional network being used in both with dropout. A linear layer is also used after the LSTM to make the final predictions. The number of predictions can be set in the main functions, but I choose to just use 1 being the close price.
-
LSTM batch This trains one LSTM model on the 3 stocks given above. The data will be batched all at once and the LSTM_time_series.py class will be used. This attempts to learn some cross correlation between the stocks to better predict AMAZON's
-
LSTM ensemble This trains three separate LSTM models on the 3 stocks given above. The training is completed one at a time using the LSTM_time_series_ensemble.py class. A weighted average between each models predicted stock for AMAZON will be included.
The batched data tends to provide better predictive power than the ensemble method, potentially due to a greater varaity of data present in any cross correlations between the stocks. A simple superposition of predictions requires rescalling of each model too, making it more complicated.
A potential future script could train more stocks as well as various combinations of batched versions of the stocks and return a larger ensemble. This ensemble could also be weighted in a non-uniform way, potentially using a NN to choose weights for a more accurate output.