Stock Prediction

This LSTM network collects stock prices of 20 companies in NYSE, and tries to predict these future prices.

collect stock data

Collection of stocks are defined in stock_price_collector.py.

The stock prices are saved in .npy format. Two files are saved, each for training and validation of the network.

The time range for data are defined like this.

training_start_time = datetime.datetime(2013, 1, 1)
training_end_time = datetime.datetime(2015, 1, 1)
validation_start_time = training_end_time
validation_end_time = datetime.datetime.today()

The data is then saved.

np.save("training_nyse_stocks.npy", training_nyse_open)
np.save("validation_nyse_stocks.npy", validation_nyse_open)

train neural network

The neural network is defined and then trained in stock_lstm.py.

The weights from the previous training are loaded.

# load weights if exists
try:
    model.load_weights(WEIGHTS_FILE)
    print("loaded weights")
except Exception as e:
    print("could not load model")

Then, the network is trained, and then the weights are saved to lstm_weights_normalizesd_new.h5. You need to download h5py from pip and install hdf5 to use weight saving.

• On mac run

brew install hdf5
pip install h5py

validate neural network

Validation of the network can be done in validation_nyse_stocks.py.

It will output a graph of the actual stock prices and the predicted stock prices.

Other things

• Currently, the network easily gets overtrained if you train it until the early-stopping module stops the training process.

• Other files not mentioned in this readme are weights of previous trainings, and has directly nothing to do with the network.