foreverycast / lstm_forecast

Example script for forecasting of Time Series data.

LSTM forecast multi dimensional

Stock prediction forecast with volume and news sentiment

We create an example script for forecasting of Time Series data. The example is based on following LSTM article in additional we add two dimensions Volume and News sentiment. News sentiment in the example data was tacken from our side foreverycast.info, where we aggregate daily news to a simple number in range between -1 (bad sentiment) to 1 (good sentiment). The idea is to create better forecasting data based not only on one dimension.

The multi dimensional script can create a forecast for set time period. The prediction for t+1 values is always based on the previous values. To create a prediction for values more than t+1, the previous will be used. In the prediction for t+2, first the prediction for t+1 will be done. As it is hard to predict news sentiment the value for t+1 is set to 0, same for volume dimension.

Run the script

python3 lstm_many_to_one.py

Before the script starts, three variables can be set:

• File name, the name of the file with the data for learning.
• Backdays, a variable of int type. The variable determines how many "days" are relevant for the one future predicted value. (Default: 20)
• forecast_period, a variable of int type, The variable set the number of predicted "days". (Default: 10)

Learning variables

The script includes variables for learning duration. Variable EarlyStopping can be changed, if the script should run until the end of the epochs.

my_callbacks = [
    EarlyStopping(patience=200),
]

Change the number of epochs can be increased in model fit.

model.fit(X, Y, epochs=1000, validation_split=0.2, verbose=2, callbacks=my_callbacks)

At the moment it is set, that the data list should be longer than 60 rows.

To-Do

• Add type check for input data
• Add requirements.txt
• Add test.py