Predicting Stock Prices - Regression Problem

In this example, we take a look into how to use Neurons in a regression problem: predicting the mean price of a stock in a certain day.

In Machine Learning, a Regression Problem is a problem where the output variable is continuous, as opposed to a Classification problem, for example. In our case, the stock price can be any positive real number.

This time, we'll use different datasets for training and testing the model. For training, we'll use Apple's stock values from 15/02/2013 (dd/mm/yyyy) to 08/06/2016. For testing we'll use the prices from 09/04/2016 to 06/02/2018. The columns are as follows:

• 0. Date (yyyymmdd)

• 1-5. Mean price of stock in the last 5 days (dollars)

• 6. Current day mean price (dollars)

For our model, we'll use the stock price of the last 5 days as input, and we want it to predict the current day stock price.

Let's create our model.

• Open Neurons. As the 'Training File', select 'stocks_data_apple_train.txt' inside the 'Data' folder.

• Change input columns to 1-5, the last few days stock prices. Change output column to 6, the current day price.

• Press 'File'->'New Model' or just press 'Ctrl + N'. We'll name the model 'Stocks_model'.

Next, we'll add 2 Hidden Layers and add some neurons in the Input layer.

• Double click the Input Layer, change neurons to 20.

• Press 'Add'. Let's use 20 neurons. We'll name the layer 'Layer_2'. Press 'Add Layer'.

• Press 'Add' again, choose 12 neurons for this one. Name it 'Layer_3'.

Since this is a Regression problem, we'll have to change some of the training settings.

• Press 'Train'.

• Change 'Loss function' to 'Mean Squared Error'. This'll tell us how far away the predicted stock price is from the true one.

• Select the 'Mean Absolute Percentage Error' metric.

• Epochs: 200. Batch size: 5.

• Select all 'Show' options. Press 'Train'.

Take a look at our model's performance:

As we see in the first figure, our estimates usually deviate 1.3% from the true value. This seems very good, but remember these are stock prices, so a small percentage variation can be a big deal. Let's use the test dataset to verify our model's performance against new data.

We want to see how this model performs with data it hasn't seen before, so we'll use the test dataset.

• In 'Test file', choose the file 'stocks_data_apple_test.txt' that's inside the 'Data' folder.

• Press 'Test' and select 'Predicted Vs Correct Output' in 'Show Plots'

• Click 'Browse' to save testing data to a file, we'll call it 'stocks_output'.

• Press 'Test Model'

Our results are acceptable given the simplicity of our model, but are we going to get rich using it? Probably not. In the Final Note, let's see how we could make it better.

This model used only the mean stock prices of the previous 5 days to predict the current price. The truth is, there are hundreds of different indicators day traders use to try to predict price fluctuations: moving averages, Bollinger bands, trading volumes, etc. One way we could redo this problem would be using some of those indicators paired with the mean prices.

There's also a better way to approach this problem with Machine Learning (and Keras): by using a model with Long Short-Term Memory layers. These types of layers, as the name suggests, give the model a sort of 'memory' of past information. Learn more about it here. Currently, Neurons only supports Dense layers, so this isn't possible with this program.