This project is a small mix of a hobby project and a school project. The goal is to predict the behaviour of a stock in the next hour.
In this project we model the problem of predicting stock price fluctuations by simplifying the problem to classification problem. It is enough to know whether the price of a stock goes up or down in the next hour to be able to optimally exploit the stock for profit. Instead of binary classification, we opt to make the problem a multi-class classification problem by dividing the price change into 3 classes: up, down and no change. This is done to hopefully reduce the noise in the data and make the problem easier to solve.
If the price of a stock changes less than 0.2% in the next hour, we consider that the stock price has not changed. We chose 0.2%, because it balances the class labels.
As data we use data pulled from the Yahoo Finance API for Helsinki Stock Exchange (OMX Helsinki). We only take certain stocks filling our quality criteria, which leads to a total of 14 stocks. As input for our model, we give the model a timeseries of M last price changes of each stock, and we predict the class of the next price change for each stock, so a total of 14*3 outputs. Since we difference the data for the model we guide the model to keep track of the price changes, and we squeeze the data into a smaller range which should improve convergence.
We chose the following stocks, because they had no missing values, splits, and their average price was between 20 - 100 euros in the last 5 years:
- Revenio Group
- Neste
- Orion Class B
- Kone
- Olvi
- Huhtamäki
- Detection Technology
- Orion Class A
- Cargotec Corp
- Vaisala
- Ålandsbanken Class B
- Valmet
- eQ
- Ponsse
As models, we try RNN, LSTM and Transformer models. For each model type we perform hyperparameter tuning using the Hyperband algorithm in KerasTuner library.