STOCK TREND PREDICTION

Tetiana Gushchyk, email: tetianagushchyk2019@u.northwestern.edu

EECS 349 Machine Learning, Spring 2018, Northwestern University

Stock prediction is an interesting and important task. It is a challenging problem that has puzzled many software engineers for years. It is also important because if a feasible solution for the task is found, it will be valuable tool for investors. The goal of the project is to build a machine learning model that can predict change in stock price. The model takes information about historical stock price and returns a prediction about the change of the stock price in the next 30 days. Trained model should predict a price trend for the next 30 days and classify the stock into one of three categories: “buy”, “sell”, “hold”. The final goal is to make trading profit.

Approaches used for the project:

• Nearest Neighbor
• Naive Bayes
• Logistic Regression
• Multilayer Perceptron

Training dataset includes 29 attributes:

• Daily open/close price
• Daily low/high price
• Daily volume
• Dividends
• SMA (20-day simple moving average)
• EMA (20-day exponential moving average)
• RSI (20-day relative strength index)
• YTD (year-to-date change)
• 1-week low/high, 10-week low/high, 20-week low-high, 52-week low/high
• EPS (earning per share)
• Daily S&P 500 index price, YTD, 1/10/20/52-week low/high

Each trading day is labelled with prediction “buy”, “sell”, “hold” using average future price for next 90 days:

• “buy” : 90-day future average price will rise at least 3% above current price
• “sell” : 90-day future average price will drop more than 3% below current price
• “hold” : 90-day future average price will stay between [-3% : 3%]

Key results

Multilayer perceptron with 2 hidden layers and sigmoid activation function has performed better than other methods with final result of 65% accuracy. It as a fairly good result taking into account the size of dataset and quality of financial attributes. This demonstrates a potential in solving complex tasks with large datasets, but better accuracy requires better training dataset with more features relevant to stock price prediction.

Source code | Final report (PDF)