LSTockMarket Predictions

Title: LSTockMarket Predictions

Team: Dylan Brady (dbrady), Armaan Dhillon (adhillo4), Filip Kierzenka (fkierzen)

Introduction

We will train an RNN regression model with LSTMS to see if we can predict the opening prices of stocks

Related Work

We came across a paper that tried using LSTMs to predict the opening stock prices of Nike and Google- two prominent companies that are members of the S&P 500. This made us interested in seeing if we could implement an RNN to predict opening stock prices of companies we're interested in. The goal of the paper was to see how well their regression matched the actual price chart of the stock by experimenting with the number of epochs used to train, and the amount of data used. We chose to base our project off of this paper because it demonstrated how accessible predictive modeling can be for someone who has an interest in quantitative investing and familiarity with Tensorflow.

Data

CSV files downloaded from Yahoo finance that contain the opening stock price, daily low, daily high, and closing stock price for 2-3 stocks over a range of at least 5 years.

Methodology

We will create an RNN using LSTM layers in order to incorporate historical stock data in our predictions for the future testing data. The downfall of DNNs is that they don't have memory capabilities that help in storing information of the past which is intuitively essential for guessing what the future price is going to be. Perhaps after making regressions using an LSTM RNN architecture, we will see how close we can get using a classing DNN.

Metrics

Since we are trying to create a regression, our metric will mostly be mostly based on graphical data of how closely our prediction fits the actual price movements of the stock. One rough way of doing this numerically would be to calculate the percentage error for each prediction date and average that over the testing set. Perhaps as we conduct more research and get further along in the project, we'll find an accurate way of calculating the percentage accuracy of our model.

Ethics

1) Major stakeholders would be anyone who places large sums of money with a firm using quantitative methods to invest. Pension funds, retirement funds, and other institutions may invest with firms using algorithms to invest, and those algorithms could malfunction or be "broken" by certain swings in the market which could end up hurting the people who are depending on those funds for future payouts. 2) Slightly unrelated, but still in the scope of quantitative investing: high frequency trading firms that use a mix of trained and rule based algorithms are notorious for front running investments made by large institutional investors. Scott Patterson's book _ Dark Pools: The Rise of the Machine Traders and the Rigging of the U.S. Stock Market _ sheds light on this subject, and our model, while still quite innocent, falls under a larger branch of investing methods that can chip away at returns from other firms while generating large profits for themselves.

Division of Labor

Filip and Armaan will pick stocks to predict prices for using RNNs, and Dylan will try to match the success with a DNN.

Built With

• python
• tensorflow