Build a Stock Price Indicator

Photo by Rick Tap on Unsplash

Table of Contents

1. Instroduction
2. Project Descriprtion
3. File Descriptions
4. Deploying
5. Results
6. Licensing and Acknowledgements

Introduction

Investment firms, hedge funds and even individuals have been using financial models to better understand market behavior and make profitable investments and trades. A wealth of information is available in the form of historical stock prices and company performance data, suitable for machine learning algorithms to process.

Project Description

The objective of this project is to build a stock price predictor. A user is able to get stock price predictions after training the model live on historical data. The system predicts 'Adjusted Close Price'. The final application is deployed on heroku at https://emichris-stock-predictor.herokuapp.com/

File Descriptions

There are 3 notebooks available; each with names correspoding to the work being done in them: 1. Downloading stock data using yahoo finance api in pandas_datareader library 2. Predicting stock price with n-historical data 3. Predicting stock price with traditional machine learning method Markdown cells are used throughout to assist in walking through the thought process for individual steps.
The folder prediction_engine contains the necessary files, .py, .html, .pkl files needed for the application to work.

Deploying

The application is already hosted on Heroku; however, you can follow the steps below to run the application locally.

• Clone the project

$ git clone https://github.com/emichris/Build-a-Stock-Price-Indicator.git

• Create a project environment (Anaconda Env Recommended)

$ conda create --name myenv # Assuming Anaconda is installed
$ source activate myenv

• Setup dependencies

$ pip install -r requirements.txt

• Start the server

$ cd Build-a-Stock-Price-Indicator/prediction_engine
$ python prediction_engine.py

Once the app is running, go to https://0:0:0:0:5000 to view the live app.

How it works

Once the app is running, enter a valid stock symbol eg AMZN for Amazon, TSLA for Tesla; you can include a start and end date, if start and end dates are not provided, the model defaults to historical data from 1/1/2015 to 1/1/2020 to train the model. After the model trained, you can select one or more stocks you entered before and enter a date to get a prediction of their price for that date. Note the query date must be beyond the end date used to train the model.

Results

The application was tested here using the following stocks, GOOG, TSLA, AMZN from 1/1/2010 to 1/1/2020 with a split of 80% training, 20% testing. K nearest neighbor outperformed linear regression receiving r2 scores of 0.97, 0.95, and 0.95 respectively. A modified version of Linear regression which uses only N number of points in the history for fitting the model as opposed to use all the historical was implemented here; although, the model performed significantly better than the traditional linear regression, the methodology is difficult to apply on predicting long distance futures that are longer than N.

Licensing & Acknowledgements

• MIT Standard License applies.
• Follow Christian Emiyah on LinkedIn
• I want to thank Udacity for the opportunity to work on this project as part of my Data Science Nanodegree Program.