DIY MachineLearningStocks

[ machine-learning ]

I recently released a machine learning stock prediction project on GitHub, unimaginatively named MachineLearningStocks. It is a project that I’ve put quite a lot of time into, and is in fact a simplified version of a system that I’ve been using to live trade. This post doesn’t really offer anything on top of the existing readme, but I figured it would be good to have a copy (with some minor changes) here as well.

MachineLearningStocks is designed to be an intuitive and highly extensible template project applying machine learning to making stock predictions. My hope is that the project will help readers to not only understand the overall workflow of using machine learning to predict stock movements, but also to appreciate some of the many subtleties. I’ve also provided a long list of possible improvements you could make, or ideas for moving forward, to put some spice into the vanilla project.

Concretely, we will be cleaning and preparing a dataset of historical stock prices and fundamentals using pandas, after which we will apply a scikit-learn classifier to discover the relationship between stock fundamentals (e.g. PE ratio, debt/equity, float, etc) and the subsequent annual price change (compared with the an index). We then conduct a simple backtest, before generating predictions on current data.

While I would not live trade based off of the predictions from this exact code, I do believe that you can use this project as starting point for a profitable trading system – I have actually used code based on this project to live trade, with pretty decent results (around 20% returns on backtest and 10-15% on live trading).

This project has quite a lot of personal significance for me. It was my first proper python project, one of my first real encounters with ML, and the first time I used git. At the start, my code was rife with bad practice and inefficiency: I have since tried to amend most of this, but please be warned that some minor issues may remain (feel free to raise an issue, or fork and submit a PR). Both the project and myself as a programmer have evolved a lot since the first iteration, but there is always room to improve.

As a disclaimer, this is a purely educational project. Be aware that backtested performance may often be deceptive – trade at your own risk!

Star

Contents

Overview

The overall workflow to use machine learning to make stocks prediction is as follows:

  1. Acquire historical fundamental data – these are the features or predictors
  2. Acquire historical stock price data – this is will make up the dependent variable, or label (what we are trying to predict).
  3. Preprocess data
  4. Use a machine learning model to learn from the data
  5. Backtest the performance of the machine learning model
  6. Acquire current fundamental data
  7. Generate predictions from current fundamental data

This is a very generalised overview, but in principle this is all you need to build a fundamentals-based ML stock predictor.

Quickstart

For readers who want to throw away the instruction manual and play immediately, just clone the project, then download and unzip the data file into the same directory. Then, open an instance of terminal and cd to the project’s file path, e.g

cd Users/User/Desktop/MachineLearningStocks

Then, run the following in terminal:

pip install -r requirements.txt
python download_historical_prices.py
python parsing_keystats.py
python backtesting.py
python current_data.py
pytest -v
python stock_prediction.py

Otherwise, follow the step-by-step guide below.

Preliminaries

This project uses python 3, and the common data science libraries pandas and scikit-learn. A full list of requirements is included in the requirements.txt file. To install all of the requirements at once, run the following code into terminal:

pip install -r requirements.txt

To get started, clone this project and unzip it. This will become our working directory, so make sure you cd your terminal instance into this directory.

Historical data

Data acquisition and preprocessing is probably the hardest part of most machine learning projects. But it is a necessary evil, so it’s best to not fret and just carry on.

For this project, we need three datasets:

  1. Historical stock fundamentals
  2. Historical stock prices
  3. Historical S&P500 prices

We need the S&P500 index prices as a benchmark: a 5% stock growth does not mean much if the S&P500 grew 10% in that time period, so all stock returns must be compared to those of the index.

Historical stock fundamentals

Historical fundamental data is actually very difficult to find (for free, at least). Although sites like Quandl do have datasets available, you often have to pay a pretty steep fee.

It turns out that there is a way to parse this data, for free, from Yahoo Finance. I will not go into details, because Sentdex has done it for us. On his page you will be able to find a file called intraQuarter.zip, which you should download, unzip, and place in your working directory. Relevant to this project is the subfolder called _KeyStats, which contains html files that hold stock fundamentals for all stocks in the S&P500 between 2003 and 2013, sorted by stock. However, at this stage, the data is unusable – we will have to parse it into a nice csv file before we can do any ML.

Historical price data

In the first iteration of the project, I used pandas-datareader, an extremely convenient library which can load stock data straight into pandas. However, after Yahoo Finance changed their UI, datareader no longer worked, so I switched to Quandl, which has free stock price data for a few tickers, and a python API. However, as pandas-datareader has been fixed, we will use that instead.

Likewise, we can easily use pandas-datareader to access data for the SPY ticker. Failing that, one could manually download it from yahoo finance, place it into the project directory and rename it sp500_index.csv.

The code for downloading historical price data can be run by entering the following into terminal:

python download_historical_prices.py

Creating the training dataset

Our ultimate goal for the training data is to have a ‘snapshot’ of a particular stock’s fundamentals at a particular time, and the corresponding subsequent annual performance of the stock.

For example, if our ‘snapshot’ consists of all of the fundamental data for AAPL on the date 28/1/2005, then we also need to know the percentage price change of AAPL between 28/1/05 and 28/1/06. Thus our algorithm can learn how the fundamentals impact the annual change in the stock price.

In fact, this is a slight oversimplification. In fact, what the algorithm will eventually learn is how fundamentals impact the outperformance of a stock relative to the S&P500 index. This is why we also need index data.

Preprocessing historical price data

When pandas-datareader downloads stock price data, it does not include rows for weekends and public holidays (when the market is closed).

However, referring to the example of AAPL above, if our snapshot includes fundamental data for 28/1/05 and we want to see the change in price a year later, we will get the nasty surprise that 28/1/2006 is a Saturday. Does this mean that we have to discard this snapshot?

By no means – data is too valuable to callously toss away. As a workaround, I instead decided to ‘fill forward’ the missing data, i.e we will assume that the stock price on Saturday 28/1/2006 is equal to the stock price on Friday 27/1/2006.

Features

Below is a list of some of the interesting variables that are available on Yahoo Finance.

Valuation measures

Financials

Trading information

Parsing

However, all of this data is locked up in HTML files. Thus, we need to build a parser. In this project, I did the parsing with regex, but please note that generally it is really not recommended to use regex to parse HTML. However, I think regex probably wins out for ease of understanding (this project being educational in nature), and from experience regex works fine in this case.

This is the exact regex used:

r'>' + re.escape(variable) + r'.*?(\-?\d+\.*\d*K?M?B?|N/A[\\n|\s]*|>0|NaN)%?(</td>|</span>)'

While it looks pretty arcane, all it is doing is searching for the first occurence of the feature (e.g. “Market Cap”), then it looks forward until it finds a number immediately followed by a </td> or </span> (signifying the end of a table entry). The complexity of the expression above accounts for some subtleties in the parsing:

Both the preprocessing of price data and the parsing of keystats are included in parsing_keystats.py. Run the following in your terminal:

python parsing_keystats.py

You should see the file keystats.csv appear in your working directory. Now that we have the training data ready, we are ready to actually do some machine learning.

Backtesting

Backtesting is arguably the most important part of any quantitative strategy: you must have some way of testing the performance of your algorithm before you live trade it.

Despite its importance, I originally did not want to include backtesting code in this repository. The reasons were as follows:

Nevertheless, because of the importance of backtesting, I decided that I can’t really call this a ‘template machine learning stocks project’ without backtesting. Thus, I have included a simplistic backtesting script. Please note that there is a fatal flaw with this backtesting implementation that will result in much higher backtesting returns. It is quite a subtle point, but I will let you figure that out :)

Run the following in terminal:

python backtesting.py

You should get something like this:

Classifier performance
======================
Accuracy score:  0.81
Precision score:  0.75

Stock prediction performance report
===================================
Total Trades: 177
Average return for stock predictions:  37.8 %
Average market return in the same period:  9.2%
Compared to the index, our strategy earns  28.6 percentage points more

Again, the performance looks too good to be true and almost certainly is.

Current fundamental data

Now that we have trained and backtested a model on our data, we would like to generate actual predictions on current data.

As always, we can scrape the data from good old Yahoo Finance. My method is to literally just download the statistics page for each stock (here is the page for Apple), then to parse it using regex as before.

In fact, the regex should be almost identical, but because Yahoo has changed their UI a couple of times, there are some minor differences. This part of the projet has to be fixed whenever yahoo finance changes their UI, so if you can’t get the project to work, the problem is most likely here.

Run the following in terminal:

python current_data.py

The script will then begin downloading the HTML into the forward/ folder within your working directory, before parsing this data and outputting the file forward_sample.csv. You might see a few miscellaneous errors for certain tickers (e.g. ‘Exceeded 30 redirects.’), but this is to be expected.

Stock prediction

Now that we have the training data and the current data, we can finally generate actual predictions. This part of the project is very simple: the only thing you have to decide is the value of the OUTPERFORMANCE parameter (the percentage by which a stock has to beat the S&P500 to be considered a ‘buy’). I have set it to 10 by default, but it can easily be modified by changing the variable at the top of the file. Go ahead and run the script:

python stock_prediction.py

You should get something like this:

21 stocks predicted to outperform the S&P500 by more than 10%:
NOC FL SWK NFX LH NSC SCHL KSU DDS GWW AIZ ORLY R SFLY SHW GME DLX DIS AMP BBBY APD

Unit testing

I have included a number of unit tests (in the tests/ folder) which serve to check that things are working properly. However, due to the nature of the some of this projects functionality (downloading big datasets), you will have to run all the code once before running the tests. Otherwise, the tests themselves would have to download huge datasets (which I don’t think is optimal).

I thus recommend that you run the tests after you have run all the other scripts (except, perhaps, stock_prediction.py).

To run the tests, simply enter the following into a terminal instance in the project directory:

pytest -v

Please note that it is not considered best practice to include an __init__.py file in the tests/ directory (see here for more), but I have done it anyway because it is uncomplicated and functional.

Where to go from here

I have stated that this project is extensible, so here are some ideas to get you started and possibly increase returns (no promises).

Data acquisition

My personal belief is that better quality data is THE factor that will ultimately determine your performance. Here are some ideas:

Data preprocessing

Machine learning

Altering the machine learning stuff is probably the easiest and most fun to do.

Contributing

Feel free to fork, play around, and submit PRs. I would be very grateful for any bug fixes or more unit tests.

This project was originally based on Sentdex’s excellent machine learning tutorial, but it has since evolved far beyond that and the code is almost completely different. The complete series is also on his website.