How to (ACTUALLY) Learn Machine Learning (and use it for trading).

_________________

Motivated by @BeingHorizontal's noble albeit suboptimal thread on the same. https://twitter.com/BeingHorizontal/status/1248870718168690691

Learn a programming language that lends itself well to data science (like Python, R or Julia).

Also, go beyond the data science libraries - learn how to build software and data infrastructure. If you don't, you probably won't be able to do much with your ML model.

1/n

Understand what machine learning is and isn't. There is a lot of hype surrounding ML/AI and that has skewed people's expectations of it.

It's not magic - just clever mathematics that harnesses our access to computing power and lots of good data.

2/n

If you do not understand the mathematics behind ML algorithms, you are very likely to fuck up, especially when working with something as chaotic/noisy as price data.

Study Linear Algebra, Calculus, Probability & Statistics. Use @MIT OCW lectures for this.

3/n

Data Science/ML has more to it than just throwing algorithms at a dataset hoping something fits. You need to understand the dataset you're working with.

Study the nature of financial time series data and its statistical properties.

4/n

An often overlooked but essential part of the ML pipeline is feature engineering i.e. modifying input data to make it more amenable to modeling. This requires a knowledge of data mining and data transformation.

Remember, bad input features can derail your ML project.

5/n

Using ML in trading, particularly for signal detection/strategy design is far more challenging than you probably realize, particularly because most ML algorithms suck at dealing with the kind of properties financial time series data exhibits (chaotic/noisy, non-stationary).

6/n

Most blogs you'll read showcasing great results on price data are trash (at least their results are; the blog itself might be useful to learn the concept).

Realize that the odds of someone blogging on Medium about their (truly) profitable ML trading model are nil.

7/n

On a similar note, do not conflate accuracy with effectiveness when using ML for trading. If anything, less accurate ML-based trading signals perform better than ones with very high (~90%) accuracy. Think hard about how you will evaluate the performance of your model.

8/n

Know what ML algorithm suits the task at hand. Trying to use deep neural networks for a problem you have 100 data points for will only yield trash.

You'll probably get much more mileage out of simpler models if you understand the process and your data [SEE ATTACHED PICTURE]

9/n

Go through JP Morgan's guide on Machine Learning and Big Data (this is also a great starting point if you know ML but aren't sure how to apply it to financial datasets).

Link: https://www.cfasociety.org/cleveland/Lists/Events%20Calendar/Attachments/1045/BIG-Data_AI-JPMmay2017.pdf

10/n

Read @lopezdeprado's paper "THE 10 REASONS MOST MACHINE LEARNING FUNDS FAIL". It's highly likely to foreshadow the mistakes you're going to make using ML for trading.

Link: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3104816

11/n

Read the following books (in order):

- 100 Page Machine Learning Book by @burkov

- Introduction to Statistical Learning

- Elements of Statistical Learnings

- Deep Learning by @goodfellow_ian

- Reinforcement Learning by Sutton

12/n

Most importantly, quant trading is a multi-faceted endeavor and trying to simultaneously learn a new skill set and apply it to a challenging problem is going to be painful.

Find good teams/groups/people/mentors to work with. It'll vastly accelerate your progress.

n/n