A system to data mine alpha and make money.
When you have a trading idea, it's usually something you can write in words such as: "If I take out a trade when X happens, will I make money?" There is not a good backtester or computer program to then take this, digitize it, and answer your question in a consise way. That's what the Alpha Engine is for. The goal is to take a question about a trading signal and return a decent answer which allows you to decide whether or not it is in fact a useful signal.
From there, the Alpha Engine can be automated by things like genetic algorithms that come up with the questions automatically and systems that filter only certain results based on the metrics that come out of it.
The goals of the project are:
- Simplicity
- Extendability
- Speed
Therefore, the core focus of the engine itself is purely on the question / answer loop. Once we achieve the 3 goals of the project, we can move onto the 'fun' stuff like extending the search space, etc.
The idea is to take raw price data, put it into a Numpy array, and calculate the t+1 turns. This is the main data that we work with.
From there, we can calculate transforms and additional signals like EMA, etc.
After that, we come up with our question. Our question takes the form of a piece of boolean logic. This logic is used to select the places in data where it is true. Aka, our entries. This just takes advantage of Numpy slices: https://numpy.org/doc/stable/user/absolute_beginners.html#indexing-and-slicing
From there, we know the immediate t+1 returns. We then can have some basic rules for 'exits.' Let's assume it's just time based for simplicity now (5 days or so). This is easy because then we can just roll the returns forward t+5 time steps and do a cumulative sum to get the total returns for the time period.
This method makes it extremely fast and vectorizable to search and calculate massive amounts of data.
For example, the shape of a single stock of daily returns over 20 years is ~(5000, 8). This assumes the features of year, month, day, open, high, low, close, and forward returns.
Across the stock universe, there is probably about 6000 stocks on the NASDAQ and NYSE. This creates a Numpy array of (6000, 5000, 8). This, in machine learning and data mining terms, is a very small matrix.