The machine learning (ML) literature sorts problems into two categories: (1) regression -- the response is a continuous numerical output (2) classification -- the response is discrete, often binary When applying ML to financial markets, you can treat return prediction as a regression problem or as a classification one (up or down). What have you found to be more effective? In general, what ML algorithms have you found useful? I have not used ML in trading yet, but in my research on predicting daily returns using medium size data sets of about 2000 points, I have found that support vector machines often outperform multiple linear regression.
Ideally your model will predict both the mean and volatility of returns. Exponentially weighted realized volatility and GARCH models have been used to forecast return volatility. I wonder if ML approaches can do better at forecasting volatility. A trader should be aware of upcoming events such as earnings announcements, central bank meetings, or economic releases that can boost volatility.
Even if you could ML such things as events, earnings, meetings, etc...it`s still next to impossible to forecast upcoming volatility and base your stops on it.ML will still curve fit stops to the past optimums. That`s all there`s about trading failure - the unknown volatility.You could predict that - your a billionaire in no time.
Regression is useful to do measurements that need to have some granularity, ie. measurement of slope and variance. Classification can be thought of as probability of something belonging to different classes, ie. binary outcomes. So generally, both categories can be useful. Even though they can sometimes overlap some, one of them is usually much more useful for some classes of problems than the other. Depends on the problem and the output you need in order to solve that problem, which is often a subproblem of a much bigger and more complex problem. If you're not exactly sure of the problem you're trying to solve, well, that's a different problem! *puts in box "Tool In Search Of A Problem"*
Hey -- not to get all geeky or anything, but there is a material issue here, between regression and classification. Regression depends on the Central Limit Theorem -- the Gaussian/Normal Distribution of bell-shaped curve fame. Symmetric, defined deviation frequencies -- even its exceptions (Student's t) note the CLT basis. "Classification" carries no such ties, and can be used with central-tendency data, uniform-distributions, bimodal, multi-modal -- bears no impact. A second difference (again, just to get nerdy-geeky-mathy) is that, as part of the normal distribution, you have ready access to a continuous function. Open any stat book, and there will be interpolation tools to go into data where the data are not defined in a continuous fashion. (Technically, a regression on option *strikes* would fit this, as the range between the strikes would have to be interpolated to one side or the other.) But with "classification" -- such "border issues" could routinely occur. Okay. I need a beer. I'll relax.....
I have found many different machine learning algorithms to be useful, but I have found that the actual machine learning part of any project has less effect on the quality of the solution than things like data collection or deciding how to apply machine learning to the problem (which is not always immediately obvious).
http://www.wildml.com/2018/02/introduction-to-learning-to-trade-with-reinforcement-learning/ Anyone interested in implementing DQN with the Microsoft Cognitive Toolkit in C# please msg me.