Hi SentDex, first of all thanks for all the material that you shared here, I'm quite new to programming and your tutorials really helped me out.
I have a passion for trading and python helped my to move from analysis on excel to operative strategies. I'm not a big fun of buy'n hold strategies but I was very impressed from your tutorial on scikit learn and stock investing. I do developed two trading strategies with python, one momentum based and one mean-reverting, for the moment working quite well.. FINGHERS CROSSED.... I could be another victim of the arcsine law!!
I do belive that even if the technical analysis is usually not used with a scientific approach - most of technical traders would be better defined as discretional traders - there are ways to get some value out of it.
So here is my question, is there a way to conceive a machine learning task in order to see if a mix of technical indicators (or better trading rules) can actually predict the market? I try to explain better, in the example of fundamentals you used stock prices and fundamental data from the stock to understand if those features could predict the outperformance of the stock over the market. Can I do the same with tecnical indicators? I guess the answear is yes but I cannot conceive a way to do it, in the case of fundamentals you gave yourself the timeframe of 1 year, so you tested how well the fundamentals would predict the performance in T1. Trading with tecnical indicators it's about understanding if and indicator or a weighted mix of indicators and take profit or/and stop loss rules can deliver reliable signal to consistenly generate profits.
How to brake down this testing objective in smaller problems? At the end of the day this should be a classification problem, what set of rules can draw a hyperplane between winning and loosing trades?
A problem in the conception of the test for example is that I don't have a fixed timeperiod to test against.. I can throw a bunch of indicators in the timeseries but how to fix a target for the prediction?
I understand that's no straitforward answear to my question and maybe I'm out on context here but I'd like to know what are your thougts about this matter.
Tanks again
Diego
You must be logged in to post. Please login or register an account.
Machine learning, while coming in and out of the spotlight for the last few decades, has been around for a very long time and has been implemented in a large variety of ways in the markets. The issue with the markets is that they are zero-sum. You don't really beat the market as most people like to say, you beat other people if you're attempting to be a trader. If the other people are doing what you are doing, you wind up not winning anything.
Thus, before embarking on an idea you have, the first thing you'd want to consider is its complexity, novelty, and what the likelihood that it's been done before is. In your case, the idea of picking which indicator to trade off of isn't even a machine learning task, it's simply a tracking task with a very rudimentary algorithm, which is also a lagging one at best, as your "best choice" would have actually been the best choice in the last week, 30 days, or whatever your window is, not necessarily into the future.
The best thing I can tell you is that, if you are going to try, you need to try blind-testing the algorithm as quickly as possible.
I can write some code that will find the best investing/trading methods over the last 3 decades, and beat the market thousands of times over, but this would be a case of data snooping. Its very hard to avoid such a thing with the line of thinking you're going down.
-Harrison 9 years ago
You must be logged in to post. Please login or register an account.