I’m working through the AWS Machine Learning Engineer Nanodegree through Udacity (https://www.udacity.com/course/aws-machine-learning-engineer-nanodegree–nd189) and today I submitted my project proposal (https://github.com/neilsimon/nd009t-capstone). It’s a little different in that I’m trying out a mechanism that is actually intended for a different purpose (it’ll be the subject of future posts), but it just so happens that it will work for this too.
So, what is unusual about the mechanism I’m using? Well, I am combining two different machine learning approaches. I am breaking up the data into different categories using a K-means categoriser and then using tabular predictors trained on the different categories, with the hope that the resulting system will give better overall performance as each predictor will be trained on a narrower set of data.
I suspect that in this case, predicting the price of gold, that my system won’t work very well. We’re up against the efficient market hypothesis, and I suspect that for the past 15 years (over which I am training/testing my system), there have been much smarter systems making sure that any potential profit to be had in making technical predictions of future gold value are already being taken.
I actually went through one of Udacity’s free courses, Machine Learning for Trading (https://www.udacity.com/course/machine-learning-for-trading–ud501) as taught by Tucker Balch (https://www.linkedin.com/in/tuckerbalch/) to get some background for the above. He’s an interesting character, and certainly helped make the material a little less dry than it would otherwise have been. To be fair, this course was a little out-of-date but for a free course, it was useful and helped me figure out how I might go about this project.
Anyway, please feel free to check out the github repo (https://github.com/neilsimon/nd009t-capstone), and by the time anyone actually reads this, hopefully I will have completed it.