Hi All, My lab partner and I have been working on some advanced implementations of machine learning in finance (for our MSFE), particularly the work of Dr Lopez Marcos de Prado, and have built a couple of open source implementations that we think the community in general will benefit from. In particular: An implementation of Meta-Labeling: This is a secondary model that determines if a primary model such as a discretionary portfolio manager, or a technical trading strategy, is correct or not. In short it helps to improve all performance metrics such as drawdowns, sharpe ratio, calmar and so on. It is also very useful when paired with bet sizing algorithms where the primary model determines the side of the trade and the secondary model the size. Slide show presentation Research report: Does Meta-Labeling Add to Signal Efficacy? Our hackathon presentation on machine learning used in portfolio management: Hierarchical Risk Parity Portfolio Optimization We are busy implementing an open source python package called mlfinlab, which contains the source code for various implementations of the ideas in the text book: Advances in Financial Machine Learning. We find the research fascinating as it incorporates several skill sets, from HFT techniques, market microstructure, portfolio optimization, to machine learning algorithms and ensembles. The following links to a post regarding where we are currently in the open source package. I would like to stress that we are not selling anything. Its all open source and a labor of love. It would be great to open up a dialog around these topics. (Sorry I posted this in programming but I realize that I should have rather posted it here, not sure how to delete the old thread?)
OP. Interesting work. What kinds of research of his did you find practically useful? Just looking at the Hudson presentation slides, shows Equal Weight TW performed much better than all of the competitors, including HRP. If you leverage up HRP, you get similar results. I.e. same Rwd/Risk ratios.
@Overnight thank you for the help. It worked and the duplicate post has since been removed. @dtrader98 thank you very much. At the moment I found the ideas around meta-labeling to be the most interesting. Mainly because of the number of applications the model has. I would like my next research to extend the project by having a look at adding bet sizing algorithms on-top of the meta model. The most exciting thing for me is that so much of de Prado's research works out-of-sample, which can't be said for all research, example mean variance portfolio optimization. Another part we would like to explore are structural breaks and micro-structure features for models. On your note of Equal weight performing better (HRP does better, did I maybe miss a slide?), the performance is largely due to the well diversified set of ETFs we used. If we had to add more ETFs that were very similar to each other then the Equal weight strategy would not be diversified and perform much worse. One of the key things to note is HRP's out-performance relative to all the other strategies employed in the slides. HRP out-performs Equal weighting on all of the performance metrics, especially the risk adjusted returns.
Thanks for this! The source code for re-barring data based on various metrics (volume, N-ticks etc) is nice and tidy. FWIW, I am not a great believer for machine learning applications to finance. There is a lot of noise in the buy-side community but I don't know anyone making money on it aside from HFT firms and they have been doing it for years. Then again, linear regression is considered machine learning too - by that metric I use a lot of ML myself.
You think? I have not read anything so don't have an opinion, to be honest, but he seems to have a history of working at various fancy places.
Count me as listening closely for a reply to sle's question. I know only the nice names and the quality presentation (cited above). I have not calculated anything here in the least iota.
I thought the book was better than a lot of stuff out there, although I did find it rather odd how he only used classification and completely ignored regression. Personally I have found there is a lot of information loss when using classification especially at non-HFT time scales. Maybe the info in the book is more useful with the assets he trades but it's definitely missing some necessary steps if trying to do ML in equities which is what I trade.