I appreciate the interest shown for Hong Kong companies in an earlier post.
In fact, according to the paper which I am now posting (in which ZGWZ post a table and don't mind me sharing the whole paper to all P123 members from ScienceDirect).
One of the key findings is that the performance of machine learning portfolios differs greatly across countries—the COMB (a combination of machine learning models tested) Sharpe ratios can be as low as 0.20 in Austria and as high as 2.65 in Hong Kong.
Intriguingly, as an important dividing line for performance runs between the developed and emerging markets. Contrary to the common narrative, return predictability is visibly stronger in developed countries. This may seem surprising since
inefficient emerging markets are frequently regarded as a reservoir of exploitable anomalies; however, the link between market maturity and mispricing is unequivocal. In developed markets, which are usually bigger, the machine learning algorithms may gain from richer datasets—which allow for better training.
Jim @Jrinne : I have also sent it to your email box and decided to post this message before your reply. Pls kindly add your comments if there are any.
Thanks for sharing this paper. Actually Adam Zaremba is my former colleague from work at university.
Below you will find other interesting findings:
Interestingly, as most of the academic papers, they do not use embargo period as a baseline method. As additional robust check, they applied one-month skip period (embargo) to assess if the model with short-lived information (e.g., short-term reversal) is still robust:
I am surprised that one of the authors is actually your former colleague.
I think you have made a good point about OLS given the less complexity (resources intensive) vs. reward (performance) and that Marco should consider the allocation of IT resources to different models/strategies.
The paper talks about skip periods rather than embargo periods. The idea is to avoid predictable returns that are difficult to trade. The part you quoted says that R-squared is not appropriate for evaluating economic effects, but P123's system already provides indicators of economic effects. I do agree with the support for custom sorting metrics, though.
