Came across this gem accidentally today. Here are the complete 20 lectures for CS 229 Machine Learning from Stanford University for my own record. I have learned most of the material before, but have forgotten most of it already.
In particular, the last lecture talks about POMDP. I have always thought it would be a great tool to be applied on financial data analysis. Long story short, POMDP is a distinctive machine learning algorithm because it inherently assumes the ‘world’ is partially-observable (MDP on its own takes care of the uncertainty part). A drawback with using artificial intelligence in trading is that A.I. are good for closed ended system but bad for open ended ones (i.e. trading). I believe POMDP can be worked on to become a valuable trading system.
I haven’t tried implementing it myself yet because of lack of a good platform and data. But since POMDP has been around for years now, I would assume somebody somewhere has tried this already. Unfortunately, I haven’t been able to find any information on this problem.
In any case, this post is merely a reminder for myself.
Related posts:

My name is Paul and I am a full-time engineer, part-time trader. Back in 2000, I deposited my $5000 interest-free student loan with an online broker. Since then, my interest in trading has become an obsession.
2 Comments
hello, paul
nice to read your good post
i have reached to your blog while finding pomdp articles
i’m trading currency (fx) and i’m trying pomdp way to that
as you know, there is various a.i. (many types of neural network , rl ,etc etc)
i have studied about that for long time
and now i’m trying pomdp , it will works well?
i hope so
good luck for your trading and see you on the market ~
Hi Paul,
just discovered your blog. Have been reading your articles with great interest. I haven’t heard of POMDP before, but they look interesting. In fact a fast search turns up several studies with stocks and POMDP.
Just one thing… You write:
Machine learning (ML) techniques, such as POMDP, (if you say “artificial intelligence”, more traditional rule-based systems could be understood) are based on statistical patterns in data, which implies that you don’t have to know everything about the system, you just need enough data.