I came across an excellent short introduction to machine learning–its limitations and potential. If you’re interested, it’s here: https://www.aqr.com/Insights/Research/Alternative-Thinking/Can-Machines-Learn-Finance?
This is interesting insight from Patrick O’Shaughnessy on Ted Seides podcast about their learning curve with machine learning and it’s limitations…
At the ~ 23:55 mark
No meat and obviously intended for an audience with no training in machine learning or statistics. Most ideas presented are not mainstream. Clearly, intended as a marketing tool.
Truth is there are some very simple, mainstream machine learning ideas that could be implemented at P123. Ideas that are just supplemental to what we are already doing.
But why would we be discussing machine learning in the forum? We know that even simple metrics that are used for the most basic of machine learning tasks (like the information ratio) will never be adopted.
I will leave the important and useful discussion about neural-net recognition of cats (in the article) to others.
-Jim
Yeah. There is some simple AND USEFUL stuff in here so: THANK YOU CARY!!!
O’Shaughnessy notes that linear regressions are not so good for nonlinear data. Hmmm. Where have I heard that?
Also, notes that non stationary data does not work. People who do time series might take note. But where have I heard that before?. We still see multiple examples of uses of time series that use non stationary data where the statistics are just wrong and completely inappropriate at P123: but the statistics look so good they have to be believed.
And no doubt, if you listen to it Patrick O’Shaughnessy is doing mainstream MACHINE LEARNING and makes no apologies. And he is sold on it working. “If the dataset is large enough a quant will find it.” Meaning the signal.
I do not think Patrick is different from his father. There are just some new methods out there.
-Jim
“All [dynamic resource allocation] examples need to deal with changing inputs and environments, which are highly dynamic and difficult to estimate and predict, as the future load is not statistically dependent on the current load,” says Eiko Yoneki, a senior researcher leading the data centric systems group at the University of Cambridge’s Computer Laboratory. “One change triggers another change, and if you want to control the system with accurate decisions, one must consider the future status of the system.” The trouble is that most existing methods rely on historical data to make predictions.
http://www.bbc.com/future/story/20190606-the-maths-problem-that-modern-life-depends-on