Hi Ron,
The short answer is yes, machine learning techniques can be very effective. For some background, I have a master's in AI from the early 90's and did my thesis on temporal learning. For the last couple of decades I haven't done much hands on programming, but have continued to invent mathematical techniques and oversee application of various statistical and learning techniques to a variety of problems. Five years ago I started applying some of these invented techniques to the problem of tactical allocation of portfolio assets. In walk-forward tests over eleven years the techniques produce an 85% CAGR in my core portfolio. In live management of the portfolio since 2011, the CAGR has been just over 65%. The difference is largely because the volume of money I need to move is too large to follow the models as accurately as I'd like. To close the gap, I'm currently building in AB with AFL to create a more flexible automated system I can scale and use to manage more portfolios. Caveat emptor: the live trading has not yet seen a true bear market.
That said, using machine learning effectively requires a good bit of judgement in selecting the technique, the inputs, and the structure of the learning system. There are many degrees of freedom when constructing the system, and it is a lot easier to make ineffective choices than effective ones. It is not as simple as throwing a bunch of data at a bunch of input and output nodes.
At the risk of telling you a bunch of stuff you already know, "neural network" is such a broad umbrella it is like saying you are going to build your software in a "programming language". If you are going to go the machine learning route for the first time, I'd recommend using a language like Python with a good library of techniques and do some exploratory work. What standard technique works best on a first pass representation of your data? Also, what inputs will work best? Raw data or pre-processed data? What kind of pre-processing? For example, I feed my system pre-processed data measuring market dynamics using a proprietary metric intended to capture the degree and direction of memory. My system never sees any raw or conventionally processed data.
Finally, and super importantly, if you are using a network learning approach, what structure does your learning network need to have? How many input nodes, output nodes, middle layers, re-entrant layers, etc. and why? These choices need to be dictated by a really good understanding of the structure of your domain and what sort of pattern you expect the learning system to capture.
__._,_.___
Posted by: rosenberggregg@yahoo.com
| Reply via web post | • | Reply to sender | • | Reply to group | • | Start a New Topic | • | Messages in this topic (2) |
**** IMPORTANT PLEASE READ ****
This group is for the discussion between users only.
This is *NOT* technical support channel.
TO GET TECHNICAL SUPPORT send an e-mail directly to
SUPPORT {at} amibroker.com
TO SUBMIT SUGGESTIONS please use FEEDBACK CENTER at
http://www.amibroker.com/feedback/
(submissions sent via other channels won't be considered)
For NEW RELEASE ANNOUNCEMENTS and other news always check DEVLOG:
http://www.amibroker.com/devlog/
This group is for the discussion between users only.
This is *NOT* technical support channel.
TO GET TECHNICAL SUPPORT send an e-mail directly to
SUPPORT {at} amibroker.com
TO SUBMIT SUGGESTIONS please use FEEDBACK CENTER at
http://www.amibroker.com/feedback/
(submissions sent via other channels won't be considered)
For NEW RELEASE ANNOUNCEMENTS and other news always check DEVLOG:
http://www.amibroker.com/devlog/
.
__,_._,___
EmoticonEmoticon