Hi Greg,
Can you please explain me your remark "A huge insight for me which got me started is realizing that ROC() is a terrible way to measure change.". Why is ROC() a terrible way to measure change ? I know there are different ways to measure change but is not ROC() the most simple en straight forward way ? Being the first derivative of a function …
Kind regards, Ton.
----- Original Message -----Sent: Saturday, February 06, 2016 9:23 PMSubject: Re: [amibroker] Artificial intelligenceHi Ron,
The first thing I tell everyone is that the algorithmic method you use is probably the least important factor to having success creating a model. Several years ago I was Chief Product Officer at a company which used machine learning to create a model which recommended treatments to patients with various illnesses. We used random forests to create the models, but we also looked at support vector machines and a couple of other techniques. I wouldn't say random forests were the key to our success. The key to our success was that we had inputs no one else had, that we structured the inputs and outputs to make patterns easier to spot, that we did hand-tuning which reflected knowledge of our problem domain, and that we used the right objective function. Given all this other work, random forests gave us our best results but we had meaningful results across a variety of methods.With the portfolio mgt that I'm doin! g now, I've mentioned that I use bayesian networks. But I believe whatever success I am having I believe is not primarily due to the bayesian. It is primarily due to having done hard work figuring out how to represent the memory a security is showing, creating an objective function out of statistics which are much better than Monte Carlo, thinking deeper about how to categorize parameters than people normally do and treating different categories differently, etc. Only because all that was done do I think the bayesian network had a chance to provide real success.I've heard of people having success evolving trading systems with genetic algorithms, using support vector machines, using fuzzy logic, using hybrids, using Bayesian (obviously), etc. A good friend of mine once had success using a genetic algorithm to evolve a neural network. For my money (literally), I prefer to rely first on sound analysis of what markets are and insight into appro! priate ways to measure change, correlations, and change in correlations in markets. For me, that involves non-standard statistics and robust statistics, but others will have their own insights. A huge insight for me which got me started is realizing that ROC() is a terrible way to measure change.The reasons I chose a Bayesian network rather than a supervised learning method are that i) Bayesian networks are not black box. I know exactly what every nodes mean and what their relationships are; ii) Supervised networks (or random forests or SVM's) need retraining and have a tendency to either overfit or overgeneralize their data. I never need to retrain my Bayesian, as the statistics self-modify on a rolling basis, and I get to pick the level of generalization I the nodes represent. iii) I understand the semantics of the parameters and that allows me to be more intelligent in how I pick values for them (and whether or not to optimize a parameter).
__._,_.___
Posted by: "Ton Sieverding" <ton.sieverding@scarlet.be>
| Reply via web post | • | Reply to sender | • | Reply to group | • | Start a New Topic | • | Messages in this topic (18) |
**** 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