Curve fitting is desirable and necessary. Strategies can't exist ......without fitting buy and sell signals to price data to maximize a fitness function's measure of an Equity Curve. Curve fitting is a good thing, not a bad thing.
If you're serious about learning about machine learning consider joining asirikuy.com. They're currently exploring ML for automated trading and using GPUs (parallel processing) to do the number crunching. The site owner is very rigorous about issues such as snooping, data mining bias, selection bias, etc. He has a strong background in math and prob/stat. Plus he's a nice guy.
Interesting. I don't really see myself using ML for trading, but rather, for other unrelated problem solving.
Well about half way through my ML Coursera course. It would seem all the attention is on Deep Learning through Convolutional Neural Networks that were popularized by Hinton from UofToronto. When it was used to win a image classification challenge in 2012 and crushed the previous best result by a wide margin. Since then it's been a land grab for the pioneers in the field and their research assistants at the various universities. Not to mention Deep Learning startups are all over the place. There was even an AI for AI startup(Whetlab) but it got bought up within a year by TWTR. I'll be taking a Deep Learning Tutorial from Stanford next and see where that takes me.
Finished the ML Coursera course which was good. Been watching/reading a lot of stuff on Deep Learning. Apparently it is necessary to have a powerful computer system as most of the neural network algorithms involve a lot of matrix calculations and are best handled by multiple GPU's. this can be farmed out to the cloud services but then that adds network delay. Nividia(https://developer.nvidia.com/devbox) is all over this and providing Deep Learning GPU systems. they certainly make some trading computers look like child's play. The other thing is DL is mostly something you need very large data sets(more data the better) to train from, so this is why the big internet companies are leading the way in DL business applications. It still seems to be early days for DL but Google just announced they are using it for practically everything. Google released their in-house software library Tensorflow to open source. Interesting times ahead..
It can't be as interesting as my class in "Intro to Telecommunications" a number of years ago. My kooky professor wasted a whole evening's lecture talking about some DARPA network that was opened to the public. It was using what seemed to be a chaotic and unstable "packet" switching technology instead of "locked on" phone circuits. He said the nodes were growing exponentially... Later on, I go to the computer lab. There was this window based app called Mosaic. I fire it up and I get nothing. I tell myself: "Well, this won't amount to anything. Mobile phones will change the world!"
This thing is hot! A self-christened MATLAB deep learning toolbox had 748 downloads within 30 days. The next hottest download is Deep Neural Network had which had 477 downloads. In comparison, the Algorithmic Trading walk-forward analysis tool had 250+. And to think that MATLAB's strengths are in numerical analysis rather than computational statistics. Barely, 100 downloads on engineering apps. Last month, the "smart guy" in my class was carrying a 6-inch thick Neuroscience book. I was wondering why a math major would punish himself with such esoterica. C4buckley was asking last month about this topic, too. I need to start hanging with the cool kids. Looks like I may be taking Intro to Telecommunications all over again. Nice post, Cdntrader. You learned me some.