NNs are curve fitting algorithms. Nothing to gain from that unless forward sample is i.i.d. and it's never that. You should do your own homework and please refrain from irrational claims with no proof/record.
I've already posted my record via a C2 screen shot a while back. You can search for it if you're interested. My annualized ROI is now much high than that original post. My ROI is a result of a lot of homework regarding machine learning/NN's etc. I don't use options/futures/etc. only ETF's. No martingale. Pure, simple, trades. Again, just because Johnny can't do it; doesn't mean it's impossible or that Jimmy can't do it. The fact is, it's being done wether you or others believe it or not. And as a result of 'doing it' for a while now, I've given up on doing the small bit of convincing I used to try to do. Now, I simply indicate that it can (if I say anything at all), and is, being done successfully. Honestly, the more that believe it to be impossible, the better it is for those of us that know better. Most recently, I just read forums like these and smile--and think back to the days when I too wondered whether the markets can be beat; but refusing to accept that they couldn't be. I appreciate the journey. Everyone is free to believe as they wish. And I have no desire to continually prove my success to those that doubt without evidence--to those that insist others can't do something simply because they themselves can't do it. Those folks, I've found, are usually a waste of time/effort. I wish all good luck though.
2 question: 1) Can you please comment on the target variable you are using, I assume you are not trying to predict the price and most likely it's a binary variable, are you predicting if next day ETF price will be up(1)/down(0) or maybe outperform(1)/(0)under perform the index? 2) Are you using Python? what library?
I haven't designed an algo yet to forecast the price, although I hope to get a round tuit one day. I forecast the change between the next day, and the following day. I run the algo after the close, so I can't take advantage of only a next day forecast. The output possibilities are 1, 0, -1. 1 is long; 0 is cash; -1 is short. The algo rarely goes to cash. I prototyped in Excel VBA w/ Solver. I started to port it over to Python, but decided Java would be better. So I started porting it over to Java when I had more ideas as to how to improve the algo further. So I'm technically still prototyping in Excel VBA. I don't use any standard libraries. I looked into them; and they can't do what I need them to do. I've concluded that only custom code can do it that way I'm doing it. I haven't gone over a month without making some sort of improvement. So, once I run out of ideas to improve it, I'll finally code it in Java...where I'll be able to add some additional improvements I can't add right now because Excel VBA is too slow/clumsy/etc.
I tried it too in the 90's and came to the same conclusion. I asked support from math wizards to feed the NN as it is not that easy to give the correct input. We never saw any noticable result. I never heard from someone able to use it for trading. So it is highly probable that is does not work for most people. NN's cannot not be used by the average Joe, it is not a miracle solution where you should just push a button and wait for the optimal system to be created by the NN. It is very complex.
Thanks, finally some quality talk. If i understand you correctly you are forecasting the change between t+1 and t+2. Do you forecast the probability of being long (1) and then put a threshold on it to split in to other classes 0 and -1? something like 0.55 (1) > 0 > 0.45 (-1). It's quite surprising that you are building the whole NN without any libraries, what is your background? Also I would like to make some assumptions regarding the factors that you are feeding as training features, I guess those are some lagged prices of particular instrument being traded and other similar ETF plus some price derivatives like MA's, volatility etc.? It would be interesting if you could comment on that.
Likewise. Correct. But to be clearer, I forecast the sign of the change, or zero (cash). Scores are calculated. But I haven't looked into whether they are similar to probabilities (yet?). You would think they would have to be, but I can't be that certain. Above a certain score: 1. Below a certain other score: -1. Else 0. EDIT: A simple NN is not difficult to code. Self-taught re: programming. Took college courses and subsequent employment after I was already self-taught. I mention that because I think being self-taught helped me to think outside of the box. (Started coding pre-teen on TS-1000 ) The algo should work on any instrument that exhibits price patterns. I currently trade NUGT/DUST, So I plug the underlying (GDX) into the algo. I only feed it EOD: DATE OHLCV data.
Is CkNN your algo? I'm surprised that your features and the model are fairly simple. I would assume that you use some optimization of parameters in the network. What error metric do you use? The reason why I'm curious is that I've been testing some nonlinear models and surprisingly they usually under perform a simple linear regression applied to clustered data, however in theory it should be the other way around. Perhaps I should have a look at NN.