Pretty much. And it's not a problem only in finance. NN is pretty much not great for anything with a decent amount of randomness in it. Throw in non-stationarity, and it's all over. On the other hand, it's great for things like like character recognition and image processing. It's all about problem domain.
Exactly. They're pretty good for pattern recognition applications like OCR, etc. But in trading there *are* statistically significant price movement patterns. For instance, multi-candle candlestick patterns such as "Three White Soldiers" etc, could be quickly and easily identified with a NN.
At a place where I used to work, the only thing we used NNs for was volatility interpolation - and I am not convinced that they were the best tool for that job either.
??? You are always curve-fitting against the statistical measure of your choice. The output of the implementation may be a bool, but the selection criteria generating the bool is not. This can be said with any optimization logic and algorithm. Am I missing anything here? Unlike you, it's a pretty retarded post.
A guy named Hessler in Chicago used to sell picks from his neural network derived system circa ~1998. It would spit out one or two stocks a day, then it would spit out two hundred about every week or two. He calculated his gains as if a user of his service was putting the same amount of money on days with one pick or hundreds which was clearly impossible with listed stocks. I generated some ways to take his outputs and select from them by hand, it was something really simple based on daily charts, I can't recall what it was, some sort of simple pattern.. whatever... and it worked well really... The user of his system was supposed to hold until he said to sell which was a disaster on occasion, I don't think he was paying attention sometimes.. I filtered his picks and bought at the end of day and sold them the next morning... it was ok for some extra income and easy to do while working... Hessler said something interesting. He said that he had searched for a long term solution but the nets never came up with anything that worked more than a couple of weeks out... he said "think about it, can you predict what you will be doing in five years? What about five days?" I thought that was very interesting and it is in fact what caused me to focus on short term trading..
Seems like I read somewhere, that some pretty intensive research showed it was futile predicting the market more than about 3 days out...
Machine learning is really not that much different than the difficultly that arises from relying on the wetware between your ears; 1) finding the inputs that are useful is not trivial. 2) finding the proper way to process them and find a statistically meaningful and useful relationship between input and output is not trivial(even a neural net does not have 1 unique architecture). 3) finding useful outputs that work with the aforementioned is not trivial. The above three criteria are important to figure out regardless of how you choose to trade. Focusing on them is a lot more important than learning whether or not a neural net is effective. P.S. For starters, a common neural net is not inherently designed to deal with non-stationary signals. It is good to begin with a basic grasp of understanding how they actually work before commenting on their efficacy.
Amen. Most throw raw, non-smoothed data into a neural net, with poor correlation between inputs and output, with ten parameters and an inevitably insufficient sample size and then can't figure out why it doesn't work. Simple numerical correlation between inputs and outputs, as you pointed out, is critical. An excellent book for those who want to learn about NN input processing is "Neural, Novel and Hybrid Algorithms for Time Series Prediction" by Timothy Masters.
Is there any evidence of success with NNs to predict ancillary series; for example predicting whether an index or stock closes up/down in a day/week/month?
The problem though is this is not a problem you solve, publish and win a nobel prize...you solve then exploit what you found to make money and keep what you found close to the vest. Machine learning/stat pattern recognition/data mining is a well studied field of which NN and genetic algorithms are just one of many algorithms depending on what your trying to do. The reason you hear more about NN/GA at the retail level is because its possible to market software with names like that. People can grasp the concept of the algorithm even though without a background in machine learning its highly unlikely you will be able to do anything with it. Just like no one bothers making bayesian network retail software because your customer base would instantly know they haven't a clue what a bayesian network is. I tried to get into this but quickly realized the time I would need to put into to understand machine learning would simply be better spent training the vastly more powerfull NN between my ears than a simplified artificial one.