Plain vanilla Neural Nets aren't a beat-all solution in finance.They are good at supervised learning. That means you have to give it trainng data with training examples with 1) a feature vector 2) a reward signal associated with each example.e.g. You can't tell a neural net "Please magically predict this stock's movement for the next 15 minutes and make me money". You can't even tell it "here's this stock's price history...will it go up or down?"Here's what you can do, if you are trying to predict what a single stock will do1) Collect a long price history of 500 stocks2) Hand-pick a feature set. For example, you could run a neural net on three features a) this stock closed yesterday above its close two days ago (1 if true, zero if false) b) the ratio of (yesterday's close / 50-day moving average) c) The ratio (yesterday's volume / 10 day moving average volume)3) Pick a reinforcement signal R e.g. 1 if this stock went up more than 0.5% that day, 0 if it went between -0.5% and 0.5%, and -1 if it went down more than 0.5%4) Convert your raw training data into features. For exampleJan 10, 2002: a=1,b=0.9,c=1.2, R=1Jan 11, 2002: a=0,b=0.8,c=1.2, R=-1...5) Train the neural net to learn a relation between the features and the data6) Try and use this to predict if a stock goes up given NEW data.Neural nets can probably do this well IF the features describe the data...that is, if the features actually have predictive power, the neural net should learn the relationship between features and reinforcement signal.But you can see several problems...what should the features be? what should the reinforcement signal be (1-day profit, 5-day profit, 10-minute profit)? do the features you picked actually have future predictive power, as opposed to just explanatory power? I plan to do research in this area...please email me if you would like to know more (cdmurray@cs.cmu.edu)

Where and when ml? Most ml people might say: PROVIDED there is a "natural, econimic, social... law" (behind your i/o relations) AND your data give you a "good" representation of such a law THEN there are ml methods that can create an approximative, predictive (hopefully computaional) model.To be added: AND if you have all relevant input data available exactly at the time you need to make yor decision (based on the prediction)If we destinguish: a) "what we know" b) "what we do not know"To get from b) to a) we might (with the 500 stock price example of cdmurray80)A) put any economic, social,.. parameters available press them into a table of i/o vectors apply any supervised ml method(-combination) and get a predictive model and (if the model is a white box) info which input parameters are "important" in a certain domain and which are only "pink noise".B) look into the history of "simple" i/o relations (often input is only time) and create features as input vector (in technical applications we often say, we search for a "fingerprint" of a material, machine behaviour, ..), as cdmurray80 introduced........My experience, for A) approaches NNs are not ideal. For B) yesAn additional difficulty in B) If you create "too many" features you might get "overfitting" and predictions become "random" (again).

QuoteOriginally posted by: DaGuyA friend argues that neural nets are just the greatest tool and that most financial researchers deprive themselves of using this panacea that "can solve any nonlinear problem". That's a bunch of huey, but I need to substantiate my argument better. Hence, I ask. Has there been any important papers in the finance literature which were based on any kind of neural net? I don't know of any? Do you? Which ones?While it is true that neural networks (at least the most commonly implemented variety) can theoretically approximate any nonlinear function, that does not imply that they can "solve any nonlinear problem". Leaving aside the data requirements that said theory implies, a single, nonlinear function may or may not be a complete solution to a given problem.-Will Dwinnellhttp://will.dwinnell.com

QuoteOriginally posted by: rmax<Anecdotal tale> The US decided use a Neural Net to try and identify Soviet Hardware. They trained the network using a set of photos depicting US hardware, and then trained it using a set of photos of Soviet hardware. When they then tried to use it, the network didn't work. Going back they tried to figure out what went on. It turned out that all the photos of US hardware were taken on a sunny day where as the photos of Soviet hardware were taken on a overcast day. The network they created could tell the difference between a sunny picture and an overcast picture.</Anecdotal tale>My objection to that old story is less about the fact that no-one is ever able to substantiate it, and more about the notion that it is typically trotted out as a criticism of neural networks, while the lesson it offers is not specific to neural networks. Any tool used incorrectly may give bad results. Any modeling technique (linear regression, logistic regression, CART, astrology, etc.), or indeed any analysis technique (PERT, linear programming, basic graphing, etc.) might be similarly characterized. The point is that, used properly, there is nothing about neural networks that should make one any more suspicious of their results than for any other technique.-Will Dwinnellhttp://will.dwinnell.com

QuoteOriginally posted by: exneratunriskIsn't "entropy ranking" used to create ID3 decision trees?Entropy is used in many machine learning methods, including as you say, ID3, to assess the quality of a candidate solution. Roughly, the idea is to minimize the conditional entropy of the target variable by optimize the splitting of the data via the predictor variables.-Will Dwinnellhttp://will.dwinnell.com

cdmurray80Do you know of paper which looks specifically at neural networks in finance?