Serving the Quantitative Finance Community

 
User avatar
OOglesby
Posts: 1
Joined: August 26th, 2011, 5:34 am

Buy-Sell-Hold classification with Neural network

March 2nd, 2012, 6:35 pm

Classifiers such as neural networks or support vector machines are not magic. I feel they both require the user to be very knowledgeable about the problem domain.QuoteI am using 50 technical indicators as input and target vector is buy-sell-hold values.I would say that many of your inputs do not relate to the behavior of the system (in this case index futures prices).I use SVMs at work for engineering problems, and I routinely get > 95% accuracy with the training data and > 85% accuracy in the real world. These classifiers only use 3 inputs, and these inputs directly relate to the behavior that I am trying to predict.
 
User avatar
humtumiit
Topic Author
Posts: 0
Joined: April 28th, 2007, 6:20 pm

Buy-Sell-Hold classification with Neural network

March 2nd, 2012, 6:44 pm

Relation of signals i.e. output and input data is not straightforward. If it would have been so easy there was no need to for ANN or SVM we could have simple relate the price input with the output i.e. signals and generated buy-sell signals with a smal algorithm. Getting input and output relationship is the reason which force me to look for such highly technical algorithms.
 
User avatar
iamquant
Posts: 0
Joined: September 26th, 2011, 5:39 pm

Buy-Sell-Hold classification with Neural network

March 7th, 2012, 9:36 pm

err why can't you make money with a 60% prediction accuracy? trading costs eat up your profit margin? not enough capital?
 
User avatar
Stale
Posts: 0
Joined: November 7th, 2006, 3:20 pm

Buy-Sell-Hold classification with Neural network

March 7th, 2012, 9:44 pm

Even though you can predict up or down, can't you think of a distribution of returns that will make your +1 or -1 useless in the market?
 
User avatar
spv205
Posts: 1
Joined: July 14th, 2002, 3:00 am

Buy-Sell-Hold classification with Neural network

March 7th, 2012, 11:07 pm

QuoteThese days almost every high frequency trading system use AI to generate signals, if the signals are not predicted precisly then how come they make profits. The claim that their system prediction accuracy is 80-90%, how do they claim?( where do you get your information from? I doubt it - the more high frequency analysis the more likely it is to use the simplest poss algos)AI is typically used as a sales "buzzword", rather than being essential to businessAs OOglesby politely said. - they are not magic. they require the user to be v knowledgeable about the problem domain.so the real point is that you have someone who knows a lot about the market then fine tuning an existing buy signal with AI...so think about what you are doing - you have a 50 dimensional statespace - lets say for the sake of argument that each variable has only 2 values - you need 2^50 =1e+15 data samples to sample each point in the state space onceIf you want you svm to work - use only a few variables...
 
User avatar
yuryr
Posts: 0
Joined: November 5th, 2007, 12:47 pm

Buy-Sell-Hold classification with Neural network

March 13th, 2012, 2:02 pm

I will never understand why one would use proper ML techniques like NN, SVM, etc. and then plug in technical analysis input.If you want your tool to dig out some signal, why obscure the signal first? or is it a part of some sort of complex genetic algorithm with predator/prey/survival of fittest?On the serious note, SVM is convex optimization tool - works well on many problems, but not all problems are convex - hence, need for ANN and even more complex things.Anyway, myy 5cents: 1) plug in raw signal, 2) tune your algo of choice in different ways (optimize cost matrix, find optimal transformation of data etc..), 3) do not use reinforcement learning for this, can resort to some simple planning problems if needed, but chances are - you don't need it (optimizing cost matrix may be sufficient) - reason is: you are dealing with a model that you can afford to fully train offline.
Last edited by yuryr on March 12th, 2012, 11:00 pm, edited 1 time in total.
 
User avatar
yuryr
Posts: 0
Joined: November 5th, 2007, 12:47 pm

Buy-Sell-Hold classification with Neural network

March 19th, 2012, 1:31 pm

that is right, but I was referring to the convexity of the optimization task that you eventually perform with SVMsIf the two sets to be separated in the original space are convex or not is irrelevant.What I meant is that not every problem can be turned into convex optimization task. Perhaps, theoretically with some regularity conditions there may exist a mapping into a convex set from any set, but in reality for a given task you may not be able to find out this mapping, because finding this mapping is equivalent to searching for structure in the data and that is the original task in machine learning, i.e. funding the structure... And in this sense non-convex methods may potentially be more powerful than SVM, even so SVMs work very well on many real world problems.