Hi, I am applying classification problem (buy-sell-hold classes recognition) to Index futures prices. Trying to use classify buy-sell-hold signals with neural network. My trained neural network does not classify buy-sell-hold signals on the testing test with great accuracy. It predict with around 60% accuracy and whcih does not help me to derive a profitable strategy. Any idea how can I improve upon the classification part where buy-sell-hold signals gets recognised with even more accuracy on testing data.

You need to think more clearly about what you are trying to do with a buy/hold/sell strategy.You first of all need to specify what you want your network to do. For example it could be to optimize a function of reward and risk in terms of a set of trades. Then you need to specify outputs (e.g. a number between -1 and +1 that you can translate into trades) that enable you to train the network to optimize that function for historic data.

I am very clearly defined buy-sell-hold value as the target vector in neural network instead of -1, 0, 1. The performance varies with the inpur variables. After thinking so much, I still could not find what should be the input which give good classification over buy-sell-hold signls.

QuoteOriginally posted by: humtumiitI am very clearly defined buy-sell-hold value as the target vector in neural network instead of -1, 0, 1.Ok, but you still need to specify how you value that output in trading in order to train the NN. What do you want to optimize?

I am using 50 technical indicators as input and target vector is buy-sell-hold values. These buy-sell-hold values are generated through some method. I am trying to minimize the misclassification. My NN predict buy-sell-hold signals with 60-70 % accuracy on test data but I want to increase the accuracy even further.

In situation where we are going to classsify trend - non trend like sitations, in such situations does simple ANN works or do I need to work with recurrent neural network or auto associative neural network?

QuoteOriginally posted by: outrunI would advice to forget about the NN and use the much better mathematically founded Support Vector Machines. Also: instead of a target vector, it might be better to use reinforcement learning because that focuses on maximizing on overall reward without explicitly knowing the reward of all little steps in between. Another option is to condition the objective (of finding the best action) onto your current position. Instead of finding the best action at some point in time, find three best actions, one in case you are short, one when long and one when having no position. This makes ik suitable for Bellman-ish dynamic programming in a second phase. Your best action might also differ based on your current position due to transaction cost and risk/return optimization.Thanks foryour nice suggestions. Generally are we able to improve the prediction accuracy on test data by using SVM? If yes, up to what percentage of accuracy we can predict signals correctly.

Can we predict signals on test data by upto say 90% accuracy in SVM?

I am working in R and used svm function to train SVM but still results are not good. It predict only 50% signals correct on the testing data. I am using 50 technical indicators as input data set and buy-sell-hold signals as target vectors. Buy-sell-hold target vector has been created mannally. I used radial function and have tested with ranges of values of parameters gamma and cost but no success. I am not sure how to choose input to the SVM such that it predict precisly. What condition input data set should satisfy?

I compared SVM results with the ANN results. Results are similar or in some case, ANN results are bit better. If we can not predict buy-sell-hold signals with good accuracy then how can we develop a profitable trading strategy. These 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?