I have been toying with neural networks to see about their functionality. The model that I have, which i did not build, seems absolutely ridiculous to me. What the model does is it trains on time series for several stocks. Then it makes its daily predictions for a stock but it uses the time series for the other stocks. In other words it makes its predictions for a single stock using the prices for stocks in the same industry for that day. So the prices for its competitors is already known. What good is that? That doesnt seem predictive at all. I dont have that information in advance. Is that how these models are suppose to work?

you can use a neural network however you want - its just a statistical estimator. It sounds like this is being used for pairs trading. you could also train a neural network to predict future price movement if you wanted to -( though I certainly don't think it would work)

QuoteOriginally posted by: spv205you can use a neural network however you want - its just a statistical estimator. It sounds like this is being used for pairs trading. you could also train a neural network to predict future price movement if you wanted to -( though I certainly don't think it would work)It is simply a model that tries to predict a daily stock price using the stock prices of competitors in the same industry. What it does though is say it wants to predict x1. It has time series for x1 and its four competitors x2, x3, x4, and x5.It trains the model for several weeks of data then it predicts a weeks closing price for x1. However it uses the closing price for x2, x3, x4, and x5 for the same week that it is predicting x1. If I know x2, x3, x4, and x5 in reality then I already know x1

Do you have any prior beliefs that their is some highly non-linear relationship(s) in the stock prices? Or some highly volatile behavior in the stock prices? Otherwise, isn't what you're describing a VAR model?The physical set up of the NN does not seem correct, as you point out, you could trick the NN by moving the stock prices that you aren't trying to forecast up a time period. Or am I missing something? (sometimes happens)

I don't really know anything about pairs trading!! butIn Pairs trading you try to identify when two stocks are out of line with each other and bet that they will converge in the future. You do not predict the future as such, you just try and identify a buy signal, using the current stock priceswith this neural network set up, aren't you doing the same: you are "estimating" what the stock price x1 should be based on historical relationships with x2,x3,x4,x5. Then if the estimate for x1 is higher than actual x1 you would buy x1 and short x2,x3,...ie you are betting that the x1 price will converge to the estimate. [maybe "predicting" is throwing you off]As you say, "If I know x2, x3, x4, and x5 in reality then I already know x1", so what if the estimate for x1 is different from the actual closing x1? isn't that a trading signal?

QuoteOriginally posted by: spv205I don't really know anything about pairs trading!! butIn Pairs trading you try to identify when two stocks are out of line with each other and bet that they will converge in the future. You do not predict the future as such, you just try and identify a buy signal, using the current stock priceswith this neural network set up, aren't you doing the same: you are "estimating" what the stock price x1 should be based on historical relationships with x2,x3,x4,x5. Then if the estimate for x1 is higher than actual x1 you would buy x1 and short x2,x3,...ie you are betting that the x1 price will converge to the estimate. [maybe "predicting" is throwing you off]As you say, "If I know x2, x3, x4, and x5 in reality then I already know x1", so what if the estimate for x1 is different from the actual closing x1? isn't that a trading signal?Point taken. I misread the post. The point is that at time t-1, I use signals from x2, x3, x4, and x5 to predict the true x1 price at time t-1 and make a determination about a time t trade.I thought we were trying to estimate x1 at time t.

QuoteOriginally posted by: BeachcomberQuoteOriginally posted by: spv205I don't really know anything about pairs trading!! butIn Pairs trading you try to identify when two stocks are out of line with each other and bet that they will converge in the future. You do not predict the future as such, you just try and identify a buy signal, using the current stock priceswith this neural network set up, aren't you doing the same: you are "estimating" what the stock price x1 should be based on historical relationships with x2,x3,x4,x5. Then if the estimate for x1 is higher than actual x1 you would buy x1 and short x2,x3,...ie you are betting that the x1 price will converge to the estimate. [maybe "predicting" is throwing you off]As you say, "If I know x2, x3, x4, and x5 in reality then I already know x1", so what if the estimate for x1 is different from the actual closing x1? isn't that a trading signal?Point taken. I misread the post. The point is that at time t-1, I use signals from x2, x3, x4, and x5 to predict the true x1 price at time t-1 and make a determination about a time t trade.I thought we were trying to estimate x1 at time t.That is exactly right Beachcomber.

QuoteOriginally posted by: rmb623I have been toying with neural networks to see about their functionality. The model that I have, which i did not build, seems absolutely ridiculous to me. What the model does is it trains on time series for several stocks. Then it makes its daily predictions for a stock but it uses the time series for the other stocks. In other words it makes its predictions for a single stock using the prices for stocks in the same industry for that day. So the prices for its competitors is already known. What good is that? That doesnt seem predictive at all. I dont have that information in advance. Is that how these models are suppose to work?you are using some scamming software? Seriously, neural networks are bad...but they aren't that bad . For instance, all the famous Backrpopagation algorithm does is to test a previous stock value with the next value and determine the difference between them - and yes...they can't predict a damn thing regardless of the algorithm being used...backprogation, rbf...etc. Just save your time and use it for something more entertaining at least...like...watching a movie.

What's the verdict on this please? Can NN be used to "predict" anything at all?

Actually, how about using NN to predict the direction of endogenous variable Y at time t based on the changes in exogenous variables X1, X2 and X3 between times t-2 and t-1? Definitely an exciting thought!