In my experience, when the neural network ends up with all zeros at the output, it is likely that the rails are being saturated. That is that the common sigmoid function processing element will only swing to 0 or +1 limits and if the input happens to be far outside of the linear input sensitivity...
Thank you for the answer. All of my inputs are scaled to the range -1 to 1 you mentioned since I'm using the bipolar sigmod which ranges from -1 to 1 ... if this was the case, the neural network would stick to either -1 or 1 but not 0. My inputs are scaled as are the outputs.
I think you mean the tanh function. Could you show a table of some of the training data? Also, what happens if you use the default output (not slope), is the data non-zero? I'm not familiar with the Sierra Charts tool, so there may be something tool specific I might miss.
I was thinking that my problem is maybe probably even more amateurish...I'm using the slope of the linear regression curve as the output which is just a value oscilating around zero as the price rise or goes down. I thought that might have something to do with my issue.
And yes, the tanh function. The Sierra Chart is just a charting software, it does have nothing to do with the NN, I probably shouldn't have mentioned the program at all.
Unfortunately, It doesn't help the troubleshooting much with only raw data. It would be more useful to post the transformed training data (with input and target variables specified).
Not price, but return as i mentioned. Price is explosive and not a good variable to use for NNs. You want to always train any type of learner with stationary data where possible. You can always transform back to price after. I have a feeling we are cluttering up the thread with conversation; any suggestions on how to limit that would be appreciated.
I'm realizing the format is a bit weird though...even lines are inputs, odd lines are the output. Would you rather the comma separated table with the output as the last column? I can do that. I'm really desperate for any help.
no, that IS one target variable. It is the linear regression slope. #2 line is the first set of inputs and #3 line is the corresponding computed slope. #4 line is another set of inputs and #5 line is the corresponding computed slope etc.
I'm confused. The table format should be input D={x1,x2,...xj} where x is your input variable columns and Y=Yt the target variable. Forget slope, just use training data that makes sense. Can you get it in this format?
Also, your .dat output has data on seperate columns making it hard to process. It would be nice to have each row correpsond to each input/output observation.
And maybe you could limit it to just 500 observations. It's taking a while to download.
ah.. i just realized why i'm having so many problems. The data has commas instead of decimals! I need to preprocess.
nput18 Input19 Output0 Min. :-1.0000 Min. :-1.0000 Min. :-0.4943881 1st Qu.:-0.3515 1st Qu.:-0.3515 1st Qu.:-0.0344828 Median :-0.2361 Median :-0.2361 Median : 0.0009466 Mean :-0.2410 Mean :-0.2410 Mean : 0.0007731 3rd Qu.:-0.1319 3rd Qu.:-0.1319 3rd Qu.: 0.0355646 Max. : 1.0000 Max. : 1.0000 Max. : 1.0000000
ok input look ok output is a llittle skewed (min,max = -.49 to 1)
Discussion between Motig and pat
Discussion between Motig and pat