Attrasoft Stock Prediction

Q & A


 
 

 Q:When attempting to forecast a result, the obvious training method is to select
the input variables and provide the known result and check the final
outcome...easy enough.  However, some inputs may be influenced by the actual
result, so should these inputs be calculated up to one period less than the
current event and then super-imposed as if it is the current event's input? or
does this not really matter.  For example, if an input is based on some
regressional analysis over x periods using historical results, then
technically the last input maybe skewed sufficiently to affect the final
outcome in a less desireable way at the time of forecast because the actual
result is yet unknown and therefore the input is a guestimate at best.  But
prior to the current event actual results are known and so can be input
calculations can be performed accurately in this case.  But my gut feeling is
that training on these inputs maybe less effective because it does not reflect
the true situation.

In this situation I have considered using an estimate of the input, ie based
on calculations up to the previous event etc., and a final input based on
actual results and then applying a Kalman filter to refine the estimate over
time or do you feel that this is a waste of time and I am off the beam?

A: Your method is fine. You might want to predict that value first, then use the predicted value for other predictions.
 

Q: ABM appears to work instantly! is it doing back prop or some clever statistical manipulation? It is so fast that I could use it in real-time analysis.

A: Attrasoft ABM 2.3 and Attrasoft Predictor 2.4 aredesigned to operate in the range between 1,000 neurons and 100,000 neurons. Already, people are asking questions about the neural net of 1,000,000 neurons. For a net of this size, speed is everything. Yes, it is doing back prop and it does not use any statistical manipulation.

Q: Another question i have is regarding the 2 variables N-day trend and Quantization. I dont completely understand your explanation in the help pages. Is it better to have as large a quantisation factor and n-day trend as possible to get the best prediction? I am not a neural net or statistical expert and your explanations are not completely "Expert free" here. All I want to know is what to do to get the best results.

A: Yes. Choose a larger "N-day trend" first, and a larger "precision/quantization" second.

Q: Also the help text implies that more variables are not necessarily going to help the prediction. I thought that the more related information a net has the better. Thank you for any help you can give me.

A: In general, more variables are better. However, it will also increase the complexity of the prediction. A certain amount of data only supports a certain level of prediction. If you do not have enough data to support your prediction, reduce the number of variables can help. Assume you have a very large amount of data and a very large neural net, then the more variables, the better.
  


Home
 webmaster@attrasoft.com