aiQUANT

The problem with moving averages

Filter Design aiquant September 22nd. 2007, 1:30pm

In the brownian path forcasting example I highlighted the danger of using a moving average vector as an input to a neural network. The values are abviously time lagged, and patterns appear in delay. The neural network inevitaby learns this lag and lag is a bad thing! I wish to develop a filter that has almost zero lag, whilst providing the same smoothing capabilities of a SMA or an EMA regardless of the filter length chosen. In the next couple of posts I shall describe this using Z-transforms.

Bookmark & Share

2 Responses to “The problem with moving averages”

1
Lumilog Says:
October 9th, 2007 at 11:20 pm

Hi aiQUANT - I’m really enjoying your articles.

I understand - intuitively - what you’re saying about the problem with time lag in inputs to a neural network. However, it would seem to me that with supervised training, you’re still training the network in how to best use this lagged information (e.g. a 50-day SMA) to predict a non-lagged output (e.g. the price move tomorrow). Right?

Thanks,
Lumilog
2
aiQUANT Says:
October 10th, 2007 at 8:37 am

Hi Lumilog,

I see where you are coming from. Consider this: We want to train a NN to forecast a sine wave using inputs that are lagged by 0 bars and 5 bars. Which input will the network attach more weight to (i.e. which input will be the networks favourite?). Lets say we now have 5 bar and 10 bar lags as inputs. Would the networks favourite input have changed?

I shall post an example of this showing how the network weights change subject to different lags in the input.

Thanks for bringing this up!
Piyush