Jay,
If I understand the last paragraph correctly (below), it seems to be stating that we should use as many influence days as possible for the short term filter, and 1-4 influence days for any other indicator that we check, which seems to include any variation of long filter, price, or volume. At least that's my interpretation.
Since our discussion last night I started 2 additional nets, following the above advice regarding the influence settings, using 30 as the influence days for the short term filter, and checking only one box per related strock. You're right, it trains much faster and "cleaner". It could be that my Neuro 60 was running so slowly because I was asking it to process too much information per related stock (too many boxes checked). An example of what I'm talking about follows for LENS. I'm also going to post this to the site for reference and additional discussion.
LENS-s/t filter-30day (TARGET)
LENS-price-3 days
LENS-volume-3days
LENS-l/t filter-3days
WSX-volume-3 days
WSX-l/t filter-3 days
RUT-volume-3 days
RUT-l/t filter-3 days
COMP-volume-3 days
COMP-l/t filter-3 days
Influences:
Volume: Daily stock volume.
Price: Daily Closing Price.
Long Term Filter: 50 day moving average of closing stock price.
Short Term Filter: Short term price variance.
Influence Period: Number of days that NeuroStock will look back to compute trends and patterns for the following day prediction.
A good general setting for NS6S is 20 days for the short term price variance and 1 day for the other indicators.
You don't have to use all parameters if you don't want for a given stock or indicator. In general, stocks can use as many Short Filter days as possible and one to three for Long or Price. A few stocks benefit from the Volume information, while most other indicators do well with one to four days of Long Filter.
|
