Len,Jay,
I sure hope you want some company on this thread. I've been using NS for about a year now, and generally agree that it's a great product...I'm up to NS80...but it sounds like you guys have some good insights regarding its use.
From the sound of things, you guys know each other and have shared quite a bit of info already privately, but I'm assuming you started this discussion to have more ideas/inputs/...
Assuming this is the case, but understanding I may duplicate things you have already covered, I'll start the discussion with some of my own observations.
1. Just because you have a good model does not GUARANTEE profitability. However, it sure makes it more likely. One of my questions is how to construct models that have good confidence AND work. For example, I currently have an SPX model that has had 70-80% confidence on a BUY signal all through this market turmoil. On the other hand, I have another individual stock model that only gets to 40-50% confidence, but I can count on it's productivity. Any comments?
2. What do you guys think about have a training vs. evaluation period. I always use a verify period of from 60-90 days, but the possible cost is that the net is not using the most recent data, so if the dynamics change, it might not pick it up quickly. (Having said that, having a good diagonal on the evaluation data as well as on the training data sure does help MY confidence!)
3. I found your discussion of using indices for industry groups in lieu of other relateds to be an interesting alternative. I had thought about this but never tried it. Would you care to elaborate on how you handle this approach? I.e., do you use only a single related index or do you have several versions of it in the model?
4. One thing that I have not had time to test thoroughly is the choice of influence periods. As you correctly pointed out, the manual is pretty sparse on this issue. So, I always do the same thing...
Short term filter, 20 days, issue of interest
Long term filter, 1 day, price, issue of interest
Long term filter, 1 or 3 day, price, related issues
I think I lifted this from one of the manual pages and have not explored things like volume or having more than one short or more than one long for a particular issue. Have either of you? What are your strategies for these choices?
5. How many days of data have you found to be necessary? I've been using about 3 years, as recommended. But could use more, if necessary. What have you found?
Well, I have many more questions for discussion, but I guess I should see if you respond to this much before continuing. However, for interest, here is the report for the VERIFY period for DAL... This model has about 60% confidence right now, with 91 hours on an 80 node net... Normal strategy
19980527 buy 47.1285 114.312 5387.35
19980610 prof tk 47.1285 126.938 5982.4
19980623 buy 48.3184 123.812 5982.4
19980714 sell 48.3184 142.188 6870.29
19980805 buy 58.4085 117.625 6870.29
19980818 sell 58.4085 124.938 7297.41
19980827 buy 71.5432 102 7297.41
Current Price: 108 7/16
Sorry for the long-ish posting, but I've been building questions/ideas for a year!
Dave |
