Len,
Another for what its worth comment on annealing--I've had a net I've been training for about 30 hours or so that was totally missing an almost 50% upswing that occurred over about the last 4-5 weeks. This was the latest time period that led into my verify period of about a month. Anyway, Sells were being recommended all the way up to the top of the move from about $14 to about $21, with holds being indicated right at the top for the last week and a half or so. I have now let it run using only annealing for an additional 6 hours or so at the present time with the following results: all of the sell recommendations are gone during the upswing, with a buy recommendation being near the start, a couple of sporadic buy signals along the way, and a sell signal just prior to a series of holds for a few days near the top. The hold point prices, I believe are at or near a new 52 week high which I don't think the net will have ever seen from the training data. It is now giving a sell signal that I tend to agree with.
Unfortunately, I sold about 15% too soon of the recent plateau, based on another net I'd trained with different indicators. I'm going to switch to back prop and annealing and let it run itself for another 20 or 30 hours and see if the verify period results will become a better diagonal.
I guess the point is, that for this particular stock, annealing has APPEARED, for now anyway, to have helped tweek this into a more accurate and profitable network - at least if I'd had this one a few days ago, I would have cleared about 15% more when I sold.
Most of the time, as you have posted previously, I don't see much difference before or after annealing runs itself through a couple of times - maybe cases like this are where it's worth the time. Who knows?
Have a good Thanksgiving.
Bill |
