Silicon Investor (SI) -- The First Internet Community

STOCKTALK

We've detected that you're using an ad content blocking browser plug-in or feature. Ads provide a critical source of revenue to the continued operation of Silicon Investor. We ask that you disable ad blocking while on Silicon Investor in the best interests of our community. If you are not using an ad blocker but are still receiving this message, make sure your browser's tracking protection is set to the 'standard' level.

Strategies & Market Trends : Neural Nets - A tool for the 90's -- Ignore unavailable to you. Want to Upgrade?

To: F Robert Simms who wrote (273)	11/9/1998 8:23:00 AM
From: Patrick Slevin	Read Replies (1) \| Respond to of 871

So it appears the models are relying overmuch on trend, you would say? It seems if they were trained on a cross-section of the early 90's through late last year there would be no periods of significant bearishness that could be translated into a bear market. This is offhand, I haven't looked at any charts to refresh my memory but as I recall it was a straight up bull market for years. Either that, or your choice of variables/indicators is bogus and needs work, I guess. I had the same problem, I had some nifty looking systems in early August but when I returned from an extended holiday they had all broken down. Insofar as I was away and not trading them I had the advantage of stepping back at the time and theorizing that the change in market conditions was never in the models' training. I'm not quite certain how to handle it; more training, or less training. Going back somewhat farther gives it more history on bearish events but some of the market conditions have changed permanently; like the installation of chillout periods after 1987. (I never thought circuit breakers made valid sense but that's another topic). But on the other hand, shorter training gives the models more weight (I would think) to periods of extreme volatility such as has been the effect recently. The downside there is the danger of walk forward testing...curve-fitting. I suppose if this was easy everyone would do it.