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: LastShadow who wrote (746)	4/6/2000 9:43:00 AM
From: Optim	Read Replies (1) \| Respond to of 871

Hi Lastshadow, Just wanted to say that I can already see that our methods of using neural networks are very different. I am gaining a new appreciation for your methods, as it has never occurred to me to build a network for general sector use that can be tailored for individual issues with some adjustment. I did model 3 stocks as a custom index once, but perhaps I didn't pursue it far enough. I can see now why price shocks may not be as much an issue for you as a basket of issue will be less susceptible to them. Where you really surprised me is with the 2-3 month training period. Is this enough data for the networks to stabilize? I have been using a minimum of 18 months data with only 2-3 inputs or else I overfit the training period and kill my out of sample results. Our approach is probably different because the package I do most of my work in, BioComp Profit, uses a GA to build it's networks. It is very difficult to reproduce results as the GA takes a new path (inputs, structure, etc.) every time it is started. Profit is also very limited in building trading strategies. I've often imported Profit's neural signal to Metastock to build better trading strategies. For stops I've found the Kase DevStop method to work very well as it adjusts for recent volatility. This is one of the areas a trading strategy can make up for deficiencies in the neural network. I agree that modeling for consistency is better that modeling for the occasional 'pops'. I hope you don't regret offering your advice on building a neural network. I know you're a busy man, so if I can help in any way let me know. Optim

To: LastShadow who wrote (746)	4/6/2000 7:10:00 PM
From: xu, b.	Read Replies (2) \| Respond to of 871

If the object is to show how to create a generic net and then finesse it to the specific secotr or stocks, then a more robust example would be to use a dozen stocks with one from eah major industry. The performance of the net will then tell you if your initial characteristic set and pattern discrimination method is better suited to one or another. Good point. Let's keep this objective. First thing first, let's agree on stocks. I propose to pick 3 from semi and 3 from bio. To make it generic, Could we pick one for large cap, one for mid-cap and one for small cap each. Here is the matrix looks like: Semi Bio Large 1 1 Mid 1 1 Small 1 1 I don't care which particular stock. But let's agree that we want to compare with sector indices AND/or SPX. We should have some funs with lag/lead among the same group as Dan suggested. Secondly, output? Do we use Ruggerio's suggestion in his book? Or you guys prefer another one. I leave that to you since you guys are much more experienced than I am with the nets. Thirdly, inputs? SPX, SOX, XCI, (maybe a Taiwan semi index as they have the largest foundries? I have no clue where to get it easily), OBV, AMA, SD or ATR or (H-L)/C. I leave transforms to you two Last and Dan if you would agree. add some? take some away? Forth, Net structure. I only have access to BP. Let me know which one gives best results and how to compare the results. I can help evaluate results. I have no problem with frequent retraining. What else missing?