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: Carl R. who wrote (146)	6/29/1998 11:33:00 AM
From: Optim	Read Replies (1) \| Respond to of 871

Woah, wait a minute. With Neurostock Len described having to train new nets for a minimum of 100 hours. You talk about 30 hours, which is a big improvement. But Brainmaker talks about being 10-100 times faster than the competition and training networks in minutes: First let me clarify what I said earlier. In 30 hours NGO bred a pool of 50 networks 5 times (or generations). That means 5 sets of 50 networks, or 250 networks in total. It works out to about 7.2 minutes each net. It then finds which of these works best, and ranks them. Then the top ten candidates are passed to the Profit system, where you can evaluate them, look at their out of sample charts, optimize them, etc. Plus they have their CNaps board which is really fast if you've got the big bucks. They also claim to be incredibly easy to use, which obviously NGO is not. The accelerator card is probably overkill for most users. So far I have found that NGO will breed a reasonable solution in 3-5 generations. This usually means about a day's worth of number crunching on my 233 Pentium II system. NGO is a hard program to learn, mainly because it is overwhelming at first. If you use the Profit front end, a default selection of parameters is used, which you can change later on as you learn more about the program and NN's in general. All packages are difficult, but it is a tradeoff between ease of use and functionality. Neurostock is a really easy program to learn. But it lacks the flexibility that would help make it a more robust system (customizing data transforms, changeing predicted output variable, etc.) Some users may not use all the features in MS Word, but when they need them they are happy that they are there. I gather that you like NGO/profit now that you are used to it. Would you recommend it? Having used Neurostock, Neuroshell Trader, Profit/NGO, and demoing a number of other packages, I would recommend each according to the users skill set. I started with Neurostock last year around this time. Within a month or two though I became frustrated with the whole 'black-box' approach to NS, as it is hard to help the network find good data when you don't know what it is looking for. I progressed to the Genetic version of Neurostock in the hopes of having the network find which inputs where best. Around that time Ward Data Systems was promoting their Neuroshell Trader package, which is basically a neural net built around a full featured TA package. I thought that this would be the next logical step for me, as it would shield me from the complexities of tuning neural parameters (# neurons, momentum, learning rates, etc.), and still allow me to create my own trading strategies and target outputs. I can't stress that enough. You need to customize your system to your style of trading. If you are a daytrader you could predict the difference between tomorrows open and high. Or if you are a longer term investor, you could predict a moving average 10 days out. Anyways, NeuroShell Trader allowed me to do this, but I found it very difficult finding good, predictive inputs. I did manage to come up with some very good nets (>100% a year, out of sample), but it was a tedious and time consuming process. So now I have graduated to Profit/NGO combination. The copy I purchased allows you to feed it 64 inputs and have it sift through them to choose which are most valuable for a given target output. This saves me a lot of time over the trial and error approach I use with NS Trader. I am not saying that Profit/NGO is the best system out there, as I am sure there are better ones. But for my level of knowledge, and my current needs (minimal amount of development time) NGO has been a good match. I sure that in time I will push the limits of Profit/NGO also, although with the flexibility built in to NGO it will take me a while to get up to speed. I just ordered some additional indicators from Jurik Research to help process my data before it is fed into any of the above packages. Until you try these packages, it is hard to say which will meet your needs. It is an expensive learning experience, but my hope is that one day this will all pay off. So in short, in my opinion Profit/NGO is a great package but it is up to you to find a packages that suits your needs. Hope that helped... :) Optim