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 (157)	7/2/1998 11:56:00 AM
From: Optim	Read Replies (3) \| Respond to of 871

As I said before I am a NN newbie Maybe I can help you out here. I'm no expert, but I have been at it long enough to pick up some of the lingo... :) There are a few main points to keep in mind when designing a neural net: 1) Know in advance what you want to predict and make sure it sufficiently profitable. In Profit you can check this in the data transformation screen by choosing the Profitability option. It will run your predicted variable though the data and come up with a dollar figure. Mess around with variable until you have one that is profitable and smooth if possible (ie graph it in Profit). A smooth output is usually easier to predict. I have used a 1 day percent change, smoothed with a 3 day Adaptive moving average, predicted 3 days into the future with good success. You can also forward predict technical analysis indicators. 2) Try and tranform all your input data so that it appears to oscillate, or flucuate up and down. For example use the % change in price instead of raw price. This helps the network pick up on relationships in the data and not learn that todays price is close to yesterdays. Good inputs for me include MESA indicators from Metastock, Momentum-type, Stochastic %K, Market Breadth (Advance,Decline,NewHighs,NewLows), and any related fundamental issues (ie for techs, tech indexes, currency rates, etc). Try to find something that leads a change in your issue. For example Bonds and the S&P500 are highly correlated. If bond yields go down, stocks go up. Feed this type of info to the networks. 3) Try not to train on too much data. This is the opposite from what most traders believe in. Most test a system on 10 years or more of data. But non-linear relationships in the data 10 years ago probably don't apply tody. I generally get my best results with less than 2 years of data. Just be careful as the life on these models trained on less data is shorter than one trained on lots of data. My models trained on 1-2 years of data are generally profitable for 4-6 months before they need to be retrained. 4) Try not to feed the network too many inputs. I try and keep the number of inputs down to less than 8, with the optimum being around 5. This limits the 'degrees of freedom' that the model has. Put simply, it forces the network to learn only valid relationships, and not just chance patterns or noise. A really great tool for this type of thing is DDR from Jurik Research (http://www.jurikres.com). It takes a number of inputs and combines them into a smaller set of inputs, while losing as little data as possible. The result is a neural network that has less error, and is more likely to continue working in the future. You can read more about it at the website. 5) Keep at it. A lot of people get discouraged with these tools initially. It is definitely overwhelming. I knew nothing about neural networks, or trading for that matter, less than 2 years ago. If you persist, NN's are a great tool and will pay for themselves many times over. Very few keep at it long enough to reap the rewards. If you have any questions, post them to the thread. Though an active message base we can all learn how to use these tools more effectively, and share in the successes and failures of others. Optim