AIM
Brad Lewis, Portfolio Manager for Fidelity Investments, uses AIM to help Manage a 2.5 billion dollar portfolio of equity mutual fund assets. The fund is one of the most successful in the country.
Lewis attended an AIM training school, where "half the time was spent learning how to devise the correct inputs." Lewis now has four costing models, consisting of 54 individual companies each with 3 individual time horizons. "That makes 152 models in all," says Lewis. "I have macros written that step through all 152 models overnight without human interferance."
To set up the models, Lewis says he spent a lot of time making sure the input variables were correct. He found that because he was acting on 250 records, he had to restrict the number of inputs to between 8 to 10. After 13 inputs he began to see overtraining.
Lewis runs his 152 models on a monthly basis, then refers to them as needed, Each model gives him a forecast for expected returns in the form of a single output. "I'd rather build more models with single outputs than try for more complex models," he comments.
"AIM is very, very fast," he adds, Further, "AIM is a good place to start when you need a decent solution. There are some critical modules for which we still use some neural nets, but you have to check your input variables and your hypothesis for neural nets as well as for AIM. If it doesn't come out well in AIM, it won't come out well on a neural net."
ModelWare
Jeff Vogel, owner of Trading Corporation in New York City, uses TERANET's ModelWare for financial modeling and short-term forecasting. "The importing of data into ModelWare is very organized because it works from a spreadsheet," Vogel says. "You can get good results with smaller amounts of data. Also, the results are reproducible. ModelWare offers a suite of tools - statistics, drivers and correlations that are handy for financial modeling. The statistics features give me minimum, maximum and standard variations. I often use the normalization tool, which shows me a percentage of change.
"ModelWare runs quickly on a 486. It's a great program for control, sorting, looking at expected values and a range of expected values. I often use output from ModelWare and put it into another model. The tool has good online help. Also, the people at ModelWare have been very helpful. They spent a lot of time with me; I had access to the president."
BrainMaker
Dr. Steven Berkov works in the emergency room at Kaiser Hospital in Walnut Creek, California. He explains that at Kaiser, as in many hospitals, when people come to ER and do not appear to be in critical condition, they are signed in and then may wait over an hour to see a doctor who will order lab tests. The lab tests may take another hour and a half to two hours to arrive. By the time the patient receives treatment, he or she may have waited several hours or longer.
Berkov built a neural net to eliminate the first waiting period. The net gets accurate tests ordered without the patient seeing a doctor.
Berkov used a broad range of ER cases and built a training set in BrainMaker. "I tested eight different neural network models with different inputs and different architectures," he says. "With each run, I was able to see which inputs and outputs would be most useful. Initially there were hundreds of inputs, but I ended up with 67 inputs that proved to be the most important. There are 38 outputs which are specific lab tests.'
In the process of training the net, Dr. Berkov said that one of his biggest challenges was getting the net to recognize gender. "It took a long time to get the net to not order pregnancy tests on men, he notes. The completed model, however, got the facts straight. According to Berkov, it is now more than 90% accurate and closer to 100% accurate most of the time.
"The net tends to order less lab work than nurses and even doctors," says Dr. Berkov. "It has potential to be a big money-saver. From statistics I ran, our hospital could have saved half a million a year in lab tests if we were using this system, and this is just a first pass at building the net."
Dr. Berkov says that he found the network easy to build in BrainMaker. He has applied to Kaiser to build a more sophisticated version of the network and he hopes that the software will be implemented throughout the Kaiser Hospital system.