I. Drug Side Effects Prediction Wins Science Fair
One young scientist (age 15!) used neural network models to predict the side effects of new drugs based upon the known side effects of existing drugs. Frederick High School freshman Jennifer Gee won first place in the Frederick County Maryland Science Fair, Computer Science Division, with her project. Jennifer coded the characteristics of 60 known and widely used drugs as inputs to her models. She then applied her models to 20 new drugs. She built a classification model for each side effect (e.g. headaches) using the NeuroShell Classifier. Since Jennifer had a small training set, she needed the “one-hold-out” feature of the genetic method in the NeuroShell Classifier to be able to use all of her 60 drugs for both training and evaluation. Jennifer averaged about 88% accuracy in her models.
II. Air Quality Forecasting
Dr. Charles Piety, a professor in the University of Maryland Meteorology Department, has been using the NeuroShell Predictor and the NeuroShell Run-Time Server to create neural networks for air quality forecasting in the Baltimore-Washington and Philadelphia areas. Dr. Piety has discovered that the nonlinear relationship between meteorology and pollution is better captured by the use of neural nets rather than the use of simple linear models, including linear regression models. Dr. Piety says that the best feature of his WSG software is that it is “easy to install, quick to run, and user friendly.” He is also extremely pleased with the technical support he received from our support team. We’re happy to hear that!
III. Classification of Mild Brain Injury
We recently received an email from Dr. Timothy Belliveau, a specialist in neuropsychology at the Hospital for Special Care in New Britain, CT., about a project he and two other collaborators are exploring. Dr. Dennis Johnston, also in neuropsychology, and Jeff Axt, information systems, are both working with Dr. Belliveau on this enterprise. They have started extensive research using the NeuroShell Classifier for diagnostic classification of mild brain injury. They have based the model on a training set of cases with historical data, neuropsychological test results, and clinical diagnosis.
According to Dr. Belliveau, the preliminary results have been very encouraging. The group expects to submit an article for publication later in the year. After that, they plan to move toward a second phase of cross-validation with two other data collection sites.
IV. Yale MBA to Start Hedge Fund Using NS DayTrader Pro
Eric Hoyle, CFA, is a Duke University graduate and a Yale MBA who remembers his Yale professors “sneering” at the very mention of technical analysis. (We can just imagine their faces if someone mentioned neural networks!) He also remembers a paper he wrote at Yale in which he argued that, on a risk-adjusted basis, the returns achieved by hedge fund managers were not worth the costs. However, now Eric has done so well with the models he has created with the NeuroShell DayTrader Professional, he is starting a hedge fund using it as one of his major tools.
Eric’s change in attitude was brought about by his experience after business school. Eric explains that “in business school we were taught that everyone used fundamentally the same tools and had the same information, therefore, no analyst could consistently outperform. However, I see the NeuroShell Trader as a new tool. It helps to model relationships traditional analysts would miss, or not discover until it was too late to trade on them.” He feels, as a hedge fund manager, he adds value because “the key is selecting the right inputs, which requires an understanding of the markets. The number of indicators and relationships a trader could attempt to model are limitless; it is a full time job. I hope to provide an opportunity for investors to take advantage of my knowledge of the markets and the power of this program.”
Prior to starting the hedge fund, Eric worked as a fixed income analyst for the mutual fund subsidiary of Raymond James Financial, and before that, on the trading desk of Ferris, Baker Watts. It is very clear that Eric knows what he is doing when it comes to trading.
Before purchasing the NeuroShell Trader Professional, he experimented with a competitor’s neural network program and found it interesting, but frustrating to use. “The amount of time required to create a basic model was formidable. Creating a complex trading strategy would have been out of the question.” Later, when he saw an ad for the debut of the NeuroShell Trader program, he decided to try it. Contrary to his prior experience, he found that the NeuroShell Trader was “user friendly and had the ability to easily create complex trading strategies.” His results, he says, “have been phenomenal. I’ve done very well.”
Like so many of the successful NeuroShell users we talk to, Eric says he does not use rocket science in what he puts into the Trader. He builds trading strategies and optimizes them to maximize return. Then he “eye balls” the out of sample trading to be sure he wants to go forward and trade the system, comparing it to a buy and hold strategy. “It’s important to feel like the system has learned some relationships between the variables I’ve selected and not just made a few great trades.”
Eric models U.S. Equities and tries to diversify his trading across markets and industries as a risk control measure. Typically, his models employ just one neural net in the trading strategy. The inputs to his nets are simple technical indicators from our price momentum category, often applied on an inter-market basis – to other issues and indices. He does not have a favorite or consistently good indicator. He has found that simple is better: simple indicators, and no more than 5 or 6 of them to assure generalization. When Eric optimizes, he just uses the parameter search function.
We asked Eric about how much data he uses for modeling. “In the beginning, I was probably using too much data for my models,” he explained. “I wanted lots of out of sample results to be confident that my models would work well going forward. However, market relationships are dynamic, and as the relationships change your models need to change. I found that models that tested well over a long period actually had less chance of continuing to be profitable in the future. My best results were achieved with nets using roughly 9 months of data, and perhaps 2 or 3 walk forward tests of just 1 to 3 months each.”
“That was BEFORE the NeuroShell DayTrader,” he went on. “The DayTrader is great because it allows you to use a lot more data over shorter time periods. Shorter periods are particularly advantageous when the markets are rapidly changing. I now train my nets on only a month or two of data, using 10-minute bar increments. I may not technically be ‘day trading,’ however, day trading is not necessarily my goal. My trades typically last two or three days and I will have exposure to both the long side and the short side of the market. I am careful not to be too heavy on either side and I diversify as much as possible. I’m not concerned about being absolutely market neutral.”
We asked Eric if he had any advice for other NeuroShell users. “In addition to keeping the number of inputs to your nets low,” he suggested, “use a small number of hidden neurons in your neural network models. This helped me create much better trading systems. At first, I felt I was making my system less robust, however, I think using less neurons forces the model to generalize better. There is a lot of noise in the market, and you do not want your system to trade based on noise. Let your system learn the basic relationships and trade on them.” (Editors note: adjusting the number of hidden neurons is a feature only found in the NeuroShell Trader Professional and NeuroShell DayTrader Professional.)
At the time of this interview, Eric’s company had yet to be named, but we will add that information when we know it to the Real Traders section of www.neuroshell.com, where this article is to appear. We do know that it will be based in Eric’s hometown of Philadelphia.
V. NeuroShell DayTrader Not Just for DayTraders
Note this comment in the article above about Eric Hoyle:
“The DayTrader is great because it allows you to use a lot more data over shorter time periods. Shorter periods are particularly advantageous when the markets are rapidly changing. I now train my nets on only a month or two of data, using 10-minute bar increments. I may not technically be ‘day trading,’ however, day trading is not necessarily my goal. My trades typically last two or three days …”
That comment made us realize that we perhaps have been placing too much emphasis on the NeuroShell DayTrader Pro being for daytraders!
Eric’s comment notes that with the NeuroShell DayTrader Pro, it is possible to get a lot more data in a shorter period of time. You’d like more data for more robust models, but you’d like not to model ancient history. The NS DayTrader solves that problem.
There is another reason why end-of-day position traders may want to use it too. Many trading systems rely on intraday information, even if the practitioners don’t consider themselves daytraders.
For example, many systems don’t depend upon orders placed at the end of the day for market order or even limit order fills the next day. The order might be placed after evaluating early trading. It might be placed near what looks like a low point of the day. Some schemes may look back during the same day to decide how to set stops, or how to exit. Some depend upon knowledge of the day’s open, or the low or high so far for the day. All of those things are difficult or impossible without the NeuroShell DayTrader.
VI. What are Open, High, Low, and Close?
So now that we’ve made the point that intraday information isn’t just for daytraders, we need to remind users of the NeuroShell DayTrader Pro that Open, High, Low, and Close mean something different now. Low, for example, refers to the low of the last bar (say a 10 minute bar), not the low of yesterday or today. If you want to use the latter, you need to use the new indicators in the “Intraday Basic” category. For example, the “Day Open” indicator is today’s open if the parameter “days back” is set to 0. It is yesterday’s open if “day’s back” is set to 1.