It’s Time to Look Seriously at Artificial Intelligence Systems

By Lou Mendelsohn and Jon Stein

Artificial intelligence systems are information processing models which mimic how the human brain processes information. Unlike conventional, rule-based technical trading systems popular in the 1980s, neural systems do not need predefined trading rules or “optimization” of technical indicators to generate trading signals. Instead, through an iterative “training” process, neural systems “learn” the underlying associations and causal relationships within technical, as well as fundamental, data affecting a specific stock’s or commodity’s price. Then, with a high degree of accuracy, neural systems can forecast future prices and trading signals for that market.

Artificial intelligence systems are also called neural networks, neural computers, adaptive systems, naturally intelligent systems, or neural nets. They are modeled after the structure and function of the brain. Because they can generalize from past experience, neural systems represent a significant advancement over rule-based trading systems, which require a knowledgeable expert to define “if-then” trading rules to represent market dynamics.

It is practically impossible to expect that one expert can devise trading rules which account for, and accurately reflect, volatile and rapidly changing market conditions. Inflexible, rule-based systems simply are not dynamically adaptive, despite periodic re-optimizations of a system’s indicators.

While today’s trading systems, utilizing historical optimization procedures, risk becoming “over-optimized” or “curve-fitted” when too many technical indicators or rules are employed, neural systems gain in predictive-ness as more data inputs are used during training. However, it’s not as easy as it sounds. Developing a profitable artificial intelligence trading system is very much an art and not a science that can be followed cookbook-style. There are many serious design issues that must be addressed when developing and training a neural trading system, if it is to be predictive, and most importantly, profitable.

Within the next several years, artificial intelligence systems will become widely used by “cutting-edge” traders throughout the world. Before long, today’s emphasis on single-market technical analysis will loose its appeal, as traders apply neural trading systems to the task of forecasting prices and trends in the financial markets.

How are neural network systems “trained” to find the underlying market patterns and hidden relationships? What types of intermarket technical data and fundamental information are utilized during training? What are the steps to be followed in developing and training neural trading systems? Finally, what kinds of artificial intelligence systems are available for traders? These are just a few of the questions that need to be answered as traders become more acquainted with this new trading technology.

Neural systems consist of layers of neurons which are connected to each other. There are typically three types of layers, an input layer which receives information such as technical price data that is supplied to the system for analysis, a hidden layer which is used by the system for internal processing, and an output layer which provides forecasted output generated by the system such as tomorrow’s price range and trading signal for a particular market.

Typically, neurons within a layer do not connect to each other. Neurons between layers communicate with one another by having specific mathematical weights (or connection strengths) assigned to their connections.

For example, you may want to develop a trading system to predict the next day’s Treasury Bond prices. Designing the appropriate architecture for your neural system is quite exacting, with more than a dozen different neural designs available. One type of neural system that I have used extensively for financial forecasting applications is known as a “feed forward”, “back propagation” system with “supervised learning”.

Before training, the neural system has a “blank mind”. Then you provide the system with an extensive amount of inter-market technical data related to T-Bonds, including various currencies, Eurodollars, the U.S. Dollar Index, the S&P 500, as well as fundamental data such as the Fed Funds rate. Neural systems carry the concept of “intermarket analysis” to its logical conclusion by being able to mathematically analyze and weigh the relative impact that each input market has on the predictiveness of the system.

These inputs must be preprocessed or “massaged” using various statistical procedures, in order to meet the system’s training requirements. Then they are paired with actual daily prices on Treasury bonds (the desired output). It is critical that the system’s architecture, learning method, input data, outputs, and massaging techniques are judiciously selected in order for the system to train properly.

Learning is accomplished through a complex, mathematical, iterative process whereby the neural network is “trained” on the input data using statistical error analysis.

During training, whenever the system’s projections are incorrect, the connection weights between neurons are modified to minimize such errors during subsequent iterations. Each input/output pair of data is called a fact. The system learns by having these error signals propagate backwards through the neuronal layers to prevent the same error from happening again each time a fact is presented to the system during training, hence the name back propagation with supervised learning.

This iterative process is repeated until the system is successfully trained. Once this is accomplished, it is a simple task for the system to provide its forecasted predictions.

After training, you can test the system on new input data to access its output predictions under real-time conditions. This is analogous to today’s “walk-forward” or “out-of-sample” testing. Depending on the results, it may be necessary to modify the system’s architecture, data inputs, massaging techniques, or desired output before retraining.

To run a trained neural network in real-time, you would perform daily updates on the input data, so that the system can then generate its output predictions. At various time intervals, perhaps quarterly, you should retrain your system on up-to-date data, as well as experiment with new data inputs and preprocessing methods.

Where is this state-of-the-art technology in trading systems heading? Right now, the field is in an embryonic stage. While there are many off-the-shelf generic neural software training simulators available (such as Brainmaker by California Scientific Software and Neuroshell by Ward Systems Group), none are specifically geared toward the financial markets. As a result, working with these programs can be challenging for traders without backgrounds in engineering, math, or computer science.

Pre-trained artificial intelligence trading systems designed for stock and futures trading are just now becoming available. These application programs will let traders apply artificial neural systems to the financial arena, handling all data conversions, input preprocessing, and system design in a user-friendly, menu-driven format. Before long, sophisticated neural systems, perhaps even employing innovative architectural designs and learning paradigms, will be applied to a wide range of financial forecasting applications.

Lou Mendelsohn, president Market Technologies Corporation, Wesley Chapel, Florida, designs and tests artificial intelligence trading systems for the financial industry.