Applying Artificial Neural Systems to the Financial Markets

By: Lou Mendelsohn

For the past decade, stock and futures traders have come to rely upon various types of technical analysis software programs to make trading decisions. Now, for the first time, a totally new technology known as artificial neural systems is being applied to this task by sophisticated traders.

Neural trading systems, by contrast with conventional analytic methods, do not employ predefined trading rules or optimization of technical studies to arrive at trading signals. Instead, through an iterative training process, neural systems learn the underlying associations and causal relationships within technical and fundamental data impacting on the price of a specific equity, commodity, or stock index. Once a neural system has been designed, trained, and tested, it can be used to predict subsequent prices and the future trend direction for a given market.

Artificial neural systems are often referred to as neural nets, adaptive systems, neural networks, neurocomputers, and naturally intelligent systems. Since their design is inspired by how the human brain processes information, neural systems actually learn to generalize from past experience. Neural systems, thereby, represent a completely new form of computer-based intelligence.

How can neural systems be trained to find the underlying market patterns and hidden relationships within today’s global markets? What kinds of data are utilized during training? Finally, what are the steps to be followed in developing and training neural trading systems? These are just a few of the issues that should be understood, if this next-generation technology is to be used effectively by traders.

Inter-Connected Neurons

Neural systems are made up of layers of interconnected neurons. There are typically three types of layers: (1) an input layer comprised of preprocessed technical and fundamental data, (2) a hidden layer which is used by the neural system for internal learning, and (3) an output layer which generates the predicted outputs.

While neurons within a given layer do not communicate with one another, neurons within adjacent layers do, with mathematical weights or connection strengths assigned to their connections.

Before training, these weights are randomized so that the neural system starts with a blank mind. Proper selection of the system’s architecture, learning method, input data, preprocessing techniques, and outputs are critical for successful training to take place.

Massaging Techniques

The data inputs are determined, then preprocessed or massaged using various statistical and analytic techniques. This is the most challenging aspect of neural system design. These inputs are paired with known values that the system is to forecast (the desired output). Each input/output pair of data is called a fact.

Training involves an iterative process whereby the neural system learns the underlying patterns within the data by comparing its forecasts with the known values to compute error signals. These signals then propagate backwards through the layers, modifying the connection weights between neurons to reduce such errors during subsequent iterations. When the system’s overall error level is minimized, it is fully trained.

At this point, the system is evaluated on new input data under real-time conditions. This is analogous to today’s walk-forward or out-of-sample testing. Depending on the results, the choice of layers, neurons, data inputs, massaging techniques, or learning algorithm may need to be modified. This redesign process is extremely research intensive, involving time consuming trial and error, in which the system is designed, trained, tested, redesigned, retrained, and retested, etc.

Forecasting Prices

Once this process has been completed, the final neural system will be able to forecast prices and signals in real-time with a high degree of accuracy. The trader would only have to provide it with the necessary input data each day. Then the system generates its output predictions for the following day. At various time intervals, even as often as every two weeks, the system can be retrained on up-to-date data, to allow it to increase its intelligence.

Where is this technological breakthrough in market analysis likely to lead? While there are quite a few plain vanilla neural training simulators already available, presently very few sophisticated financial application programs have been commercially developed for use by traders.

With a fairly steep learning curve for most traders, pre-trained turnkey neural trading systems, such as the VantagePoint-Analyst which I developed for the futures and index markets, are the easiest way for traders to become acquainted with this new technology, without having to start from scratch designing and training their own neural systems.

Such pre-trained systems handle all decisions about data inputs, preprocessing methods, and system architecture design, offering the trader a user-friendly, menu-driven format for daily updating of new price and trend predictions.

Over the next several years, hybrid artificial intelligence-based systems, even combining expert systems, genetic algorithms, and neural nets will be amalgamated for price forecasting in the financial markets. Before long, this emerging technology will result in a significantly broadened definition of technical analysis from what it is today.

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