Klimasauskas Group^TM
Leading the Edge in Emerging Technologies

We have worked extensively in applying Neural Networks, Genetic Algorithms, and Fuzzy Logic to a variety of financial applications. Here is a small sampling.

Automated Stock Portfolio Management. One of our staff was a member of a team assigned to develop an automated stock portfolio management system. The objective of the system was to provide weekly buy / sell lists to balance a $100 MM equities portfolio. One of the system requirements was that 100% of the money had to invested at all times. Input data consisted of a variety of technical and fundamental indicators. There were several challenges in developing the system. The data was "noisy" (very uncertain), and there were far more potential variables than made sense. After testing several approaches to variable selection, we discovered that Genetic Algorithms provided the best solution. In fact, the GA was able to identify synergistic subsets of variables that out-performed any other sets of variables. These synergistic subsets also demonstrated excellent performance in the following prediction periods. Another issue with modeling the data was that there were high concentrations of data in small areas of the space where the markets were moving sideways. Many neural networks, in particular, Back-propagation, use the distribution dependent metric sum-squared error. This resulted in sub-optimal performance in areas where the underlying equities were making larger moves. To address this, we invented sub-sampling methods that enabled us to develop models with better predictive performance. Several of these innovations were later incorporated into NeuralWare's NeuralWorks Predict product. How well did it work? The annualized excess rate of return was 2.6%, adjusted for market performance and relative portfolio volatility, and using historical levels of slippage and commissions. Barra analysis indicated that the primary contributor to performance was selectivity. This outperformed 85% of mutual fund managers.

Market Timing with Multiple MACD. A list of moving averages were converted to MACD oscillators by taking all possible differences. These were then combined together in different ratios (with both positive and negative coefficients) to produce a combined signal that was interpreted by a trading strategy. This approach was based on the notion that market behavior is the result of the combination of multiple cycles. The MACD oscillators helped to decode the cycles. Several interesting results emerged from this (see: Klimasauskas, 1993a): The importance of the evaluation function in shaping solution performance, the value of non-standard genetic operators (such as average cross-over) in reducing search time, and islands of evolution for maintaining population diversity.

Risk of Ruin. Have you every wondered whether your trading system is better than chance? What is the probability that someone with a dart board could consistently do better at trading than your system? What are the worst unrealized draw-downs that you should expect? Simulation is one way to find out. Computing risk of ruin is the simplest of simulations. More complex simulations can provide statistical support for how much better than chance (or the market) you can expect a trading system to perform. It can also give you critical insights into the level of realized and unrealized draw downs to expect, and the probability that you lose everything.

Risk Management. Whether identifying true name fraud in internet credit applications, selecting transactions for high probability of fraud, identifying likely candidates for pre-approved credit, or determining insurability, the lessons learned in developing these trading applications apply. Variable selection on noisy data, sub-selection for overall good performance, simulation to provide a statistical basis for deployment. If you are involved in any of these applications, contact us today.

In other applications, our staff has been involved in developing value propositions for cost-justification and auditing large multi-year projects. We have applied genetic optimization to project scheduling to maximize internal rate of return (IRR) subject to constraints on maximum negative cash flow. We have also built economic models of multi-use real estate projects, alternative energy projects, and estate planning analysis including use of irrevocable charitable remainder trusts as a method for maximizing total cash flow.

Key Benefits

• High returns based on stock selectivity.
• Optimal market timing with good forward performance.
• Confidence in the statistical distribution of possible outcomes based on simulations.
• Improved lift and reduced false positive rates, which translate directly to improved bottom line performance.

Capabilities

Klimasauskas Group staff have extensive experience in developing hybrid solutions using Neural Networks, Genetic Algorithms, and Fuzzy Logic. In these projects, we were able to leverage our understanding of these technologies to leverage the benefits while mitigating the weaknesses of existing solutions.

• Genetic Algorithm for synergistic high-performance variable selection.
• Data sub-sampling for improved prediction performance.
• Genetic optimization of multiple MACD oscillators for novel market timing signals.
• Simulation for validating system performance and estimating draw-downs and returns.

Selected Publications

Klimasauskas Group staff have extensive experience in developing novel solutions for real-world financial applications. We understand both the financial applications as well as the technology to improve them. (Klimasauskas Group Staff in bold.)

• Klimasauskas, C. (1997).Data Pre-processing for Profit. Computerized Trading, Jurik, et. al. (ed.). New York: Prentice Hall. In press.
• Klimasauskas, C. (1995a). Developing a Multiple-Indicator Market Timing System: Theory, Practice, and Pitfalls. Virtual Trading, Lederman, Jess (ed.); Klein, Robert A. (ed.). Chapter 6. pp. 127-213. Chicago: Probus Publishing.
• Klimasauskas, C. (1995b). Training Neural Nets to Predict Profit. The Magazine of Artificial Intelligence in Finance. Fall 1995, Volume 2, Number 3, pp. 21-25.
• Klimasauskas, C. (1994). Neural Network Techniques. Trading on the Edge, Deboeck, Guido J. (ed.). Chapter 1. pp. 3-26. New York: John Wiley & Sons.
• Klimasauskas, C. (1993a). Developing a Multiple MACD Market Timing System. Advanced Technology for Developers. October-November-December 1993 Special Issue. High-Tech Communications: Sewickley, PA.
• Klimasauskas, C. (1993b). Implementing a Trading Strategy in Excel. Advanced Technology for Developers. April, 1993. High-Tech Communications: Sewickley, PA.
• Klimasauskas, C. (1993c). Estimating Risk of Ruin. Advanced Technology for Developers. March, 1993. High-Tech Communications: Sewickley, PA.
• Klimasauskas, C. (1993). Issues in Predicting Time Series. Advanced Technology for Developers. January, 1993. High-Tech Communications: Sewickley, PA.
• Klimasauskas, C. (1992a). Why Can’t I Make Money Using my Network Predictions. Advanced Technology for Developers. December, 1992. High-Tech Communications: Sewickley, PA.
• Klimasauskas, C. (1992b). Hybrid Neuro-genetic Approach to Trading Algorithms. Advanced Technology for Developers. November, 1992. High-Tech Communications: Sewickley, PA.
• Klimasauskas, C. (1992c). Accuracy and Profit in Trading Systems. Advanced Technology for Developers. June, 1992. High-Tech Communications: Sewickley, PA.
• Klimasauskas, C. (1992d). Genetic Function Optimization for Time Series Prediction. Advanced Technology for Developers. July, 1992. High-Tech Communications: Sewickley, PA.
• Klimasauskas, C. (1992e). An Excel Macro for Genetic Optimization of a Portfolio. Advanced Technology for Developers. December, 1992. High-Tech Communications: Sewickley, PA.