Encoding Trader 'Horse-Sense':
Ideas and Experiments using Historical Foreign Exchange Data
Selene Makarios
Stanford University AI Lab
Trade with the trend. Cut losses short. Let gains run.

In classical Artificial Intelligence, one important paradigm is the logical or algorithmic encoding of common-sense information and heuristic procedures, and their incorporation into computational processes. The three trading dictums bulleted above, which we will characterize here as "horse-sense" trading principles, are often found explicated in lay-audience treatments of trading, and elsewhere. We are investigating whether these heuristics can be gainfully incorporated into algorithmic trading. This approach is qualitatively different from traditional price-series modeling and prediction techniques involving best-fit autoregressive moving-average or conditional heteroscedastic models, etc.

We present results of experiments with one formalization of the trading horse-sense principles. Using 53 foreign exchange currency-pair price histories, the experiments offer evidence that suitable encoding and application of the three principles can result in a parameterized trading model in which historically-inferred parameter values yield algorithmically-driven returns exceeding those of the analogous strategy with randomized parameters, that is, a strategy that ignores the historical data. Both the historically-based and randomized strategies have equal position bias. All comparison testing between the historical and randomized strategies was done out-of-sample.

We also discuss how the algorithmic machinery developed can be used to construct a statistical discriminator of historical forex price-charts versus price-charts generated by Wiener/Martingale processes having means and variances comparable to those of the historical data. A total of 36 random price-charts were used, including series both with normal and with log-normal error term distributions. For the populations of historical and randomly generated price-charts tested, the discriminator produces figures of merit which reject the null hypothesis of a common mean for the two different groups, at above the 99% confidence interval. These results could be construed as evidence against the Random Walk Hypothesis.

Finally, we describe an interesting relation observed between the performance of the algorithmic system on price signals and the raw information content of the signals, as heuristically determined.

Bio: Selene Makarios is a Research Associate with the Stanford Artificial Intelligence Lab. Her work there in knowledge representation and computational logic has included techniques for declarative control of automated theorem provers, techniques for formal reasoning via reification of structured propositions, and logical techniques for dealing with systems of action and change. Her work also included providing the first complete model-theoretic foundation for a quantified, generalized form of McCarthy's Context Logic.

Selene graduated from M.I.T. with a Bachelor's degree in Mathematics. She worked for several years in commercial software development of computer animation and hypermedia systems, before taking a research position at the University of Maryland in artificial intelligence and parallel and distributed computing. She simultaneously enrolled as a graduate student in the Computer Science Department and earned her Ph.D.

She spent a year in a post-doctoral position working in parallel algorithms and computational physics, after which she moved to Silicon Valley, where she has held an evolving series of engineering, research, technologist, and entrepreneurial roles in artificial intelligence, distributed computing, and web-based systems.