Chapter 08 – Perils and Promise of Evolutionary Computation on Wall Street



Using Genetic Algorithms, Optimization Models, and Evolutionary Computation on Wall Street

“Be careful what you ask for — you might get it.”

My enthusiasm for machine learning, described at the end of the previous chapter, led me to kiss many artificial intelligence ( AI ) frogs. This included many flavors of inductive and explanation – based learning, as well as connectionist ideas, such as neural nets, that were based on simulating simple nervous systems. There were some interesting notions, but nothing came close to reproducing that “Wow!” Macsyma moment, until I found artificial evolution and genetic algorithms (GAs).

These techniques used populations of solutions, and applied digital versions of the principles of evolution to select the fittest, and to combine the best of the bunch for successor generations of hybridized and mutated solutions. There were some remarkable examples — robots that started out wandering aimlessly and bumping into things evolved before your eyes into what looked like precision drill teams. Symbolic regressions “discovered” complex algebraic relationships instead of just calculating coefficients on an assumed model structure. There were very capable network controllers and logic circuits, all of which emerged from a clearly useless population of random initial solutions.(1)

Promotion of Evolutionary Artificial Intelligence and Genetic Algorithm Applications for Investment Management

I became a major cheerleader for learning in finance using artificial evolution. I attended academic conferences where I met the leading lights in the field and hired their grad students, and where I got to stay in college dorms. It was a refreshing change from all of the four-star hotels favored for investment management conferences. Actually, sharing a bathroom with 25 other people is not so refreshing, but it was a change.

One of the founding fathers was Dave Goldberg, the big dog of evolutionary computation (EC), at the University of Illinois in Champaign-Urbana (also home to 2001: A Space Odyssey’s HAL 9000). In 1989 Dave wrote the classic and still best-selling text, Genetic Algorithms in Search, Optimization and Machine Learning (Addison-Wesley, 1989). I borrowed code from Dave for the experiments described in this paper, and got to speak at the GA and EC conferences. Eventually, Dave felt sorry for me getting in the bathroom line with all the grad students, and reserved one of the prime speaker slots at another conference, arranging to have me quartered in accommodations with my own plumbing at the spectacular Jumer’s Bavarian Grotto in Champaign-Urbana.

Fans of the Madonna Inn in San Luis Obispo, California, which features the Cave Man Room, the Spanish Inquisition Room, and the Li’l Abner Suite, among too many others to mention, would feel right at home in Jumer’s. I gather that Jumer’s has gone more upscale and mainstream, but back in the mid-1990s it was one of the kitsch capitals of the United States. You should trust me on this — I have been to Gatorland, home of the Gator Jump-a-roo, nine times.

Jumer’s had an all-medieval, all-Teutonic all the time theme going on: lots of stuffed actual bears; lots of weaponry on the walls, including in the guest rooms; battle axes, chain mail, and heavy purple draperies everywhere you looked; many, many suits of armor; and an actual stuffed horse, wearing more armor. All were well secured to the walls and floor to discourage University of Illinois students trolling for dorm decor items. I didn’t want to leave for the GA and EC events down the road at the much duller university, but I did, and hired more grad students and borrowed more code.

My willingness to show up at Jumer’s earned me an invitation to do a keynote talk on evolutionary computation and finance at the 2002 Genetic and Evolutionary Computation Conference (GECCO) in New York. GECCO was the major confab for this branch of the AI world, and there were more than a few fellow travelers on the EC – finance trail. I was willing to talk about it, since I’d begun to have some doubts and felt that maybe I’d learn more than I gave away. This chapter is based on that talk.

The AI Spring? — Precursors to Using Genetic Algorithms on Wall Street

The AI winter was not just a write-off for the venture capitalists who had drunk too deeply of the Kool-Aid; it spawned many genuine innovations, which came from questioning in a scientific way what had gone wrong. The symbolic predicate calculus logic programming view of AI had its limitations. Learning how to solve problems in the really messy, noisy, dynamic world was different from theorem proving and chess.

Scientists looking for successful models of learning and adaptive behavior do not have to look far. Birds do it, bees do it, even monkeys in the trees do it. But they all do it using wetware that we understand well enough to appreciate the crucial lessons for the next generation of AI paradigms. There is massive parallelism. Computation is going on all over the place, not in one instruction stream. Brains do not have accumulators.

AI went parallel. Thinking Machines, founded by computational superstar Danny Hillis (son-in-law of Marvin Minsky, the pope of symbolic AI), gathered some of the leading lights to build and program massive machines with up to 64K (65,536) processors. That is a lot more than one, but still a lot less than the 100 billion neurons in the brain.

You don’t need a machine with a billion processors to try out solutions that would use them. A simulator will do fine, if not as fast. For theory buffs, this is an example of the idea of a universal computation; a Turing machine or its equivalent can emulate anything you want. The Nintendo 64 emulators you can run on your PC to play Pac-Man are another.

The neural net movement exploited this idea, seeking to realize learning by mimicking structure and function. Another branch of the turn to biologically inspired approaches to learning used the intriguing idea of mimicking evolution. The mechanics of evolution at the chromosome level — the processes of mutation and crossover, dominant and recessive traits — are understood well. John Holland of the Santa Fe Institute proposed the idea of genetic algorithms, using computers to emulate evolution of solutions to problems, in order to use computer programs to evolve better programs.

Genetic Algorithms Test Whether Financial Models and Trading Rules Have Predictive Ability or Provide Alpha Returns Over the Portfolio Benchmark

Genetic algorithms are a tool for machine learning and discovery modeled on the time-tested process of Darwinian evolution. Potential forecasting models and trading rules are modeled as “chromosomes” containing all of their salient characteristics. A population of these solutions is allowed to “evolve,” with the fittest solutions rewarded by inclusion in subsequent generations. Each individual’s fitness is calculated explicitly as a payoff.

For example, fitness can be measured by predictive ability or alpha (excess return over the benchmark). Solutions with the lowest fitness become extinct in a few generations.

Variety is introduced into the population of solutions by mimicking the natural processes of crossover and mutation. Crossover effectively combines features of fit models to produce fitter models in subsequent generations. In crossover, we blend chromosomes (bit strings) defining two successful models in the hope of developing a still better model.

Mutation stirs the pot and introduces variations that would not be produced by crossover. Here we randomly alter any bit in any chromosome to create a mutation. Most fail badly, but a few survive. As long as we are playing God, we can give the fittest members of the population a free pass into the next generation without participating in the breeding and selection cycle.

Using Genetic Algorithms for Financial and Investment Management Applications

Genetic algorithms have been used successfully in many contexts, including meteorology, structural engineering, robotics, econometrics, and computer science. The genetic algorithm is particularly appealing for financial applications because of its robust nature and the importance of the payoff in guiding the process.

The genetic algorithm is robust in the sense that very few restrictions are placed on the form of the financial model to be optimized. …

>>>>>> READ MORE HERE < <<<<<<

All notes for this chapter about Computerized Investing, Genetic Algorithms, Artificial Intelligence Applications, and Evolutionary Computation on Wall Street:

This article originally appeared in the Winter 2003 issue of the Journal of Investing. It is reprinted in Nerds on Wall Street with permission. To view the original article, please go to iijoi.com

1. A good starting point for genetic algorithms is ILLIGAL, the GA lab at the University of Illinois (www.illigal.uiuc.edu/web/). For genetic programming, John Koza’s work is found here: www.genetic-programming.org/. The Santa Fe Institute, where many of these ideas first got started, is still in the game: www.santafe.edu/research/topics-innovation-evolutionary-systems.php

2. There are videos of some these early examples accompanying John Koza’s books, Genetic Computation I and II (Cambridge, MA: MIT Press). A search for “genetic algorithm demonstrations” turns up hundreds.

3. First Quadrant ( www.firstquadrant.com ), in Pasadena, California. Assets under management were in excess of $20 billion; clients were primarily pension funds.

4. Thanks to Andy Lo of MIT for this clear view of the central issues, discussed in many other contexts in his fine text Econometrics of Financial Markets (written with John Y. Campbell and A. Craig MacKinlay; Princeton University Press, 1996) and popular work A Non-Random Walk Down Wall Street (written with A. Craig MacKinlay; Princeton University Press, 1999).

5. Actually, just the variable specification portion; as will be discussed later, the first two segments were fixed.

6. Using the formula n !/[( n – r )! r !] and recognizing that there are 192 variations on each variable.

7. For information on these models, see “A Disciplined Approach to Global Asset Allocation,” by Robert D. Arnott and Robert M. Lovell, Jr., Financial Analysts Journal (January/February 1989).

Wall Street Analytics

Easily find all of the
lowest cost no load
mutual funds and ETFs
Available at major ebook stores:

Amazon -- Kindle/MOBI
index funds list investing books

Apple iBookstore -- iPad/EPUB
index mutual funds investment guide

Barnes & Noble -- Nook/EPUB
no load mutual funds investment guide

Smashwords -- EPUB, MOBI, PDF
low cost mutual funds investment guide

no load index mutual funds
by David Leinweber