As of the early 2000's, Internet has opened paths to a new form of collective intelligence, viz., human or humanoid computing. Be reassured, reader, this is not a way to turn your brain into a computer per se. What it does is coordinate thousands of connected people together to assemble a computer ‘power’ that exceeds, for certain complex problems, what is already available today via supercomputers. The specialty biochemistry was first on the track, and now we see financial affairs joining in. So, let’s see what happens when amateurs take over a market trading room.
The basic idea behind market economy is what James Surowiecki called the wisdom of crowds: large numbers of people adjust prices more precisely than a single actor, even if the latter is perfectly well informed while the crowd is not.
Spontaneous mobilization of collective intelligence is one of the bases of today’s economics. But it is also possible to develop and mobilize their collective computing power in a more specific way. And, although the idea itself is fairly old, it is opening up some awesome possibilities today.
Human computing with a large scale organization
Let’s begin with a story that goes back to the 19th century. After France’s inglorious defeat in 1870, the country was hell-bent on preparing for revenge. The commanders in chief identified one of the reasons for losing the war: French horses were too heavy, too slow on the ground manoeuvres faced with German counterparts. The challenge to “improve homebred horses” became a leitmotiv, to ensure the needs of the military, but also in agriculture, transport in general both civilian and military. The question was, how do you go about improving horse breeding characteristics?
In 1891, the French Government came up with the idea to use the greed of horse-racing betters, in those days a very popular habit. A Government decree was issued that required the Pari Mutuel (France’s original national betting service) to “improve work-horses and military mounts” (to quote the decree authors), reporting to and under the control of the Ministry for Agriculture of the time.
Emile Riotteau, author of the text invented what probably was the ancestor of large scale human computing. It could already have warranted the name of human computation – there was no underlying theory and, of course, no computers, instead the expression of an ingenious stroke of intuition which, in essence saw the problem sub-contracted to a crowd of amateurs (the betters), the horse trainers and breeders – a problem that was far too complex for specialists. “Horse races are not the objective but the means, under the control of public authorities, to select thoroughbred steed, either to perpetuate their race or improve other native races through additional nervous influx” (from the Foreword to A History of horse racing from Antiquity to the Present, 1914). Today, we could describe the 19th Century Pari Mutuel as a supercomputer in which the CPU (central processing unit) is the better.
It was noted, for example, from 1871 to 1890 that the horse that came first at the Jockey Club Prize, Chantilly ran the track taking an average 2 min 42s whereas between 1891 and 1914, the average track time dropped to 2 min 36 s. The 3 fastest ever horses in this event ran the race in 1905, 1911 and 1913.
In today’s world, overwhelmed as it is in a digital wave, a world where computers are beginning to beat Chess Grand Masters, the idea of human computing is resurfacing. As of the end of the 1990s, various Internet companies started enrolling Internauts to help out on processing, selection and data improvement. Google, Facebook and Twitter consider users as both potential clients and as cogs in a computing process that enriches a data base that becomes more and more ‘relevant’, i.e., customer focused. In this way, the large number of social network users mechanically improve the media (indeed, there is probably a square law effect here).
Today’s human computation is a special, noteworthy variation of collective intelligence described back in the 1990s. The aim is to implement a task or run a computation by deliberately considering the human agents as a resource in a process that is beyond their individual reach. The agents serve the machine, and it is therefore no longer necessary to understand all the inner cogs or even its ultimate objectives.
Mobilizing this form of ‘intelligence’ is no longer reserved exclusively for the Internet majors. In 2009, some biochemical research scientists realized that their professional computers were unable to solve complex problems and so they set up a collaborative, community-based work platform where amateurs were invited to tackle protein folding problems. This Fold-it [in English] project rapidly became the community platform, the motto being « Solve puzzles for science ».
Thousands, literally of active player connected to the platform totally outperformed the powerful computers the scientists had, and they delivered unusual, original and relevant solutions. The results were published in Nature in 2010. Fold-it, a game, now constitutes a supercomputer of a totally original design: it is built round an assembly of human CPUs.
When the amateurs take over a market trading room
For a financier or a ‘quant’ [a quantitative analyst/trader] in a trading room, it is very important to be able to access high level computational power to determine their trading strategy. The gigantic amount of data in a market place and the stupendous number of possible correlations takes any calculator-based approach systematically to its limits. It is necessarily attractive to learn that with a computer of the power like Fold-it, there could be a competitive advantage in market trading operations. This was the context which saw the launching of Krabott [in French].
This programme invites amateur traders to join in a ‘game of strategy’ on a level of complexity comparable to Fold-it, replacing the trading room floor intelligence by a crowd of anonymous participants who otherwise are more likely to be fans and adepts of on-line poker or game like World of Warcraft than aces of differential equations. Krabott is like an aircraft in which the passengers are collectively invited to take the cockpit controls and land the aircraft … better than the pilots could do themselves. Krabott is a trade-room game in which the players are the breeders for trading strategies each of which is called a Krabott. Players choose their own Krabott among set of available Krabotts. The players then observe how their Krabott performs/behaves in the monetary market (FOREX). Each Krabott will take positions simultaneously on 4 different currencies: the Swiss Franc (CHF), the US dollar (USD), the Yen (JPY) and the Sterling Pound (GBP), buying and selling rates when he closes his orders (weekly, Friday evening at latest). Gains or profits (P) and losses (L) accrue to, or are deducted from the players’ account scores. The players then decide to discard the least efficient Krabotts or to place them temporarily in a test or stand-by zone to observe their behaviour without impacting on their individual scores: the standby zone is the nursery.
Last but not least, the players can reproduce two Krabotts by crossing their parameters and thereby create a new single strategy that can be tested ‘live’, in the trading room and be reassessed. The cross-over mechanism was inspired by cross-fertilization techniques in biology and is in fact a genetic algorithm in a variation proposed by Alex Kosorukoff in 1999 as the “Human Based Genetic Algorithm”. Kosorukoff showed statistically how human dexterity proved better compared with machines when carrying out blind selections of the best reproduction breeders in a given population
Genetic Algorithm– principle
Krabott reproduction is the key to human computation. The main feature of the game is its capacity to integrate human players in the crucial phase of parameter optimization in the genetic algorithm, viz., reproduction. My thesis [in French with an English abstract]defended for my PhD in Sept.2013 on “Human computation applied to algorithmic trading” demonstrated that amateur players obtained better trading results that machines when it came to identifying and selecting Krabott best qualified for reproduction.
During our experimentations, we arranged for a competition between a machine (computer) capable of testing and assessing approximately 100 000 different strategies over a 9-month period and a set of around 100 players who explored some 1 000 strategies manually. The results were undisputed, inasmuch as the players despite their exploratory capacity being 100 times less than that of the machine, constantly came up with more efficient (better profits) strategies that those identified by machine.
Comparative P&L between human computing and 100% machine computation, 1% gain for the players at close of trading compared with 1% loss for the machines.
Our next step, in the view of this encouraging result, was to automate the design for a sort of meta-Krabott, or meta-community-signal that would rely on analysis of a multitude of strategies identified by the human players.
The “signal” consolidating the position of the 6 best Krabotts of the preceding week, also largely outperforms every other classic form of optimization (in our experimentations, the initial capital outlay was doubled up with astonishing regularity over a 5 month ‘live’ test period).
The paradox here is that the Krabott experiment showed that if the players had the opportunity to explore elsewhere in the ‘field of possibles’, they generally proved inept to choose the best strategies by themselves. And curiously enough, the average performance of the players was somewhat mediocre: the reason was that since they were impatient to obtain results, they tended to place their stakes more easily on strategies that were immature or too risky.
Another important point in the experiment is the absence of back testing. This approach offers a systematic technique to be used when designing trading strategies, either with hedge investment funds or with the banks. Each new strategy is tested and optimized over years of back-tests before it is actually used in a real ‘live’ context. Of note also, a strategy has a life expectancy of several months to a few years before being abandoned given that the market conditions vary rapidly over time and obsolescence is rapid.
In the case of the Krabott meta-signal, the component strategies are updated every week and each Krabott taken singly is tested over short periods (3 to 12 months max.) by the players. Our approach consists of challenging the trading strategy contents at a very high frequency, this enabling the signal to adapt to market changes and leads to a totally unique dynamic strategic approach to trading.
The ‘field of the possibles’ is in fact unlimited and still remains wide open to exploration, all the more so that it is counter-intuitive; We can readily imagine that in numerous applications, the idea of bringing in players to solve a complex problem is especially weird. Krabott, in the wake of Fold-it, notwithstanding demonstrate the intrinsic power of the approach. We should also be asking ourselves about the nature off such collective intelligence. Where does it originate and how does it form and develop? Is the human brain somehow “wired” for collective thinking? These are questions today that do not carry an answer and thus will lead on to basic research that, will allow us, one day perhaps, to come up with new sophistications. But to stay modest for the time being, let me just end by predicting that collective intelligence has a bright future ahead.
News, in shortKrabott was the subject of a thesis defended in November 2013 at Mines ParisTech (publication under press). Since 2014, the company AVIOMEX SAS houses the Krabott system. In 2014, Krabott trading meta-signals are commercialized for the benefit of an association whose aim is to aid isolated farmers in difficulty, viz., « Solidarité paysans ».