July 31, 2009
AAAI Asilomar Meeting
John Markoff's NYT article "Scientists Worry Machines May Outsmart Man" touched off a mini-firestorm this week. The article refers to a meeting of AI scientists held at Asilomar (a conference center near Monterey) in February to discuss societal implications of future AI technology. The provocative headline may have had something to do with the reverberations in other media outlets, and on the blogs. The most egregious example I saw was an entry by Dan Smith on popsci.com:
The long-awaited robot-led holocaust may happen any day now. That seems to be the finding of a secret conference of the world's top computer scientists, roboticists, and artificial intelligence researchers.
As it happens, I participated in this Asilomar meeting (which was not at all secret), and can assure anybody reading this there was no finding that a "robot-led holocaust" is imminent. (And who exactly has been awaiting this?)
Ironically, one of the goals of this meeting was for the major AI professional organization to develop a stance to engage the broader public in a reasoned discussion about societal implications of AI technology. Certainly artificial intelligence and other means of increased computational scope in our world has transformative effects on the economy and the world, and preparing for these is only responsible behavior. Many AI scientists are uncomfortable with the existing public dialog, which is to a large extent driven by science fiction writers and singularity prophets. Accordingly, the discussion tends to emphasize broad utopian or distopian visions, rather than nearer-term practical implications of technology. As a result, the "debate" often takes a hysterical (therefore non-constructive) tone.
As current President of AAAI, my colleague Martha Pollack (Dean of the School of Information at U.Michigan) was called on to represent the AI field in this brouhaha, and pressed valiantly to counter the panic reflex. A telling episode was her appearance on Fox News, where the host framed the discussion with: "I'm scared. Tell me why I shouldn't be." Martha's appearance was a success in that by the end he said he was no longer in a fright state. But overall it is pretty depressing that the public discussion is at the level of whether we should panic or not. (This is not at all to pick on Fox News. They were "fair and balanced" in this instance, and completely typical of how the media handle AI.)
July 30, 2009
Cost/Benefit of High-Frequency Trading
On Marginal Revolution, Tyler Cowen discusses high-frequency trading and gets to the nub of the issue:
The philosophical question is why it might possibly be beneficial to have market prices adjust within five seconds rather than within fifteen. One second rather than five? 0.25 rather than one?
Of course, it's an economic question, not just philosophy. A commenter named "a student of economics" argues persuasively that there is in fact no benefit at these short time scales:
The social benefit is the net present value of having "accurate" prices a few fractions of a second earlier. This benefit is a function of the discount rate, which is customarily expressed in annual terms, and the value of the changes in decision-making that might result under the new, more "accurate" prices.
I submit that the social benefit is trivial.
The private benefit is in the billions, but that is almost 100% rent-seeking. It is gained almost entirely at the expense of another party who's computer or trading algorithm might be fractionally slower.
This is an nice example of almost pure rent-dissipation. It's a case where the invisible hand of the market does more harm than good, by directing enormous capital resources and some of our most brilliant minds toward an activity that creates essentially zero value.
We should set up the rules of the game to maximize incentives for value creation and minimize incentives for rent dissipation. That suggests banning or heavily taxing HFT.
Rather than banning or taxing, we could adopt market rules that eliminate the benefit of HFT, as in the discrete-time trading scheme (one-second call markets) advocated in my previous post.
Countering High-Frequency Trading
The recent NYT article by Charles Duhigg on high-frequency trading (HFT) has set off a flurry of argument about the benefits and threats of this activity to financial trading systems. The revelation that some systems provide advance information (exposing incoming orders 30-500 milliseconds before they are submitted to the general market) to select HFT systems has drawn particular fire. Some have suggested that rapidity of response capability per se could open up manipulation possibilities or is otherwise destabilizing. We have also seen questions about whether diverting trade surplus toward whomever builds the biggest fastest network is an efficient use of resources, and the implications for perceptions of fairness across the trading public.
Let us start from the premise that asymmetry of information about incoming orders is inherently undesirable. Leveling the playing field in information promotes efficiency and lowers the cost of entry for the broader investing public.
The root of the problem, in my view, is the system's support for continuous-time trading. In a continuous market, trades are executed instantaneously whenever there are matching orders, and introduction of an unmatched order likewise causes an instantaneous update to the information available to traders.
An alternative would be a discrete-time market mechanism (technically, a call market), where orders are received continuously but clear only at periodic intervals. The interval could be quite short--say, one second--or aggregate over longer times--five or ten seconds, or a minute. Orders accumulate over the interval, with no information about the order book available to any trading party. At the end of the period, the market clears at a uniform price, traders are notified, and the clearing price becomes public knowledge. Unmatched orders may expire or be retained at the discretion of the submitting traders.
Even with a period as short as one second, the call market totally eliminates any advantage of HFT systems. It does not eliminate the opportunities for algorithmic trading in general--just those that come from sub-second response time. No party has privileged information about order flow, and no party benefits by getting a shorter wire to the "trading floor".
Moreover, the call market eliminates the disparity in risk incurred by those who dangle limit orders in continuous trading systems. Given the latency in retracting an order, an HFT system can take advantage of existing orders when new information becomes available. Because of this, systems need to provide extra incentives for the so-called "liquidity providers" who are willing to maintain a limit order. In the discrete-time systems, nobody can hang back and prey on liquidity providers. In order to trade, one must incur the one-second (or whatever) exposure. But everyone is doing that, and no information about orders becomes available anyway in that interval, so this exposure spreads out the risk evenly.
Finally, call markets also have efficiency advantages over continuous mechanisms. Aggregating orders over time eliminates some of the noise in matching based on arbitrary arrival sequences, thus producing higher surplus overall, and less price volatility. There is a classic tradeoff here between efficiency and execution delay. But with clearing intervals as short as a second, it seems hard to argue that this delay has very much real cost.
Switching over to discrete-time systems entails some technological changes, all well within feasibility in my view. The conceptual change of mindset required of the trading community may be a more serious challenges, but the benefits could be immense.
July 29, 2009
2009 Trading Agent Competition Results
The tenth annual Trading Agent Competition was completed earlier this month at the IJCAI-09 conference in Pasadena, California. TAC-09 featured three games: the supply chain management (SCM), market design (reverse TAC, or "CAT"), and Ad Auction (AA) games. Preliminary rounds began in June, involving 42 teams (13 in SCM, 14 in CAT, and 15 in the new AA game) from 14 countries. A full list of participating teams is available.
Based on the qualifying and seeding rounds, the 14 CAT agents, 15 AA agents, and 12 of the SCM agents proceeded to the final tournament, on 13–14 July. The first day (semifinals) coincided with the Workshop on Trading Agent Design and Analysis (TADA-09), where researchers presented their latest results on trading agent technology.
The competition was as always an exciting event. Although the table below lists only the top-scoring few agents for each game, many entries exhibited strong trading strategy. Complete SCM-09 score tables are also available.
|CAT||1 Jackaroo||University of Western Sidney|
|2 CUNY.CS||City University of New York|
|3 IAMwildCAT||University of Southampton|
|SCM||1 Deep Maize||University of Michigan|
|2 TacTex||University of Texas|
|3 MinneTAC||University of Minnesota|
|Ad Auction||1 TacTex||University of Texas|
|2 AstonTAC||Aston University|
|3 Schlemazl||Brown University|
A few articles and other descriptions of TAC-09 agents are currently available. We will post additional reports as the participants provide post-tournament agent descriptions and analyses.
TAC-09 would like to thank our bronze sponsor, Advanced Technologies Integration, Inc.
The University of Michigan also acknowledges the generous support of Microsoft, Yahoo!, and Google enabling the production of the TAC Ad Auctions game.