A year ago, I spoke to the University of Pennsylvania’s Michael Kearns about whether we might be able to do something to help prevent a much worse reprise of the May 2010 flash crash. The short answer is that no, we can’t — or won’t, in any case. But the longer answer is that there is something we could do, if we just had some will and a lot of money:
You can imagine trying to build an ambitious, reasonably faithful simulator of our current markets. You’d have high-frequency algos, shorter-term stuff, dark pools, multiple exchanges, etc. A giant sandbox.
If you do a simulation and you try some perturbation or stress, and it tells you that a disaster happens, then it’s worth thinking hard about our current markets. But if you don’t find a disaster, that’s not reassurance that some other disaster won’t happen.
I’m proposing a quant version of the stress tests that were proposed for banks.
A car company, before they roll out a product, have a lab environment where they put it through tests. And in reality problems which weren’t tested for get discovered. We’d be much better off with a simulator.
We have no such lab for our financial markets. This strikes me as a little off.
People on Wall St think about simulation, but not for catastrophe prediction, just for their own trading purposes.
Kearns’s idea didn’t get anything like the traction it needs, and it’s not going to happen. But now a new paper from the UK’s Government Office for Science, written by Dave Cliff and Linda Northrop, lays out the case for building such a simulator over the course of 47 very interesting pages.
First of all, they write that the whole global economy “dodged a bullet” on May 6, 2010: if the Flash Crash had just happened a couple of hours later — and there’s no reason it couldn’t have done so — then the US markets might well have closed before the Dow had a chance to recover. The US sell-off would have triggered big market swoons in Asia and Europe, with very nasty consequences for, among many other things, Greek debt dynamics.
More generally, they write,
The global financial markets have become high-consequence socio-technical systems of systems, and with that comes the risk of problems occurring that are simply not anticipated until they occur, by which time it is typically too late, and in which minor crises can escalate to become major catastrophes at timescales too fast for humans to be able to deal with them.
Cliff and Northrop say that we should do exactly as Kearns suggested:
The proposed strategy is simple enough to state: build a predictive computer simulation of the global financial markets, as a national-scale or multinational-scale resource for assessing systemic risk. Use this simulation to explore the “operational envelope” of the current state of the markets, as a hypothesis generator, searching for scenarios and failure modes such as those witnessed in the Flash Crash, identifying the potential risks before they become reality. Such a simulator could also be used to address issues of regulation and certification. Doing this well will not be easy and will certainly not be cheap, but the significant expense involved can be a help to the project rather than a hindrance.
There are many reasons why this is not going to happen, starting with the fact that no one, right now, can afford to do it. Cliff and Northrop rather hopefully say that if this market simulator is expensive enough, then lots of Wall Street players will pay up to have access to its results — but in reality they’re much more likely to do everything they can to stop it from being built in the first place. Because if it is built, the certain consequence will be more regulation:
It may also be worth exploring the use of advanced simulation facilities to allow regulatory bodies to act as “certification authorities”, running new trading algorithms in the system-simulator to assess their likely impact on overall systemic behaviour before allowing the owner/developer of the algorithm to run it “live” in the real-world markets. Certification by regulatory authorities is routine in certain industries, such as nuclear power or aeronautical ￼engineering. We currently have certification processes for aircraft in an attempt to prevent air-crashes, and for automobiles in an attempt to ensure that road-safety standards and air-pollution constraints are met, but we have no trading-technology certification processes aimed at preventing financial crashes. In the future, this may come to seem curious.
Even if regulators don’t have to sign off on trading strategies on an algo-by-algo basis, there’s really no point in building a hugely expensive and complex market simulator if the results of the simulations don’t result in constraining market participants somehow. And I can assure you that no amount of “you’ll all be safer” pleading with banks will persuade them that more regulation and constraint is ever going to be welcome.
Gillian Tett, too, is skeptical that anybody’s going to go ahead with a project of this magnitude:
Most regulators still prefer to forget May 6 rather than admit in public that they are struggling to understand how modern markets really work. And that, sadly, is unlikely to change, unless there is another flash crash.
And there’s a bigger reason, too, why it’s not going to happen: for all its ambition, a financial-market simulator wouldn’t actually address any of the causes of the financial crisis we just had, and probably wouldn’t address any of the causes of the next one, either. As Cliff and Northrop write,
The concerns expressed here about modern computer-based trading in the global financial markets are really just a detailed instance of a more general story: it seems likely, or at least plausible, that major advanced economies are becoming increasingly reliant on large-scale complex IT systems (LSCITS): the complexity of these LSCITS is increasing rapidly; their socio- economic criticality is also increasing rapidly; our ability to manage them, and to predict their failures before it is too late, may not be keeping up. That is, we may be becoming critically dependent on LSCITS that we simply do not understand and hence are simply not capable of managing.
We could try to spend hundreds of millions of dollars simulating and examining the fine-grained architecture of securities trading and high-frequency algorithms; and even if we were incredibly successful in that endeavor, there would still be hundreds if not thousands of other large-scale complex IT systems which can and probably will fail catastrophically at some point. We can’t simulate them all. So why pick on the stock market?