Why the banks collapsed, and how a paper on Haskell programming can help stop it happening next time

Trading Risk

The financial system exists to trade three kinds of thing: money, commodities and risk. Money and commodities are the easy bit. Either I have £1,000 or I don't. Similarly with commodities: either I have a barrel of oil or I don't. But risk is a very different matter. In theory risk is easy to quantify: just multiply the probability of something by the cost, and you have the expected loss. But in practice its not so simple because the probability and cost may be difficult to quantify, especially for rare events (like, say, a global credit crunch). Many of the factors that go into a risk model are subjective, so honest people can have genuine disagreements about exactly what the risk is.

The Slippery Slope

Unfortunately risk assessment is not value-neutral. Risk has negative value: you have to pay people to take it off you. The higher the risk, the more you have to pay. And because the amount of risk is always debatable this is a very slippery slope; the people paying others to take the risk away have every incentive to present a lower estimate. Everyone can see that everyone else is doing the same, and so methods of hiding or downplaying risk migrate from dodgy dealing to open secret to standard practice.

Specific examples abound throughout the recent history of the finance industry;

• The retail mortgage houses that originally lent to "sub-prime" clients would hire valuers who were known to be on the generous side with their valuations. So any valuer who wasn't so generous found their income drying up. Background checks on clients were cut back, then eliminated. Eventually borrowers were simply told what income to claim on the forms, regardless of what they actually earned.
• These loans were then bundled up and sold. The idea was that the buyers would each get a share of the incoming loan repayments. Rights to this stream of money were divided into "tranches", the idea being that, for instance, Tranche 1 would get the first 34% of whatever money was due, Tranche 2 would get the next 33%, and Tranche 3 would get the last 33%. When some borrowers defaulted (as some always do), Tranche 3 would lose out first, then Tranche 2. Tranche 1 would only fail to get all their money if the overall repayment rate fell below 34%, which had never happened. The game here was to persuade a credit rating agency that Tranche 1 was so safe that it was worthy of a "Triple A" rating, because that meant that banks, insurance companies and similar big financial institutions could legally buy this debt without having to put cash aside to cover potential losses. The rating agencies earned fees for evaluating securities, so just like the house valuers they found it paid to be on the generous side.
  
• All these institutions had risk management departments who were supposed to watch out for excessively risky behaviour. But in practice they found it very difficult to blow the whistle. Risk managers tell stories of being given two days to review a deal that took ten people a month to negotiate, and of accusations of "not being a team player" when they questioned over-optimistic claims. This story from the New York Times has some details. Look through the comments after the story as well; many of them are by people with their own tales of risk.

None of this is new; similar behaviour has contributed to past financial crises. In theory more regulation can prevent this, and everyone is now planning or demanding lots more regulation (even the banks). But in practice regulation has failed repeatedly because the regulators always concentrate on the way it went wrong last time. Regulations have to be written carefully, and the companies being regulated have to be given time to understand changes and put their compliance mechanisms in place. This prevents regulators from moving as fast as the institutions they regulate.

The regulators also don't have visibility of the information they need to assess systemic risk. Systemic risk arises because financial companies are exposed to each other; if one institution fails, others have to write off any money it owed them, possibly pushing them into bankruptcy as well. Regulators try to make companies insulate themselves by avoiding excessive risk and keeping some cash on hand, but without a clear picture of the risks being run by each company they have no way to tell if this is enough.

The basic problem, I believe, is the food chain of risk management within each institution. At the top are the negotiators and fund managers who design and package the securities. Then the lawyers are bought in to specify the precise terms of the deal. Somewhere along the way the "quants" will be asked to develop mathematical models, and at the bottom coders will be given the job of turning the models into executable code that will actually determine the real price and risks. It is this food chain that needs to be rethought, because its hiding important information.

This 2000 paper by Simon Peyton Jones, Jean-Marc Eber and Julian Seward shows a way forwards. It describes a Domain Specific Language embedded in Haskell for describing the rights and obligations imposed by a contract. Arbitrarily complicated contracts can be built up using a small collection of primitives. Aggregations of these contracts can also be created, as can risks of default and bankruptcy. This created quite a stir in the quantitative analysis world when it was presented, as it was the first time anyone had proposed a formal language for describing contracts. Today the list of commercial Haskell users includes a number of financial institutions using this kind of technique to model their market positions.

But on its own this is just a faster and more efficient way of getting the same wrong answer. It doesn't solve the underlying problem of concealed systemic risk. The solution has to be for big financial companies to reveal their positions to the regulators as formal models of the contracts they have written. At the moment they don't even have to reveal all their contracts, but merely knowing the legal terms of a contract is only the first step. Those terms have to be converted into a mathematical model. That model probably already exists, but only as an internal artefact of the parties to the contract. What ought to be happening is that the contract is specified in a well-defined mathematical language that can be converted into a model automatically. If the regulators have this information about all the contracts entered into by all the finance companies then they can model the impact of, say, a downturn in the housing market or a jump in the price of oil, and if they see systemic risk looming then they can order the companies involved to take corrective action. Unlike the various Risk Management departments they will be able to see the whole picture, and they don't have to worry about being "team players".

ISPs and Bandwidth

This Slate article reminded me of some stuff about how the Internet works that isn't widely appreciated. For a few years now ISPs have been complaining about "bandwidth hogs" while at the same time advertising high bandwidths for fixed prices. Comcast has had its knuckles rapped for discriminating against particular traffic. But this hides the real issues, which are to do with the structure of the market for long-haul infrastructure.

Your ISP has to pay for connection to the Internet in exactly the same way that you do: it pays a big telecom company like AT&T or Qwest. The only exception is if your ISP is one of those companies, but even then the retail ISP business will be a separate division that has to "buy" bandwidth from its parent. These "tier 1 ISPs" don't publish price lists, but the general pricing structure looks a lot like the ones described in the Slate article, and for the same fundamental reason: capacity is limited. So your local ISP may buy, say, 1 terabyte per day upstream and 2 terabytes downstream, with any traffic over those thresholds being charged per megabyte.

Aside: Actually the whole thing is much more complicated: a retail ISP will generally buy connections to more than one upstream ISP, which may be either tier 1 or "tier 2" (with long-haul bandwidth but not global presence). Often it will have multiple connections to each of these ISPs, each of which may have different price plans. Any ISP may also have "peering" agreements with other ISPs of the same size. These are free, but are not allowed to carry "transit" traffic destined for anywhere else. Everyone always tries to offload traffic onto someone else as fast as possible, even if the resulting routes are not ideal. Managing this mess to keep the customers happy at minimum cost is a key skill in the ISP business

The retail ISPs are therefore caught between a rock and a hard place. They are in a commodity business, but the traditional retail price plan of "all you can eat at a given bandwidth" doesn't match their cost structure. Its a general rule that if you are in such a market and have a competitor who's price plan does match their cost structure then you are bound to make a loss, because the customers who find you cheaper are going to be the ones who cost you more than they pay, while the ones who would balance this by paying more than they cost find your competitor cheaper, so they go there instead.

Thus ISPs will gradually converge on pricing plans that are simplified versions of the cost structure of their industry. This will probably be based on a combination of peak-time limitations and traffic caps that give people an incentive to shift their heavy usage off-peak. The winners will be the ones who can innovate. The challenges are:

• At any given point in time the network has a fixed bandwidth. Hence the challenge is not to reduce the total amount of data moved but to even out usage. Similarly heavy users are only a problem when they start pushing out other customers who might collectively pay more.
  
• You can't force users to track their usage in detail. Price plans that suddenly cut a customer off until tomorrow (or next month) are scary and unfriendly. Plans that charge extra for heavy usage are even worse, especially for families with teenagers. Throttling is more user-friendly.
  
• Negotiate with your upstream ISPs to bring your costs into line with your pricing structure. They, too, need to shift usage off-peak and will be prepared to offer pricing plans accordingly. However their idea of "off-peak" may not be the same as a retail ISP with lots of home users.
• Transparency will happen whether you want it to or not. At least some end users are smart enough to detect traffic shaping and other tricks, and their results will be picked up by price-comparison sites for everyone else to read.
  

Application-specific traffic shaping (as tried by Comcast) won't work. Customers and regulators both hate it, but more importantly it gets you into an arms race between stealth P2P protocols and your packet inspection software that you can't win. However there is another option: offer your customers the opportunity to do their own traffic shaping. For instance you could have a metered high-priority service for everyday browsing combined with an unmetered low priority service for bulk downloads. The challenge is to give the customer a simple easy-to-use system that distinguishes between the two and (automatically as far as possible) uses the right one.

In many ways the situation reminds me of 1997, when large ISPs first started to deny peering agreements to smaller competitors and made them pay for transit agreements instead. There were many calls to ban the practice, and grave predictions of the "balkanisation of the Internet". But in practice the economics of Metcalf's Law guaranteed that a well-connected network would have more value than a disconnected network, and after that it was just a matter of how that extra value was distributed. I believe that much the same thing will happen with bandwidth. The Internet has most value when its pipes are full of traffic, and if there is demand for more bandwidth then there will be money to be made by providing it. After that, its just a matter of working out who pays how much for what. As long as the market remains competitive it will converge on the optimum solution, probably quite rapidly.