What Have We Learned?
Ten technologists discuss what the events of the past few months have taught us about risk management as it is practiced today.
Last year’s financial crisis in Asia was quickly followed by another in Russia, which was followed by serious losses at almost every major world bank. All this raises a question: Wasn’t risk management supposed to prepare us for all this? Some people have argued recently that most risk models completely missed the most important market phenomena. For example, we learned that if everybody is chasing the same carry trade and hedging or diversifying it in the same fashion, the hedge will fall apart when everybody rushes to the exits. How can models account for this kind of phenomenon? What does our recent experience teach us about the nature of correlation? And how can credit models be improved in light of recent experiences in Asia and Russia, and in the hedge fund world?
Correlation assumptions are the forgotten stepchildren of risk calculations. People focus a lot on whether volatility should be scaled up to two, three or five standard deviations. But stressing correlation assumptions is as important as stressing the size of shocks in volatility assumptions. In some cases, recent correlations between markets have turned out to be dramatically different from anything we’ve seen before.
This is particularly important in spread or arbitrage trading because that, by definition, is a correlation-based strategy. In the case of Long-Term Capital Management, it was convergence trades. Correlation assumptions are the first things you should stress-test if you’re betting on mean reversion in spread trades. The key point is that these trades assume that the underlying distributions of asset prices are stable, and that deviations of spreads from their historical means are “outliers” in a stable distribution. In times of structural shifts in markets, however, these distributions can actually change (either permanently or for a long time), so that what appeared to be an outlier is in fact the mean of a new distribution. Relative value trading runs into big trouble here.
One of the most difficult things in stress testing is getting people to suspend incredulity. If you come up with a scenario that’s relatively extreme, senior managers can say: “We can run that scenario but it won’t happen.” As a result, you attach such a low probability to it that you discount it and don’t take any action. So it becomes a meaningless exercise. You don’t reduce positions, you don’t hedge; you just run this scenario because somebody told you to.
You can believe that spreads will converge over a longer time frame, but if that’s outside the time frame of market patience, it doesn’t do you any good. LTCM thought it had taken care of that because it had locked in its investors. But it couldn’t lock in its creditors. One could argue that even Drexel Burnham was a viable concern when it went under. If creditors aren’t going to be patient, the liquidity issue bites with a vengeance.
Market risk measurement is based largely on volatility and correlation forecasts, which in turn are based largely on historical experience. This is why people say that objective risk measures such as value-at-risk measure only “normal” market risk. It turns out, for example, that in most markets only about 60 percent of the volatility variations are explained by the best possible volatility forecasts available, leaving substantial room for “abnormal,” or unanticipated, risks. Does this mean, as some people have concluded, that we should throw out VAR and all the other objective risk measurement techniques because they are imperfect? (Weather forecasts are quite imperfect as well, but they still seem to be useful.) We need to learn to accept residual uncertainties in our risk management systems, while striving continually to improve them.
Correlations can be quite unstable over time and changing market conditions. I remember being in a trading room during the crash of 1987, when the Australian equities market was about to open, after New York had closed. Consider the correlation estimate for the All-Ordinaries index vs. the Dow Jones Industrial Average on that day, compared with the same correlation on the previous day: Everyone knew at that point in time, of course, that the correlation was about 1.0.
The tendency for correlations to converge toward unity when there is a market break is one reason why it’s important that VAR software show the user both diversified and nondiversified VAR numbers. The diversified risk number is for the “normal” correlations, but it is the nondiversified risk number that is important in many types of market disasters, because correlations tend to move toward 1.0, usually because all markets are simultaneously affected by a single investor-confidence factor.
I think hedge funds should be subject to risk-auditing requirements. The simple disclosure of numbers such as VAR is not quite enough, because it essentially puts the fox in charge of the henhouse. There should be some additional assurances that the VAR numbers being generated and shown to investors have accurately depicted risks in the past and can be expected to continue doing so in the future. The disclosure should also state that the risk systems in use meet the industry standards in capturing the risks of the instruments the hedge fund intends to trade.
president and CEO
The recent problems show how young risk management and financial engineering really are. In the civil engineering profession, after the Tacoma Bridge falls down, everybody analyzes what went wrong and all that gets integrated into the civil engineers’ building manual and it never happens again. But in the risk management profession the things that happen to Metallgesellschaft and Barings happen again and again.
One message to risk managers resounds after each derivatiaves blowup: Learn how to ask questions. If you’re brought a risk report that says your VAR is $18 million, you should ask: “What scenarios were used to create that VAR? Did you look at a scenario in which all correlations break down?” This is not a new insight. For years, people have known that when there’s a crash, there’s a flight to quality or a herd instinct, which is another way of saying that all correlations become 100 percent.
Recent events have also shown that people who measure credit risk separately from market risk are completely wasting their time, because the two are intimately linked. You saw it recently in Korea. The biggest effect of the market devaluation was its enormous impact on credit exposures. You wouldn’t have had the same credit exposures if you hadn’t had the devaluation. But if you went to a Korean bank and asked, “Does your market risk department talk to your credit risk department?” the answer would be no.
Part of the problem is that people schooled in finance and financial theory have not taken risk management seriously enough. They’ve looked at risks of individual securities, but not at how to measure the risk of a bank.
Mamdouh Barakat Risk Management
The increased liquidity in the market and the increase in unconstrained capital flows make the chances of a catastrophic event more likely, because the flow of funds is so large that it can overwhelm the economic fundamentals. A lot of the recent big movements have been one-way, whether one-way down or one-way up. The foreign exchange market, for example, was originally a place for transactions based on economic objectives of exporters and importers. There was usually some strong economic rationale for market movements—for instance, that one country’s balance of trade was improving. But all that has now been overshadowed by unconstrained capital flows (or hot money) in and out of markets.
A lot of hedge funds seem to be making the same bets. To a certain extent, it’s a self-fulfilling prophecy. If a large number of hedge funds decide a country is going to devalue its currency, by virtue of the amount of capital they have at their disposal (magnified with leverage), the currency is likely to devalue. But if the hedge funds have to unwind their positions quickly (because of a credit crunch, for instance), then it becomes a self-fulfilling prophecy in the opposite direction, because the market prices, assisted by rumors, would move against a large forced seller.
Some models look at a series of bonds and project the underlying economic expected return on a trade. Many people will not buy a particular bond that the model tells them is cheap because they’re afraid they will not be able to sell it easily as a result of a lack of liquidity. If you buy a bond simply because the model says it appears cheap relative to other bonds, there is no guarantee that tomorrow it will not be even cheaper. This would not matter much if you are able (and happy) to hold onto the bond until maturity. However, if you have to meet margin calls, since you may be using the “cheap” bond as collateral, or simultaneously shorting a “bearer” bond, this could wipe out the maximum capital you wish to put into the trade by triggering a stop-loss reversal of the position, thus crystallizing your losses.
Hitting a stop-loss does not make the model wrong or faulty. The model would have done its job in telling you the bond is relatively cheap—but acting on it without evaluating the possible reasons for the different relative values would be a mistake. These market price anomalies could exist for long periods of time, and anomalies like these may even accentuate market behavior. If a bond looks cheap because it is rumored to be illiquid, for instance, then new investors could buy the more liquid “expensive” bond, making the expensive bond even more expensive.
Models, after all, are just tools. If a speedometer shows you are doing 40 miles per hour, that in itself does not indicate that you are too fast or too slow. You need to know what is happening in the street around you.
The standard stress test is based on statistical phenomena and works well most of the time. But in situations such as capital flight, you must use a variety of different methods so you don’t miss anything. That involves more than statistical measurement and moving parameters around—it means trying out specific scenarios.
You have to think about what kind of market distortions can take place in times of stress. Weaker markets can weaken and stronger markets can strengthen. The flight out of Russian debt even affected convergence trades on euro currencies, where spreads widened because people were buying Deutsche marks instead of pesetas.
For example, if you buy emerging market debt and you’re shorting U.S. Treasuries as a hedge against global interest rates, you have to ask, “What will the effect be on my emerging markets portfolio—along with the Treasury hedge—if there is a flight to quality?”
One of the standard stress tests we show prospective clients is equity volatility during the 1990–91 Persian Gulf crisis. Last year’s volatility totally eclipsed the volatility we saw then. Who thought this would happen? But you can’t live your life in only worst-case scenarios. You just have to make sure you can survive them.
senior vice president and COO of the risk management division
There is a story in the Bible about some people who wanted to build a tower reaching into the heavens, hence challenging God’s power. The obsession with modeling and the power given to models is the modern version of the Tower of Babel.
Financial institutions attributed excessive power to models and forgot a simple fact: Models are simplifications of reality and as such can only approximate it. This mistake was probably compounded by regulators putting too much emphasis on risk models.
The market believed in these very same models. It forgot that, ultimately, the marketplace is composed of human beings acting in an irrational manner that cannot be modeled. It’s too simple to say that the answer to model risk is good stress testing. If you apply extreme stress testing to determine capital charges for your institution, you will most likely reduce the attractiveness of most business lines.
A lot of people are suggesting that more stress testing will be able to prevent all this. Stress testing should serve as a way to help investors make informed decisions. We shouldn’t consider it normal for public firms such as Merrill Lynch or Credit Suisse First Boston to suddenly disclose massive losses. It all stems from a lack of transparency. Ultimately, what matters is improved disclosure vis-à-vis the investor. Let’s have better financial disclosure so investors can make more informed decisions. Instead of merely requiring firms to show VAR in their quarterly statements, they should be required to disclose their exposures to market events that adversely affect their holdings.
The LTCM crisis does not demonstrate the shortcomings of risk management as it is practiced today. Rather, it was the misuse of available risk management tools that led to the losses. There may have been an attitude of invincibility at LTCM that developed along the way and exacerbated the situation, but there was certainly an awareness of the risk. The risk management gurus at LTCM could have simply plugged in 30 years’ worth of market observations to catch a glimpse of the worst-case scenario if market conditions turned against them and the spreads deviated even further. This type of scenario analysis is standard functionality in most top-tier mission-critical trading systems available today.
In the derivatives industry, hedging is not an exact science. A hedge needs to be evaluated continually to ensure that it is doing the job it was intended to do—namely, reducing risk within the bounds of a firm’s or trader’s risk appetite.
Model risk is an entirely different issue. If LTCM was using models that were not properly calibrated, or if it was engaging in deals that were so illiquid that model calibration was not a possibility, then the real fault lies with LTCM’s risk management policies and not the market’s standard approach to risk management. All firms need to evaluate liquidity and compensate for the potential lack thereof when contemplating the exit strategy of their positions.
Financial institutions will have the information and tools necessary to reduce the risk profile of their portfolios dramatically, if correlation is properly measured on a current basis, if the financial instruments involved are understood, and if a market-accepted and market-tested VAR calculation is employed. Correlations tend to break down at the wrong time—usually when the markets are hectic and moving dramatically. This is why no firm can simply look at a single VAR number, but has to do many simulations with different correlations based on what could happen under numerous concurrent or sequential scenarios.
A standardized approach to disclosure by financial institutions would go a long way toward improving the evaluation of credit worthiness. As far as modeling is concerned, stricter limits and more realistic scenario analysis are key to making the most of the sophisticated models available today.
manage and quantitative analyst
There was clearly a lot more correlation between institutions and countries than participants expected. The models did not account for that. The dimensions of the problem are much broader than traditional models have accounted for, and, as a result, not everything is accounted for. If a model is good, it will capture most of the problem, but it will never capture the whole problem.
It is particularly important to understand how correlations between components affect market volatility, a catch-all number designed to reflect uncertainty. When market components are weakly correlated, movements between these components tend to offset each other to some degree. When markets move, however, these components become strongly correlated, causing the market volatility to increase significantly.
Another problem is that, frequently, the people who construct the models are not the ones who actually use them. The senior managers who interpret these numbers may not be familiar with the limitations and assumptions that go into the model.
At LTCM and elsewhere, people might have become complacent. Their success may have blinded them to the actual risks they were taking on. They were only too happy to be realizing the returns.
There are certain inherent limitations in many statistical or parametric risk management models. For example, JP Morgan’s RiskMetrics uses a U.S. Treasury curve and a swap curve, but it doesn’t capture curves from other significant markets, such as mortgage-backed, agency, commercial and many others. As a result, you lose the resolution in these market segments—or else you’d have to create a gigantic correlation matrix (for these markets)—and that would be too expensive and ultimately impractical.
We have to remember that VAR is a snapshot of a frozen moment in time, while risk management is the movie. You need to supplement VAR with all sorts of business judgments. Translating some of these business developments into a scenario analysis is the best strategy to fill in the gaps, but many systems are not good at allowing users to define and transform their market views. In some cases, the market will gap in a way that goes beyond a simple shock in market prices. It is not just a matter of changing bid and offer prices. A particular scenario may, in fact, change the entire nature of the market. Longer-term transactions on benchmark securities, for example, may not be available for you to unload. You have to be able to model a market taking into consideration what happens if nobody is willing to buy or sell at a particular maturity.
Axiom Software Laboratories
We need to take another look at methods used to calculate portfolio risk. One of the main criteria has been counterparty ratings, but in the LTCM case nobody paid specific attention to how that counterparty organized its portfolio, or how big the leverage was in the portfolio and where the money had gone.
We should also take another look at correlation—not as it is understood in a statistical mathematical sense, but in terms of a rare confluence of events. You can’t estimate correlation for these events because you don’t have enough data. But there are several mathematical distributions that allow you to model events that have low probabilities under normal distributions. Using these new techniques, you can find events beyond three, four or five standard deviations. These distributions should not be applied every time but to relatively rare events, such as modeling potential political crises in Russia. To do this, you need to be able to identify the types of risk factors for which the rare-event methodology occurs and distinguish these factors from other risk factors where you might apply other—or even normal—distributions. A lot also depends on the flexibility of model. You must be able to manipulate the data and link your portfolio to the history of risk factors and be able to distinguish how these risk factors should be treated.
|Was this information valuable?|
Subscribe to Derivatives Strategy by clicking here!