RiskMetrics and CreditMetrics: Industry Standard Risk Tools
We talked about Value at Risk (VaR) earlier in the book. You know the concept: estimate how much you can lose from your portfolio over a given time, with a given confidence level. Cool idea. But where do you get the actual numbers? Volatilities, correlations, credit data? Chapter 42 is about two systems that try to answer that question: RiskMetrics and CreditMetrics. Both came from JP Morgan, and both became industry standards.
RiskMetrics: VaR Data as a Service
In October 1994, JP Morgan launched RiskMetrics. Think of it as a data service. You have a portfolio, you know your deltas and gammas (your exposures). But you need volatilities and correlations for the underlying assets. RiskMetrics provides those.
The datasets cover almost 400 instruments: foreign exchange, bonds, swaps, commodities, equity indices. They give you term structure data across many currencies. There are three flavors: one-day horizon data, one-month horizon data, and data designed to meet Bank for International Settlements (BIS) requirements for internal risk models.
So the bank handles its own portfolio sensitivities, and RiskMetrics supplies the market parameters. Nice division of labor.
How RiskMetrics Measures Volatility
The simplest way to measure volatility is to look back over some period (say three months), calculate daily returns, and take the standard deviation. You get a number. Easy enough.
But this approach has two big problems.
First, how far back do you look? Three months? Six? What happened three months ago might be irrelevant today. But the more data you use, the smaller your sampling error. There is no perfect answer.
Second, there is the plateauing effect. Say one big move happens. That big return stays in your window for the entire three months. Then one day it drops out, and your measured volatility drops suddenly, even though nothing changed in the market. The volatility plot shows flat plateaus with sudden cliffs. It looks weird and it is misleading.
RiskMetrics fixes this with exponentially weighted moving averages (EWMA). Instead of giving equal weight to every day in the last three months, it gives more weight to recent days and less weight to older ones. The formula looks like this:
$$\sigma_i^2 = \lambda \sigma_{i-1}^2 + (1 - \lambda) R_i^2$$
Where $\lambda$ is a decay factor. JP Morgan chose $\lambda = 0.94$ for the one-day horizon and $\lambda = 0.97$ for one month. Each new return gradually fades in influence rather than suddenly falling off a cliff. No more plateauing. The resulting volatility series is much smoother and more responsive.
Correlation Estimation
Same story with correlations. The simple approach is equal-weighted covariance over some lookback window. Same plateauing problem. Same fix: use exponentially weighted averages.
$$\sigma_{12,i} = \lambda \sigma_{12,i-1} + (1 - \lambda) R_{1,i} R_{2,i}$$
But correlation estimation has its own special headache: time zones. Two assets might be perfectly correlated, but if they trade in different time zones, your daily data will show them as uncorrelated. You are measuring synchronicity of data collection, not synchronicity of actual price movements.
Also, there is no guarantee that your exponentially weighted covariance matrix is positive-definite. If it is not, your entire VaR calculation can produce nonsense. This is a known issue and something you need to watch for.
CreditMetrics: Measuring Credit Risk
RiskMetrics handles market risk. CreditMetrics handles credit risk. The risk that someone does not pay you back.
CreditMetrics was also proposed by JP Morgan. It has two main goals: create a benchmark for measuring credit risk, and increase market liquidity. The logic is simple. If you can measure credit risk systematically, people will understand risky bonds better. When people understand things better, they are less afraid to trade them. Less fear means more liquidity.
The CreditMetrics dataset has four components:
Yield curves. Risk-free yields for major currencies, at maturities from 1 to 30 years.
Spreads. For each credit rating (AAA, AA, A, BBB, etc.), the extra yield above the risk-free rate. Riskier bonds pay higher yields. A BBB bond yields more than an AA bond, which yields more than the risk-free rate. That extra yield compensates you for the chance of not getting your money back.
Transition matrices. What is the probability that a bond rated AA today will be rated AAA, A, BBB, or default in one year? The transition matrix gives you all those probabilities. Most of the time, a bond stays at its current rating. But there is always some chance of upgrade, downgrade, or default.
Correlations. Are defaults of different companies related? If company A goes under, does that make company B more likely to default? CreditMetrics decomposes each company’s stock returns into parts correlated with major market indices, plus a company-specific component. This lets you estimate correlations between any two companies in your portfolio.
How CreditMetrics Works
Let us walk through a concrete example. You own a zero-coupon bond rated AA, maturing in three years. You want to know what this bond could be worth in one year.
Three things affect the value: time passing, interest rate changes, and potential rerating.
Time effect. In one year, your three-year bond becomes a two-year bond. Simple enough.
Interest rate assumption. CreditMetrics assumes forward rates do not change. You calculate the forward rates from today’s yield curve and use those. This is called “rolling down the curve.”
Credit rating changes. The transition matrix tells you the probability of each future rating. Say there is an 87% chance the bond stays AA. There is a 5.5% chance it gets upgraded to AAA. There are smaller probabilities for each downgrade.
For each possible rating, you know the corresponding spread, so you can calculate the bond’s value. Combine all these with their probabilities and you get a distribution of possible future values.
Here is the key insight: this distribution is highly skewed. Most of the time the bond is worth about the same. But there is a small tail where it defaults and you lose big. This is very different from the Normal distribution you might assume in simple VaR. The distribution is far from bell-shaped, so classical portfolio theory must be used with care.
Portfolios of Risky Bonds
One bond is straightforward. A portfolio of risky bonds is harder. With two bonds, each of which can be in eight states (AAA through Default), you have 64 possible joint states. You need to know the probability of each.
This is where the correlation data comes in. You decompose each company’s stock returns into index-related and company-specific components. From there you can estimate the correlation between any two issuers and calculate the probability of any joint state.
The output of CreditMetrics for a portfolio gives you risk measured by standard deviation. It is a relative measure. Because credit risk distributions are so skewed, standard deviation is not a perfect absolute risk measure. But it works well for comparing the relative riskiness of different instruments or portfolios.
The Bigger Picture
RiskMetrics and CreditMetrics are complementary pieces of the VaR puzzle. RiskMetrics handles day-to-day market fluctuations. CreditMetrics handles the risk that your counterparties go bust.
Wilmott makes a good point here: with something as important as Value at Risk, you should explore every available measure. No single methodology captures all the dangers in your portfolio. Use multiple tools, understand their limitations, and build the most complete picture you can.
In the next chapter, we look at what happens when the market does not just fluctuate but completely crashes.
Previous: Credit Derivatives: CDS and CDO