Building Credit Risk Economic Capital in Practice

Executive Insight

Credit risk modeling promises that diversification reduces portfolio loss. Yet the relationship between diversification and risk is profoundly non-linear: adding a second B-rated bond to a single A-rated bond can increase portfolio Value-at-Risk at the 99.7% confidence level, even though expected loss falls. This paradox—documented in the source case study with Moody’s 1920–2012 transition data—is the central puzzle this paper resolves.

Using calibrated default probabilities, recovery rates, and credit migration matrices from Moody’s Investors Service, we construct a complete economic capital framework from first principles. The analysis demonstrates that a portfolio of 20 Ba-rated bonds achieves a maximum loss of 3.4% at the 0.3% probability threshold, compared to 49.7% for a single Baa-rated bond—a 93% reduction in tail risk. But the path from 1 bond to 20 bonds is non-monotone, and the turning point depends critically on the correlation structure.

Expected loss tells you what to provision. Unexpected loss tells you how much capital to hold. The gap between them is economic capital—and it is driven by correlation, not by default probability.

The Diversification Paradox

Consider two hypothetical bonds:

• Bond X (A-rated): probability of default $\text{PD} = 33.3\%$, loss given default $\text{LGD} = 100\%$
• Bond Y (B-rated): probability of default $\text{PD} = 50.0\%$, loss given default $\text{LGD} = 100\%$

A portfolio holding one unit of Bond X has expected loss $\text{EL} = 0.333$ and maximum loss 1.0 (total default). Adding one unit of Bond Y creates a portfolio with $\text{EL} = 0.833$ but maximum loss 2.0—and the probability of maximum loss is $0.333 \times 0.500 = 16.7\%$ under independence.

The paradox emerges when we examine the loss distribution at a fixed confidence level. At the 0.3% threshold (Basel II/III standard), the single A-rated bond has zero incremental risk (its PD exceeds the threshold). But a concentrated portfolio of 2–5 B-rated names can exhibit higher tail risk than the single A-rated bond due to correlation-amplified joint default.

This result is not academic curiosity—it drives real capital allocation decisions. The resolution requires moving from individual-bond PD/LGD analysis to full portfolio credit risk modeling.

Moody’s Transition Data: 1920–2012

The empirical foundation for the analysis is Moody’s historical credit transition matrix, covering 92 years of corporate bond ratings from Aaa to C, plus default. Key diagonal entries (probability of remaining in the same rating class over one year):

The critical parameter for economic capital is the default transition probability. For Ba-rated issuers, the 1-year default probability is 1.268%—not negligible, but sufficiently low that a diversified portfolio of 10–20 names provides substantial tail-risk reduction.

Recovery rates by seniority, from Moody’s recovery studies:

The spread between secured loans (80.6% recovery) and subordinated bonds (28.5% recovery) implies that LGD assumptions alone can swing economic capital estimates by a factor of 3.7×.

Expected Loss and the EL/UL Decomposition

The fundamental decomposition of credit risk separates expected loss (provisioned from revenue) from unexpected loss (absorbed by capital):

For a Baa-rated senior unsecured bond with $\text{PD} = 0.189\%$, $\text{LGD} = 51.4\%$, and $\text{EAD} = 100$:

$\text{EL} = 0.00189 \times 0.514 \times 100 = 0.0972$

Expected loss is a cost of doing business—it is priced into the credit spread and provisioned. Economic capital covers the unexpected component:

where $\alpha$ is the confidence level (typically 99.9% for Basel II IRB, 99.7% for internal ICAAP). The challenge is computing $\text{VaR}_{\alpha}(L)$ for a correlated portfolio.

VaR Analysis: From Individual Bonds to Portfolios

The source case study computes maximum loss at the 0.3% probability level for increasingly diversified portfolios. The results resolve the diversification paradox:

Credit Portfolio Models

Three modeling frameworks are integrated in the analysis:

Merton Structural Model

Default occurs when asset value $V_T$ falls below the debt barrier $D$ at maturity. Under geometric Brownian motion:

The distance-to-default $\text{DD} = [\ln(V_0/D) + (\mu - \sigma^2/2)T]/(\sigma\sqrt{T})$ is the key risk metric, mapping equity market data to credit risk through the firm’s capital structure.

Vasicek One-Factor Model

For portfolio-level risk, the Vasicek framework introduces a single systematic factor $Z$:

where $\rho$ is the asset correlation parameter. Under worst-case $Z$ (stressed macro state), the conditional PD can be orders of magnitude higher than the unconditional PD. This is the mechanism by which correlation drives economic capital.

CDS Fair Spread and CVA

The market-implied default probability can be extracted from credit default swap spreads. The fair CDS spread $s$ satisfies:

$s = \text{PD} \times \text{LGD} \times \text{(duration adjustment)}$

Credit Value Adjustment (CVA) extends this to bilateral counterparty risk in OTC derivatives portfolios, linking credit risk modeling to derivatives pricing.

Rating-Dependent Yield Curves

The source data provides interest rate curves by rating class that underpin mark-to-market credit risk calculation. For credit migration from Baa to Ba, the bond is repriced at the Ba yield curve—generating a mark-to-market loss even without default. This migration-mode VaR is typically 2–3× larger than default-mode VaR for investment-grade portfolios.

A downgrade from Baa to Ba widens the 10Y spread by 200 bps, generating a mark-to-market loss of approximately 14% on a 10-year bond. This loss occurs without default and is the dominant risk mode for investment-grade credit portfolios.

Institutional Implications

For Chief Risk Officers

• Diversification is necessary but non-linear: The jump from 1 to 10 names reduces tail risk by 93%; the jump from 10 to 20 reduces it by only an additional 6%. Marginal diversification benefit declines rapidly, and the correlation parameter $\rho$ determines the asymptotic floor.
• Recovery rate assumptions matter enormously: Economic capital for a subordinated portfolio ($\text{LGD} = 71.5\%$) is 3.7× higher than for a secured loan portfolio ($\text{LGD} = 19.4\%$). Model validation must independently stress-test recovery assumptions.

For Portfolio Managers

• Migration risk dominates default risk: For investment-grade portfolios, mark-to-market losses from rating downgrades exceed expected default losses by a factor of 2–3×. Economic capital models using default-mode only will systematically understate risk.
• Concentration limits: The VaR table shows that a single-name Baa exposure has 49.7% tail risk vs. 3.4% for a 20-name Ba portfolio. Name and sector concentration limits should be calibrated from the VaR table, not from notional exposure caps.

For Regulators

• The Vasicek one-factor model, which underpins Basel II/III IRB capital formulas, captures systematic risk but not sector concentration. Pillar 2 (ICAAP) economic capital should supplement Pillar 1 regulatory capital with multi-factor models that capture industry and geographic concentration.

Methodology & Academic Foundation

The analysis draws on the following academic and institutional sources:

• Moody’s Investors Service: Historical credit transition matrices (1920–2012); recovery rate studies by seniority class; annual default studies
• Merton, R. C. (1974, Journal of Finance): structural model of default; distance-to-default framework linking equity prices to credit risk
• Vasicek, O. (1987, 2002): one-factor Gaussian copula model; the asymptotic single risk factor (ASRF) model underpinning Basel II IRB formulas
• CreditMetrics (J.P. Morgan, 1997): credit migration framework for mark-to-market credit VaR; the industry standard for portfolio credit risk
• Basel Committee on Banking Supervision: IRB approach (2004, revised 2017); Pillar 2 ICAAP requirements for economic capital
• Schönbucher, P. (2003, Credit Derivatives Pricing Models): CDS fair spread derivation and CVA framework

All numerical examples use Moody’s published data. Portfolio VaR calculations assume conditional independence given the systematic factor (Vasicek model). No proprietary data or models are required—the analysis is fully reproducible from public sources.