Chapter 22: Risk Management¶
This chapter develops the mathematical foundations of modern risk management, from static and dynamic risk measures through counterparty credit risk and valuation adjustments to systemic risk and model governance. Starting from the axiomatic theory of coherent and convex risk measures, we build the full risk management stack: VaR and Expected Shortfall for market risk, spectral risk measures and Kusuoka representations for generalized tail weighting, conditional and BSDE-based frameworks for dynamic risk assessment, exposure profiles and XVA for counterparty risk, stress testing and extreme value theory for tail events, funding and liquidity risk including optimal execution, network models for systemic contagion and central clearing, and the backtesting, validation, and governance frameworks that ensure model integrity under regulatory requirements.
Key Concepts¶
Market Risk Measures¶
Value-at-Risk (VaR) is the \(\alpha\)-quantile of the loss distribution: \(\text{VaR}_\alpha(L) = F_L^{-1}(\alpha)\). Under normality, \(\text{VaR}_\alpha = -\mu + \sigma\Phi^{-1}(\alpha)\); the variance-covariance method gives \(\text{VaR}_\alpha = -\mathbf{w}^\top\boldsymbol{\mu} + \sqrt{\mathbf{w}^\top\boldsymbol{\Sigma}\mathbf{w}}\,\Phi^{-1}(\alpha)\) for a portfolio with weight vector \(\mathbf{w}\). Historical simulation estimates \(\widehat{\text{VaR}}_\alpha = L_{(\lceil n\alpha\rceil)}\) from the empirical loss distribution; Monte Carlo simulation generates scenarios from a specified model and takes the empirical quantile. VaR scales with the square-root-of-time rule \(\text{VaR}_\alpha(h) \approx \sqrt{h}\,\text{VaR}_\alpha(1)\) under i.i.d. returns. VaR is elicitable (via the pinball loss \(S(x,L) = (\mathbf{1}_{L \le x} - \alpha)(x - L)\)) but fails subadditivity: diversified portfolios can have higher VaR than the sum of components, violating the diversification principle. Expected Shortfall corrects this as the average of tail VaRs: \(\text{ES}_\alpha(L) = \frac{1}{1-\alpha}\int_\alpha^1 \text{VaR}_u(L)\,du = \mathbb{E}[L \mid L \geq \text{VaR}_\alpha]\), which under normality becomes \(\text{ES}_\alpha = \mu + \sigma\,\phi(\Phi^{-1}(\alpha))/(1-\alpha)\). The Rockafellar--Uryasev optimization formulation \(\text{ES}_\alpha = \min_v\{v + \frac{1}{1-\alpha}\mathbb{E}[(L-v)^+]\}\) enables convex optimization of ES-constrained portfolios. ES satisfies all four coherence axioms (Artzner et al.): monotonicity, translation invariance, positive homogeneity, and subadditivity. The dual representation theorem establishes that a risk measure \(\rho\) is coherent if and only if \(\rho(L) = \sup_{\mathbb{Q} \in \mathcal{Q}}\mathbb{E}^{\mathbb{Q}}[L]\) for some set of probability measures \(\mathcal{Q}\); for ES, the densities satisfy \(d\mathbb{Q}/d\mathbb{P} \leq 1/(1-\alpha)\). Convex risk measures relax homogeneity, with the entropic risk measure \(\rho_\gamma(L) = \frac{1}{\gamma}\log\mathbb{E}[e^{\gamma L}]\) as the canonical example admitting a penalty-based dual representation \(\rho(L) = \sup_{\mathbb{Q}}\{\mathbb{E}^{\mathbb{Q}}[L] - \alpha(\mathbb{Q})\}\). Spectral risk measures \(\rho_\phi(L) = \int_0^1 \phi(u)\,\text{VaR}_u(L)\,du\) generalize ES through a non-decreasing weight function \(\phi\) encoding risk aversion: ES corresponds to \(\phi(u) = \frac{1}{1-\alpha}\mathbf{1}_{u \ge \alpha}\), while an exponential spectrum \(\phi(u) = \gamma e^{\gamma u}/(e^\gamma - 1)\) provides increasing weight on worse outcomes. The Kusuoka representation \(\rho(L) = \sup_\mu \int_0^1 \text{ES}_u(L)\,d\mu(u)\) characterizes all law-invariant coherent measures, showing ES is the fundamental building block. Euler capital allocation \(\text{AC}_i = \mathbb{E}[L_i \mid L \geq \text{VaR}_\alpha]\) decomposes ES into additive risk contributions satisfying \(\sum_i \text{AC}_i = \rho(L)\). Under Basel III/IV FRTB, market risk capital is based on \(\text{ES}_{0.975}\) rather than VaR, with capital \(= \max(\text{ES}_{t-1},\, m_c \cdot \overline{\text{ES}}) + \text{SES}\).
Dynamic Risk Measures¶
Static risk measures ignore the temporal structure of risk. A conditional risk measure \(\rho_t : L^\infty(\mathcal{F}_T) \to L^\infty(\mathcal{F}_t)\) maps future losses to \(\mathcal{F}_t\)-measurable risk assessments, with conditional versions of the coherence axioms: conditional translation invariance uses \(\mathcal{F}_t\)-measurable shifts (\(\rho_t(X + m) = \rho_t(X) + m\) for \(\mathcal{F}_t\)-measurable \(m\)), conditional positive homogeneity uses \(\mathcal{F}_t\)-measurable \(\lambda > 0\), and conditional subadditivity holds almost surely. Conditional ES \(\text{ES}_\alpha^t(X) = \frac{1}{1-\alpha}\int_\alpha^1 \text{VaR}_u^t(X)\,du\) is conditionally coherent; conditional VaR generally is not. The dual representation becomes \(\rho_t(X) = \operatorname{ess\,sup}_{\mathbb{Q} \in \mathcal{Q}_t}\mathbb{E}^{\mathbb{Q}}[X \mid \mathcal{F}_t]\). Monetary risk measures satisfy monotonicity and translation invariance as the minimal axiom set, forming the foundation for both coherent and convex dynamic extensions. Time-consistency requires that risk rankings never reverse: \(\rho_t(X) \leq \rho_t(Y)\) a.s. implies \(\rho_s(X) \leq \rho_s(Y)\) a.s. for \(s < t\), equivalently captured by the recursive (tower) property \(\rho_s(X) = \rho_s(-\rho_t(X))\), which is analogous to the tower property of conditional expectation and equivalent to the Bellman principle in dynamic optimization. Dynamic VaR is not time-consistent; dynamic ES can be made time-consistent through recursive construction \(\rho_s(X) = \text{ES}_\alpha(-\rho_t(X) \mid \mathcal{F}_s)\). BSDE-based risk measures provide the canonical time-consistent framework: given
the solution \(\rho_t(X) := Y_t\) defines a \(g\)-expectation (Peng) that automatically satisfies time-consistency from the BSDE flow property. The driver \(g\) determines the risk measure type: \(g = 0\) gives conditional expectation; \(g = \gamma|z|^2/2\) gives the entropic risk measure \(\frac{1}{\gamma}\log\mathbb{E}[e^{\gamma X} \mid \mathcal{F}_t]\); convex \(g\) in \(z\) gives convex risk measures; positively homogeneous \(g\) gives coherent measures. Existence and uniqueness follow from Pardoux--Peng theory under Lipschitz and growth conditions on \(g\). The comparison theorem ensures monotonicity, and the nonlinear Feynman--Kac formula \(\partial_t u + \mathcal{L}u + g(t,u,\sigma^\top\nabla u) = 0\) connects BSDE risk measures to semilinear PDEs, with direct applications to XVA pricing and dynamic portfolio optimization.
Stress Testing and Scenario Analysis¶
Statistical risk measures (VaR, ES) are backward-looking; stress testing is forward-looking, evaluating portfolio losses under hypothetical extreme scenarios. Historical simulation uses past returns \(L_i = -P_0(e^{\mathbf{w}^\top\mathbf{r}_i} - 1)\); filtered historical simulation standardizes by conditional volatility \(\epsilon_t = r_t/\sigma_t\) via GARCH to adapt to the current volatility regime. Extreme scenarios must be severe, plausible, and internally consistent: stress loss uses either the sensitivity approximation \(L^{\text{stress}} \approx -\sum_i \Delta_i\,\Delta x_i - \frac{1}{2}\sum_{i,j}\Gamma_{ij}\,\Delta x_i\,\Delta x_j - \sum_i \mathcal{V}_i\,\Delta\sigma_i\) or full revaluation \(L^{\text{stress}} = P_0 - P(\mathbf{x}^{\text{stress}})\). The Mahalanobis distance \(D(\mathbf{x}) = \sqrt{(\mathbf{x}-\boldsymbol{\mu})^\top\boldsymbol{\Sigma}^{-1}(\mathbf{x}-\boldsymbol{\mu})}\) measures scenario plausibility, enabling maximum-loss optimization \(\mathbf{x}^* = \arg\max L(\mathbf{x})\) subject to \(D(\mathbf{x}) \leq D_{\text{target}}\). Reverse stress testing inverts the problem: given a failure threshold \(L^*\), find the most plausible scenario causing it via \(\mathbf{x}^* = \arg\min D(\mathbf{x})\) subject to \(L(\mathbf{x}) \geq L^*\); the Lagrangian first-order condition \(\nabla D(\mathbf{x}) = \lambda\nabla L(\mathbf{x})\) aligns the gradient of plausibility with the gradient of loss, and for linear portfolios the minimum-plausibility failure direction is \(\mathbf{x}^* = \mathbf{x}_0 + \frac{L^*}{\boldsymbol{\delta}^\top\boldsymbol{\Sigma}\boldsymbol{\delta}}\boldsymbol{\Sigma}\boldsymbol{\delta}\). Stressed correlations \(\boldsymbol{\rho}^{\text{stress}} = \alpha\boldsymbol{\rho}^{\text{normal}} + (1-\alpha)\mathbf{1}\mathbf{1}^\top\) capture the breakdown of diversification under crisis conditions. Extreme value theory provides the distributional foundation for modeling tail behavior beyond historical observation: the generalized extreme value distribution and Peaks-over-Threshold methods using the generalized Pareto distribution extend tail modeling. Regulatory stress tests (CCAR/DFAST, EBA adverse scenarios) prescribe macro-variable paths and require capital adequacy demonstration under multi-year stress horizons, complementing internal stress testing with supervisory scenarios.
Counterparty Credit Risk¶
Counterparty exposure is one-sided: \(E_t = \max(V_t, 0) = V_t^+\). The expected exposure \(\text{EE}(t) = \mathbb{E}^{\mathbb{Q}}[V_t^+]\) and expected positive exposure \(\text{EPE} = \frac{1}{T}\int_0^T \text{EE}(t)\,dt\) summarize the exposure profile, with the regulatory effective EPE using \(\text{Effective EE}(t) = \max_{s \le t}\text{EE}(s)\) to prevent gaming. Potential future exposure \(\text{PFE}_\alpha(t) = F_{E_t}^{-1}(\alpha)\) is the tail quantile, with the ratio \(\text{PFE}/\text{EE}\) typically 2--4\(\times\); PFE is non-additive across portfolios (\(\text{PFE}_\alpha(A+B) \le \text{PFE}_\alpha(A) + \text{PFE}_\alpha(B)\)) and is used for credit limit setting via the peak PFE \(\max_t \text{PFE}_\alpha(t)\). Product-specific profiles vary: interest rate swaps exhibit a hump shape peaking at \(t \approx T/2\); FX forwards scale as \(\sigma_{\text{FX}}\sqrt{t}\) with maximum at maturity; options show buyer/seller asymmetry. Netting reduces exposure: \(E_t^{\text{netted}} = (\sum_i V_{i,t})^+ \leq \sum_i V_{i,t}^+\). Collateral with margin period of risk (MPOR, typically 10 days bilateral, 5 days CCP) gives \(\text{EE}^{\text{coll}}(t) = \mathbb{E}[\max(V_t - C_{t-\text{MPOR}}, 0)]\). Wrong-way risk (WWR)--positive correlation between exposure and default--amplifies CVA beyond the independence assumption: \(\text{CVA}^{\text{WWR}} = \text{LGD}\int_0^T \mathbb{E}[E_t \mid \tau = t]\,dPD(t) > \text{CVA}\), with specific WWR arising from direct transaction-counterparty links (e.g., put options on counterparty stock) and general WWR from systemic correlations (e.g., FX derivatives with EM counterparties). Intensity-based models \(\lambda_t = \lambda_0 e^{\beta X_t}\) capture the exposure-default dependence explicitly, with the CVA multiplier \(\text{CVA}^{\text{WWR}}/\text{CVA}\) typically ranging from 1.2--3.0\(\times\) depending on severity.
Valuation Adjustments (XVA)¶
Classical risk-neutral pricing \(V_0 = \mathbb{E}^{\mathbb{Q}}[D(0,T)\cdot\text{Payoff}]\) assumes perfect collateralization and no counterparty risk. The XVA framework corrects this: \(V^{\text{total}} = V^{\text{risk-free}} - \text{CVA} + \text{DVA} - \text{FVA} - \text{KVA} - \text{MVA}\). CVA \(= \text{LGD}\int_0^T \text{EE}(t)\,\lambda(t)\,S(t)\,D(0,t)\,dt\) prices counterparty default risk, with the discrete approximation \(\text{CVA} \approx \text{LGD}\sum_i \text{EE}(t_i)[PD(t_i) - PD(t_{i-1})]D(0,t_i)\); CVA Greeks include credit spread sensitivity \(\partial\text{CVA}/\partial s\) and market factor sensitivities \(\partial\text{CVA}/\partial S\), enabling hedging via CDS and underlying instruments. DVA symmetrically accounts for own default using negative expected exposure \(\text{NEE}(t) = \mathbb{E}[V_t^-]\), creating the controversial "profit from own credit deterioration" (required by IFRS 13 but excluded from regulatory capital). FVA \(= \int_0^T s_F(t)\,\mathbb{E}[(V_t - C_t)^+]\,D(0,t)\,dt\) captures the cost of funding uncollateralized positions at the bank's spread \(s_F\) above risk-free, with asymmetric versions using different rates \(s_F^+, s_F^-\) for borrowing and lending. KVA \(= \int_0^T h\,K(t)\,S_B(t)\,D(0,t)\,dt\) represents the cost of holding regulatory capital \(K(t) = K^{\text{default}} + K^{\text{CVA}} + K^{\text{market}}\) at hurdle rate \(h\) (8--15%). MVA \(= \int_0^T s_F(t)\,\mathbb{E}[\text{IM}(t)]\,D(0,t)\,dt\) accounts for the cost of funding initial margin, which is always a cost and can be substantial for long-dated trades. XVA breaks classical linear pricing: \(V^{\text{total}}(A+B) \neq V^{\text{total}}(A) + V^{\text{total}}(B)\) due to netting, collateral, and capital effects, creating counterparty-dependent bid-ask spreads. The BSDE formulation
with driver \(f(t,V,Z) = -rV + \lambda_C\,\text{LGD}_C\,V^+ - \lambda_B\,\text{LGD}_B\,V^- + s_F(V-C)^+ + \cdots\) unifies all adjustments within a single semilinear PDE framework \(\partial_t V + \mathcal{L}V + f(t,x,V,\sigma^\top\nabla V) = 0\), handling interactions consistently and connecting to hedging strategies. Incremental XVA \(= \text{XVA}(\text{Portfolio} + \text{Trade}) - \text{XVA}(\text{Portfolio})\) determines trade-level pricing and profitability, with Euler allocation ensuring \(\sum_i \text{XVA}_i^{\text{allocated}} = \text{XVA}(\text{Portfolio})\).
Funding and Liquidity Risk¶
Funding constraints break the classical assumption of symmetric borrowing and lending rates: \(r^{\text{borrow}} > r^{\text{OIS}} > r^{\text{lend}}\), with the funding spread \(s_F(t) = r_B(t) - r(t)\) decomposing into credit, liquidity, and term premia. This asymmetry creates nonlinearities in pricing and can cause arbitrage to persist when \(\text{Arbitrage Profit} < \text{Funding Cost} + \text{Capital Cost}\). Regulatory liquidity requirements--LCR \(= \text{HQLA}/\text{Net Outflows} \geq 100\%\) (short-term), NSFR \(= \text{ASF}/\text{RSF} \geq 100\%\) (structural), leverage ratio \(\geq 3\%\)--constrain balance sheet deployment and increase effective funding costs. Margin procyclicality creates dangerous feedback loops: stress \(\to\) margin calls \(\to\) forced selling \(\to\) further price decline \(\to\) more margin calls. Liquidity premia compensate for illiquidity: the Amihud measure \(\text{ILLIQ} = \frac{1}{D}\sum_d |r_d|/V_d\) quantifies price impact per unit volume, with corporate bond premia of 50--100 bps for investment grade. The liquidity-adjusted CAPM (Acharya--Pedersen) adds a liquidity beta \(\beta_2\) capturing commonality in liquidity, return-liquidity sensitivity, and liquidity-return sensitivity. Brunnermeier--Pedersen liquidity spirals create feedback between market liquidity and funding liquidity, amplifying crises through forced selling and widening spreads. Market impact follows the empirical square-root law \(\Delta P \approx \sigma\sqrt{Q/V}\) where \(Q\) is order size and \(V\) is daily volume, with temporary and permanent components. The Almgren--Chriss framework optimizes execution: minimizing \(\mathbb{E}[C] + \lambda\,\text{Var}[C]\) over liquidation trajectories yields
with urgency parameter \(\kappa = \sqrt{\lambda\sigma^2/\eta}\) interpolating between risk-neutral TWAP (\(\kappa \to 0\)) and immediate execution (\(\kappa \to \infty\)), and implementation shortfall \(\mathbb{E}[C^*] = \gamma X^2 + \eta X^2\kappa/\tanh(\kappa T)\). Liquidation-adjusted VaR \(= \text{VaR} + \text{Liquidation Cost}\) integrates market impact into risk measurement, with stressed impact potentially 2--5\(\times\) normal during severe dislocations.
Systemic Risk and Network Effects¶
Financial institutions are interconnected through bilateral exposures \(\Pi_{ij}\), forming a network where nodes are institutions and weighted edges are claims, with core-periphery structures observed empirically. The Eisenberg--Noe clearing model determines payment vectors via the fixed-point equation \(p = \min(\bar{p},\, e + \Pi^\top p)\) where \(e_i\) is external assets and \(\bar{p}_i\) is total liabilities; institution \(i\) defaults if \(p_i < \bar{p}_i\), triggering cascading losses. Leverage amplifies shocks: a 1% asset decline causes a \((1/\text{equity ratio})\%\) equity decline. Contagion propagates through four channels: credit (cascading defaults when \(\sum_j \mathbf{1}_{\{j\text{ defaults}\}}\,\text{Exposure}_{ij}\,\text{LGD} > E_i\)), fire sales (price impact \(\Delta P_a = -\lambda_a\,\text{Sales}/D_a\) forcing further selling, as in the Greenwood--Landier--Thesmar model), information (signaling effects triggering bank runs), and funding (money-market stress and liquidity withdrawal). The basic reproduction number \(R_0 = \bar{d}\,p_{\text{default}} > 1\) (Gai--Kapadia) determines whether contagion spreads, with phase transitions: low connectivity yields isolated failures, intermediate connectivity maximizes contagion risk, and high connectivity enables shock absorption--the "robust-yet-fragile" property. Dynamic contagion follows epidemiological models \(dI/dt = \beta SI - \gamma I\) with tipping points and hysteresis. Systemic risk metrics include CoVaR \(\Delta\text{CoVaR}^j = \text{VaR}_\alpha(\text{System} \mid j\text{ at VaR}) - \text{VaR}_\alpha(\text{System} \mid j\text{ at median})\), marginal expected shortfall \(\text{MES}_i = \mathbb{E}[R_i \mid R_{\text{mkt}} < \text{VaR}_\alpha^{\text{mkt}}]\), and SRISK \(= k\,D_i - (1-k)\,W_i(1-\text{LRMES}_i)\) measuring capital shortfall in crisis. Central counterparties (CCPs) reduce bilateral risk through multilateral netting (\(\text{CCP Exposure} = \sum_i |E_i^{\text{CCP}}| < \sum_{i<j} |E_{ij}|\)) and standardized margining (variation margin \(\text{VM}_t = V_t - V_{t-1}\), initial margin \(\text{IM} \approx \text{VaR}_{99\%}\)), with a default waterfall (defaulter's IM \(\to\) defaulter's DF \(\to\) CCP skin-in-the-game \(\to\) survivors' DF \(\to\) assessments) absorbing losses sequentially under the Cover-2 standard. However, CCPs concentrate systemic risk as single points of failure, with margin procyclicality (\(\text{IM}_{\text{stressed}} \le 1.25 \times \text{IM}_{\text{normal}}\) as indicative cap) and CCP-bank interconnections creating two-way risk channels.
Model Risk and Governance¶
Model validation ensures conceptual soundness, implementation correctness, and fitness for purpose through independent review (SR 11-7 guidance), sensitivity analysis \(\partial\text{Output}/\partial\text{Input}_i\), benchmarking against alternative approaches, and outcome analysis including backtesting and out-of-sample performance. Models are tiered by materiality (Tier 1: annual or more frequent validation; Tier 2: annual; Tier 3: periodic) and rated via traffic-light systems. VaR backtesting compares predicted exceedances against realized: the Kupiec test \(LR_{uc} = -2\ln[(1-\alpha)^{n_0}\alpha^{n_1}/\hat{p}^{n_1}(1-\hat{p})^{n_0}] \sim \chi^2(1)\) checks unconditional coverage; the Christoffersen test checks independence of exceedances (\(\pi_{01} = \pi_{11}\)); the combined conditional coverage test \(LR_{cc} = LR_{uc} + LR_{ind} \sim \chi^2(2)\) checks both. The Basel traffic light classifies models by exceedances over 250 days at 99% VaR: green (0--4, multiplier 3.0), yellow (5--9, multiplier \(3.0 + 0.2\times(\text{exc}-4)\)), red (\(\geq 10\), multiplier 4.0). ES backtesting uses the Acerbi--Szekely identity \(\mathbb{E}[L\,\mathbf{1}_{L > \text{VaR}_\alpha}/\text{ES}_\alpha] = 1-\alpha\) and the McNeil--Frey residual test transforming exceedances to standard exponential. Risk limits translate model outputs into operational controls across a hierarchy (firm \(\to\) business unit \(\to\) desk \(\to\) trader), with VaR limits, stress limits, and sensitivity limits enforced through automated monitoring and escalation procedures. Robust risk frameworks address model uncertainty through model averaging \(\hat{\rho} = \sum_i w_i\rho_i\), envelope approaches \(\hat{\rho} = \max_i\rho_i\), Bayesian parameter integration \(p(\text{Risk}\mid\text{Data}) = \int p(\text{Risk}\mid\theta)\,p(\theta\mid\text{Data})\,d\theta\), and the maximin criterion \(\max_a\min_s U(a,s)\) for decision-making under uncertainty. Capital is set as \(\text{Capital} = \max(\text{Statistical Capital},\, \text{Stress Capital})\), with model risk buffers and prudent valuation adjustments. The three lines of defense--business units (risk ownership), risk management (oversight), internal audit (independent assurance)--provide layered governance, supported by comprehensive model inventories, change management protocols, and kill switches for predefined model override triggers. Basel regulatory requirements prescribe capital standards, internal model approval criteria, backtesting frameworks, and the use test ensuring models inform actual decision-making.
Role in the Book
Risk management integrates nearly every thread from the preceding chapters. Market risk measures build on the probability and stochastic process theory of Chapters 2--3; spectral risk measures and Kusuoka representations connect to the functional analysis foundations; the BSDE framework for dynamic risk measures extends Chapter 5's Feynman--Kac theory into nonlinear territory; counterparty credit risk and XVA draw on the credit modeling of Chapter 21 and the interest rate framework of Chapters 18--19; stress testing and extreme value theory connect to the calibration and model risk themes of Chapter 17; the Almgren--Chriss execution framework and funding liquidity models apply the stochastic control methods developed throughout the book; and systemic risk network models bring together probability, optimization, and fixed-point theory to address the stability of the financial system as a whole.