betaFeedback

WorksEconomics, Econometrics and Finance

Portfolio risk allocation through Shapley value

Abstract

We argue that using the Shapley value of cooperative game theory as the scheme for risk allocation among non-orthogonal risk factors is a natural way of interpreting the contribution made by each of such factors to overall portfolio risk. We discuss a Shapley value scheme for allocating risk to non-orthogonal greeks in a portfolio of derivatives. Such a situation arises, for example, when using a stochastic volatility model to capture option volatility smile. We also show that Shapley value allows for a natural method of interpreting components of enterprise risk measures such as VaR and ES. For all applications discussed, we derive explicit formulas and / or numerical algorithms to calculate the allocations.

Introduction

All quantitative financial risk management is scenario based. Depending on the the requirements, risk measures are calculated at various granularities: at the single instrument level, portfolio level, enterprise level, or systemic level. In the most myopic case, the risk of a single instrument, one subjects the instrument to specific scenarios in the factors affecting its value. Individual instruments and their risk metrics are aggregated into portfolios, portfolios combine into enterprises, which in turn form an interconnected global web of the financial industry. Moving up the ladder of risk aggregation introduces new categories of risk which require their own scenarios for quantification. The scenarios used for risk quantification may be microscopic (infinitesimal) or macroscopic (finite), they may be generated by a mathematical valuation model or based on historical market prices, they may be defined by the institution carrying the risk or prescribed by a regulating authority.

A typical situation arises as follows. We consider a portfolio of risky assets whose explainable risk is determined by a number of random state variables X1,…,Xn∈RX_1,\ldots,X_n\in\mathbb{R}. For mathematical modeling purposes we assume that the random variables XiX_i have finite first and second moments, E(∣Xi∣)<∞\mathbb{E}(|X_i|)<\infty, E(Xi2)<∞\mathbb{E}(X_i^2)<\infty, for all ii. The meaning of the state variables XiX_i depends on the context. They can represent the variables entering the pricing model of a derivative contract. In the context of portfolio management, one can think of each of the assets in the portfolio as defining a risk factor XiX_i. More generally, we can group subsets of assets in the portfolio into subportfolios, each of which defining a risk factor (such as industry, size, etc). Finally, the value of a financial institution itself may represent a risk factor in a model of systemic risk.

Portfolio risk is measured in terms of a suitable risk metric or a number of risk metrics. One such metric is the portfolio P&L, i.e. the change in value of the portfolio under a realized or simulated change in the market conditions. One of the key functions of a portfolio manager is to explain (attribute) the daily P&L to the risk factors XiX_i of the portfolio. Other metrics, such as value at risk (VaR), expected shortfall (ES), and stress tests are used for economic or regulatory capital calculation, margin sizing, etc. We do not require that these metrics satisfies any (possibly problematic) properties such as coherence. An important question is how to attribute, in a systematic and meaningful way, such a risk metric to each of the risk factors.

We let m(X)m(X) denote a generic risk metric, where

X=∑j=1nXj.(1)X=\sum_{j=1}^{n}X_j. \tag*{(1)}

Risk attribution is straightforward, if the risk factors are orthogonal, i.e. the XiX_i’s are mutually independent. This may be the case if the risk factors are defined in abstract terms, such as principal components, but is typically not true for financially intuitive risk factors. of non-orthogonal, subjecting a portfolio to a particular scenario may partly account for the same risk as it is exhibited in another, seemingly unrelated scenario. In particular, the risk of the entire portfolio is not simply the sum of its risk measured by all the scenarios.

In this paper we advocate the view that the concept of Shapley value studied in cooperative game theory [12] offers a “fair allocation” design for attributing portfolio risk to each of the risk factors affecting its value. Shapley value allocates overall risk to a risk factor XiX_i as the average marginal risk of each coalition of risk factors when XiX_i is added to the coalition. Earlier papers in the context of portfolio management in [14], [15], [13], enterprise risk management [19], [3], and systemic risk [19], [18], [6].

The paper is organized as follows. In § 2 we review the definition of Shapley value and some of the methods to compute it. Explicit expressions for computing the Shapley value of the variance and volatility games are discussed in § 3. In § 4, we apply the concept of Shapley value to risk attribution of a portfolio of derivatives with correlated risk factors. Finally, in § 5 we show how Shapley value can help with risk attribution at the enterprise risk management level.

Shapley value

For parts of the discussion in this paper we will find it convenient to use the concepts and language of game theory [12]. A coalitional game is a pair (N,v)(\mathcal{N},v) consisting of the set of players N\mathcal{N} and the characteristic function v:2N→Rv:2^{\mathcal{N}}\to\mathbb{R}. The only requirement on vv is that v(∅)=0v(\varnothing)=0. Any subset S⊂NS\subset\mathcal{N} is referred to as a coalition. In the context of finance, a risk metric mm defines a coalitional game whose characteristic function is v(S)=m(XS)v(S)=m(X_S).

Games and risk metrics

As discussed in the Introduction, a portfolio risk metric mm is determined by a number of stochastic risk factors XjX_j, j=1,…,nj=1,\ldots,n. In the following we will denote by N={1,…,n}\mathcal{N}=\{1,\ldots,n\} the set of indices labeling the risk factors, and view it as the set of players in a cooperative game. We consider the set of partial risk metrics m(XS)m(X_S), where S⊂NS\subset\mathcal{N}, with the convention that m(X∅)=0m(X_{\varnothing})=0.

The Shapley value [12] is a way of attributing the value of a coalitional game to each player based on the average marginal value that it contributes to each coalition. Namely, the Shapley value of player i∈Ni\in\mathcal{N} is defined by

Sh⁡i(v)=1n∑S⊂N∖{i}(n−1∣S∣)−1(v(S∪{i})−v(S)).(2)\operatorname{Sh}_i(v)=\frac{1}{n}\sum_{S\subset\mathcal{N}\setminus\{i\}}\binom{n-1}{|S|}^{-1}\left(v(S\cup\{i\})-v(S)\right). \tag*{(2)}

This can also be recast into the following more symmetric form:

Sh⁡i(v)=1n!∑π∈Sn(v(Sπi∪{i})−v(Sπi)),(3)\operatorname{Sh}_i(v)=\frac{1}{n!}\sum_{\pi\in S_n}\left(v(S_\pi^i\cup\{i\})-v(S_\pi^i)\right), \tag*{(3)}

where SnS_n denotes the set of all permutations of N\mathcal{N}, and Sπi={j∈N:π(j)<i}S_\pi^i=\{j\in\mathcal{N}:\pi(j)<i\}.

A key theorem, due to Lloyd Shapley, states that the Shapley value is the unique way of attributing the value of a coalitional game that satisfies a few natural “fairness” requirements [12]. One of these requirements is the following efficiency property:

∑i=1nSh⁡i(v)=v(N),(4)\sum_{i=1}^{n}\operatorname{Sh}_i(v)=v(\mathcal{N}), \tag*{(4)}

stating the sum of individual contributions is equal to the value of the grand coalition.

We conclude this section by deriving an approximate relation between the Shapley values of a game and its square. Assume that the characteristic function vv is the square of a characteristic function σ\sigma, i.e. v(S)=σ(S)2v(S)=\sigma(S)^2, for each S∈2NS\in2^{\mathcal{N}}. Then we can make the approximation

Sh⁡i(σ2)=1n∑S⊆N∖{i}(n−1∣S∣)−1(σ(S∪{i})−σ(S))(σ(S∪{i})+σ(S))≈Sh⁡i(σ)λ,\begin{aligned} \operatorname{Sh}_i(\sigma^2)=\frac{1}{n}\sum_{S\subseteq\mathcal{N}\setminus\{i\}}\binom{n-1}{|S|}^{-1}\bigl(\sigma(S\cup\{i\})-\sigma(S)\bigr)\bigl(\sigma(S\cup\{i\})+\sigma(S)\bigr) \\ &\approx\operatorname{Sh}_i(\sigma)\lambda, \end{aligned}

where λ\lambda represents an average value of σ(S∪{i})+σ(S)\sigma(S\cup\{i\})+\sigma(S). Applying equation (4) to both Sh⁡i(σ2)\operatorname{Sh}_i(\sigma^2) and Sh⁡i(σ)\operatorname{Sh}_i(\sigma) yields λ≈σ(N)\lambda\approx\sigma(\mathcal{N}), and thus

Sh⁡i(σ)≈Sh⁡i(σ2)σ(N).(5)\operatorname{Sh}_i(\sigma)\approx\frac{\operatorname{Sh}_i(\sigma^2)}{\sigma(\mathcal{N})}. \tag*{(5)}

Calculating Shapley value via Monte Carlo simulations

Calculating the Shapley value of a game is challenging, owing to the computational complexity of its definition. The only method that is universally applicable is Monte Carlo simulation, first considered in [11] in the context of the voting game. Refinements have been discussed in [4] and, for supermodular games, in [10].

For the purposes discussed in this paper, especially when the number of players is moderate, the following variation on the original algorithm [11] performs sufficiently well. Rather than presenting a pseudocode, below is the actual Python 3 code assuming that the package numpy has been imported as np.

def calc_shapley_vals(char_fct, num_players, sample_size):
    N = set(np.arange(0, num_players))
    shapley_vals = np.zeros(num_players)
    for i in range(num_players):
        N_minus_i = list(N - {i})
        shap = 0.0
        for j in range(sample_size):
            k = np.random.randint(0, num_players - i)
            S = np.random.choice(N_minus_i, k, replace=False)
            S_union_i = np.append(S, i)
            shap += char_fct(S_union_i) - char_fct(S)
        shapley_vals[i] = shap / sample_size
    return shapley_vals

This code assumes of course that the characteristic function char_fct has been defined within the appropriate context.

Variance and volatility games

The variance and volatility games [5] (see also [19] and [16]) are the basis for applications of Shapley value in portfolio management. In this section we review their properties and derive explicit allocation formulas.

Analytic expressions for the variance and volatility games

Consider nn random variables XiX_i, i=1,…,ni=1,\ldots,n distributed so that their first and second moments exist. We consider the variance game defined by the following characteristic function. For S⊂NS \subset\mathcal{N}, v(S)v(S) is the variance of XX conditioned on the complement of SS,

v(S)=Var⁡(X∣XSc)=Var⁡(XS),(6)\begin{aligned} v(S) &= \operatorname{Var}(X\mid X_{S^c}) \\ &= \operatorname{Var}(X_S), \tag*{(6)} \end{aligned}

where

XS=∑j∈SXj.(7)X_S = \sum_{j\in S} X_j. \tag*{(7)}

Explicitly,

v(S)=∑j,k∈SCov⁡(Xj,Xk).(8)v(S) = \sum_{j,k\in S} \operatorname{Cov}(X_j,X_k). \tag*{(8)}

Note the decomposition:

v(S∪T)+v(S∩T)=v(S)+v(T)+2∑j∈S∖T, k∈T∖SCov⁡(Xj,Xk)=v(S)+v(T)+2Cov⁡(XS∖T,XT∖S),(9)\begin{aligned} v(S\cup T)+v(S\cap T) &= v(S)+v(T)+2\sum_{j\in S\setminus T,\,k\in T\setminus S}\operatorname{Cov}(X_j,X_k) \\ &= v(S)+v(T)+2\operatorname{Cov}(X_{S\setminus T},X_{T\setminus S}), \tag*{(9)} \end{aligned}

for all S,T⊂NS,T\subset\mathcal{N}. In particular, if i∉Si\notin S, then

v(S∪i)=v(S)+Var⁡(Xi)+2Cov⁡(Xi,XS).(10)v(S\cup i)=v(S)+\operatorname{Var}(X_i)+2\operatorname{Cov}(X_i,X_S). \tag*{(10)}

Also notice that

∑π∈SnCov⁡(Xi,XSπi)=∑π−1∈SnCov⁡(Xi,XSπ−1i)=∑π∈SnCov⁡(Xi,XSπi),\begin{aligned} \sum_{\pi\in S_n}\operatorname{Cov}(X_i,X_{S_\pi^i}) &= \sum_{\pi^{-1}\in S_n}\operatorname{Cov}(X_i,X_{S_{\pi^{-1}}^i}) \\ &= \sum_{\pi\in S_n}\operatorname{Cov}(X_i,X_{S_\pi^i}), \end{aligned}

where Sπi={j∈N:π(j)>i}S_\pi^i=\{j\in\mathcal{N}:\pi(j)>i\}. As a consequence, from (3),

Sh⁡i(v)=1n!∑π∈Sn(Var⁡(Xi)+2Cov⁡(Xi,XSπi))=1n!∑π∈Sn(Cov⁡(Xi,Xi+XSπi)+Cov⁡(Xi,XSπi))=1n!∑π∈SnCov⁡(Xi,Xi+XSπi+XSπi)=1n!Cov⁡(Xi,∑π∈SnX).\begin{aligned} \operatorname{Sh}_i(v) &= \frac{1}{n!}\sum_{\pi\in S_n}\left(\operatorname{Var}(X_i)+2\operatorname{Cov}(X_i,X_{S_\pi^i})\right) \\ &= \frac{1}{n!}\sum_{\pi\in S_n}\left(\operatorname{Cov}(X_i,X_i+X_{S_\pi^i})+\operatorname{Cov}(X_i,X_{S_\pi^i})\right) \\ &= \frac{1}{n!}\sum_{\pi\in S_n}\operatorname{Cov}(X_i,X_i+X_{S_\pi^i}+X_{S_\pi^i}) \\ &= \frac{1}{n!}\operatorname{Cov}\left(X_i,\sum_{\pi\in S_n}X\right). \end{aligned}

This leads to the following equation, established in [5]:

Sh⁡i(v)=Cov⁡(Xi,X).(11)\operatorname{Sh}_i(v)=\operatorname{Cov}(X_i,X). \tag*{(11)}

We are unaware of a closed form expression for the Shapley value of the volatility game with characteristic function given by

σ(S)=Var⁡(XS).(12)\sigma(S)=\sqrt{\operatorname{Var}(X_S)}. \tag*{(12)}

However, approximation (5) leads to the following intuitively appealing formula:

Sh⁡i(σ)≈Cov⁡(Xi,X)σ(X),\operatorname{Sh}_i(\sigma)\approx\frac{\operatorname{Cov}(X_i,X)}{\sigma(X)},

or

Sh⁡i(σ)≈ρ(Xi,X)σ(Xi).(13)\operatorname{Sh}_i(\sigma)\approx\rho(X_i,X)\sigma(X_i). \tag*{(13)}

where ρ(Xi,X)\rho(X_i,X) denotes the correlation coefficient between XiX_i and XX.

The attribution of risk factor XiX_i is proportional to the volatility of that factor, with the proportionality coefficient reflecting the correlation between the factor and the overall risk. Below we verify numerically that this approximation is remarkably accurate.

Numerical simulations

In Table 1 below we present the results of a test of these results based on simulated data. In particular, we test the accuracy of the analytic approximation (13) for the volatility game.

SVCVSSCS
1,268.11,191.98.96.6
1,868.71,837.010.610.2
3,371.23,355.517.018.7
3,992.84,042.318.422.5
1,260.31,231.17.96.9
1,371.31,310.78.77.3
3,018.93,034.114.816.9
4,226.84,306.918.724.0
799.7740.36.64.1
3,436.63,476.416.219.4
1,002.0930.37.55.2
198.4145.33.70.8
693.2686.04.43.8
2,443.92,468.211.713.8
1,741.41,775.38.49.9
524.9494.84.02.8
306.9286.33.31.6
-284.0-320.80.5-1.8
-221.5-242.30.1-1.4
869.5877.04.74.9
640.7648.73.63.6
84.560.42.00.3
-529.1-564.8-1.0-3.1
47.235.21.10.2
451.1463.62.32.6

Table 1. Shapley value attribution to the variance (the first two columns) and volatility (the last two columns) games using Monte Carlo and analytic expressions.

The data has been prepared as follows.

  1. We generate a 25×2525 \times25 random (positive definite) covariance matrix Σ\Sigma.

  1. We generate 500 independent samples XiX_i from the multivariate Gaussian distribution N(0,Σ)\mathcal{N}(0,\Sigma),

Xi∼N(0,Σ),i=1,…,500.X_i\sim\mathcal{N}(0,\Sigma),\quad i=1,\ldots,500.

We chose 500 observations, as this corresponds to the observation window of 2 years, frequently chosen in the financial industry for VaR calculation.

The meaning of the columns in Table 1 is as follows:

  1. SV represents Sh⁡^i(v)\widehat{\operatorname{Sh}}_i(v),

  1. CV represents Cov⁡(Xi,X)\operatorname{Cov}(X_i,X),

  1. SS represents Sh⁡^i(σ)\widehat{\operatorname{Sh}}_i(\sigma),

  1. CS represents ρ(Xi,X)σ(Xi)\rho(X_i,X)\sigma(X_i),

for i=1,…,25i=1,\ldots,25. For the Monte Carlo computations of the Shapley value (columns one and three), we used 100,000 simulations.

P&L attribution of a derivatives portfolio

In this section we present the Shapley value approach to the problem of risk (and daily P&L) attribution of a portfolio of derivatives to a set of possibly non-orthogonal risk factors. The challenge here is to distribute fairly the portion of portfolio P&L that could in principle be attributed to several risk factors such as delta and vega.

SABR model

In order to motivate the discussion, we first consider a concrete example, namely a single European option on an asset whose volatility follows the SABR model [8]. The dynamics of the SABR model is specified in terms of two state variables: forward FtF_t and instantaneous volatility parameter σt\sigma_t, and is given by:

dFt=σtC(Ft)dWt,dσt=ασtdZt,(14)\begin{aligned} \mathrm{d}F_t &= \sigma_t C(F_t)\mathrm{d}W_t,\\ \mathrm{d}\sigma_t &= \alpha\sigma_t\mathrm{d}Z_t, \tag*{(14)} \end{aligned}

where WtW_t and ZtZ_t are two Brownian motions with

dWtdZt=ρdt.(15)\mathrm{d}W_t\mathrm{d}Z_t=\rho\mathrm{d}t. \tag*{(15)}

The function C(x)C(x) determines the backbone of the volatility smile, and is often assumed to be of the form C(x)=xβC(x)=x^\beta, where β≤1\beta\leq1.1

In practice, the SABR model is used via an asymptotic approximation in combination with the standard Black-Scholes model. Namely, let BB denote the Black-Scholes pricing function in the normal model, i.e.

B(τ,F,K,σ)={στ(d+N(d+)+N′(d+)),for a call option,στ(d−N(d−)+N′(d−)),for a put option,(16)B(\tau,F,K,\sigma)= \begin{cases} \sigma\sqrt{\tau}\left(d_{+}N(d_{+})+N'(d_{+})\right), & \text{for a call option},\\ \sigma\sqrt{\tau}\left(d_{-}N(d_{-})+N'(d_{-})\right), & \text{for a put option}, \end{cases} \tag*{(16)}

where N(x)N(x) denotes the cumulative normal distribution, and where

d±=±F−Kστ.(17)d_{\pm}=\pm\frac{F-K}{\sigma\sqrt{\tau}}. \tag*{(17)}

Then the time tt value VtV_t of an option struck at KK with expiration TT is given by

Vt=B(T−t,Ft,K,σimp(T−t,Ft,K,σt)).(18)V_t=B\left(T-t,F_t,K,\sigma^{\mathrm{imp}}(T-t,F_t,K,\sigma_t)\right). \tag*{(18)}

Here σimp\sigma^{\mathrm{imp}} is an explicit approximation to the (analytically unknown expression for the) normal volatility implied by the SABR model, see [8], [9] for details. For simplicity, we have disregarded discounting.

This valuation formula is the basis for calculating various sensitivity measures of the option value used to measure its exposure to the factors driving the model. Key among them are the delta

Δ=∂B∂F+∂B∂σ∂σimp∂F,(19)\Delta=\frac{\partial B}{\partial F}+\frac{\partial B}{\partial\sigma}\frac{\partial\sigma^{\mathrm{imp}}}{\partial F}, \tag*{(19)}

and vega

Λ=∂B∂σ∂σimp∂σ,(20)\Lambda=\frac{\partial B}{\partial\sigma}\frac{\partial\sigma^{\mathrm{imp}}}{\partial\sigma}, \tag*{(20)}

calculated as the sensitivities to the “naive” scenarios Ft→Ft+dFtF_t\to F_t+\mathrm{d}F_t, σt→σt\sigma_t\to\sigma_t, and Ft→FtF_t\to F_t, σt→σt+dσt\sigma_t\to\sigma_t+\mathrm{d}\sigma_t, respectively. It was argued in [1], [7] that these scenarios are inadequate if ρ≠0\rho\ne0, which may lead in practice to mishedging of the portfolio.

Instead, we proceed as follows. We decompose the Brownian motion ZtZ_t into WtW_t and a Brownian motion Wt⊥W_t^\perp, independent of WtW_t: Zt=ρWt+1−ρ2 Wt⊥Z_t=\rho W_t+\sqrt{1-\rho^2}\,W_t^\perp. Then, under the forward scenario Ft→Ft+dFtF_t\to F_t+\mathrm{d}F_t, the volatility parameter σt\sigma_t is bound to move as follows: σt→σt+ρα/C(Ft) dFt\sigma_t\to\sigma_t+\rho\alpha/C(F_t)\,\mathrm{d}F_t. Denoting by dσt⊥=ασtdWt⊥\mathrm{d}\sigma_t^\perp=\alpha\sigma_t\mathrm{d}W_t^\perp the component of dσt\mathrm{d}\sigma_t uncorrelated with dFt\mathrm{d}F_t, we obtain from Ito’s lemma:

dσtimp=−∂σimp∂τdt+(∂σimp∂F+∂σimp∂σραC(Ft))dFt+∂σimp∂σdσt⊥+12σt2(C(Ft)2∂2σimp∂2F+2ρC(Ft)∂2σimp∂F∂σ+α2∂2σimp∂2σ)dt.\mathrm{d}\sigma_t^{\mathrm{imp}}=-\frac{\partial\sigma^{\mathrm{imp}}}{\partial\tau}\mathrm{d}t+\left(\frac{\partial\sigma^{\mathrm{imp}}}{\partial F}+\frac{\partial\sigma^{\mathrm{imp}}}{\partial\sigma}\frac{\rho\alpha}{C(F_t)}\right)\mathrm{d}F_t+\frac{\partial\sigma^{\mathrm{imp}}}{\partial\sigma}\mathrm{d}\sigma_t^\perp +\frac{1}{2}\sigma_t^2\left(C(F_t)^2\frac{\partial^2\sigma^{\mathrm{imp}}}{\partial^2F}+2\rho C(F_t)\frac{\partial^2\sigma^{\mathrm{imp}}}{\partial F\partial\sigma}+\alpha^2\frac{\partial^2\sigma^{\mathrm{imp}}}{\partial^2\sigma}\right)\mathrm{d}t.

where τ=T−t\tau= T - t. This leads the to the following risk decomposition for the P&L:

dVt=(−Θt+12σt2(C(Ft)2Γt+2C(Ft)VAt+α2VOt))dt+ΔtBdFt+Λtdσt⊥,(21)dV_t = \left(-\Theta_t + \frac{1}{2}\sigma_t^2\left(C(F_t)^2\Gamma_t + 2C(F_t)VA_t + \alpha^2VO_t\right)\right)dt + \Delta_t^B dF_t + \Lambda_t d\sigma_t^\perp, \tag*{(21)}

where

ΔB=Δ+ραC(F)∂B∂σ∂σimp∂σ(22)\Delta^B = \Delta+ \rho\frac{\alpha}{C(F)}\frac{\partial B}{\partial\sigma}\frac{\partial\sigma^{\mathrm{imp}}}{\partial\sigma} \tag*{(22)}

is the modified (Bartlett) delta. The other greeks are defined as follows:

Θ=∂B∂τ+∂B∂σ∂σimp∂τ(23)\Theta= \frac{\partial B}{\partial\tau} + \frac{\partial B}{\partial\sigma}\frac{\partial\sigma^{\mathrm{imp}}}{\partial\tau} \tag*{(23)}

is the time decay,

Γ=∂2B∂F2+∂B∂σ∂2σimp∂F2(24)\Gamma= \frac{\partial^2 B}{\partial F^2} + \frac{\partial B}{\partial\sigma}\frac{\partial^2 \sigma^{\mathrm{imp}}}{\partial F^2} \tag*{(24)}

is the gamma,

VA=∂2B∂F∂σ+∂B∂σ∂2σimp∂F∂σ(25)VA = \frac{\partial^2 B}{\partial F\partial\sigma} + \frac{\partial B}{\partial\sigma}\frac{\partial^2 \sigma^{\mathrm{imp}}}{\partial F\partial\sigma} \tag*{(25)}

is the vanna, and

VO=∂2B∂σ2+∂B∂σ∂2σimp∂σ2(26)VO = \frac{\partial^2 B}{\partial\sigma^2} + \frac{\partial B}{\partial\sigma}\frac{\partial^2 \sigma^{\mathrm{imp}}}{\partial\sigma^2} \tag*{(26)}

is the volga. Formula (21) decomposes the option P&L in terms of independent risk factors dFdF and dσ⊥d\sigma^\perp, time decay, and second order greeks. It was shown in [1] and [7] that this approach to calculating delta and vega leads to more robust hedges than those obtained by means of the traditionally calculated sensitivities.

Alternatively, we can represent WtW_t in terms of ZtZ_t and its independent complement Zt⊥Z_t^\perp as Wt=ρZt+1−ρ2dZt⊥W_t = \rho Z_t + \sqrt{1-\rho^2}dZ_t^\perp, which yields

dVt=(−Θt+12σt2(C(Ft)2Γt+2C(Ft)VAt+α2VOt))dt+ΔtdFt⊥+ΛtBdσt.(27)dV_t = \left(-\Theta_t + \frac{1}{2}\sigma_t^2\left(C(F_t)^2\Gamma_t + 2C(F_t)VA_t + \alpha^2VO_t\right)\right)dt + \Delta_t dF_t^\perp+ \Lambda_t^B d\sigma_t. \tag*{(27)}

Here,

ΛB=Λ+ρC(F)α(∂B∂σ∂σimp∂F+∂B∂F)(28)\Lambda^B = \Lambda+ \rho\frac{C(F)}{\alpha}\left(\frac{\partial B}{\partial\sigma}\frac{\partial\sigma^{\mathrm{imp}}}{\partial F} + \frac{\partial B}{\partial F}\right) \tag*{(28)}

is the modified (Bartlett) vega. Formula (27) is a decomposition of the option’s risk in terms of an alternative basis of independent risk factors, namely dF⊥dF^\perp and dσd\sigma. Note that theta and the second order greeks enter both decompositions in the same manner.

These two forms of risk decomposition show that part of the option’s volatility sensitivity can be viewed either as a component of its delta or of its vega:

dMt≡ΔtBdFt+Λtdσt⊥=ΔtBdFt+ΛtBdσt−((ΔtB−Δt)dFt+(ΛtB−Λt)dσt)=ΔtdFt⊥+ΛtBdσt.\begin{aligned} dM_t \equiv\Delta_t^B dF_t + \Lambda_t d\sigma_t^\perp\\ &= \Delta_t^B dF_t + \Lambda_t^B d\sigma_t - \left((\Delta_t^B-\Delta_t)dF_t + (\Lambda_t^B-\Lambda_t)d\sigma_t\right) \\ &= \Delta_t dF_t^\perp+ \Lambda_t^B d\sigma_t. \end{aligned}

In the case of ρ≠0\rho\ne0, there is no obvious way of attributing the P&L to the delta risk and the vega risk2. We propose that the “fair attribution” given by Shapley value is a natural approach.

Notice that the Shapley value of a game with two players is given by

Sh⁡1(v)=12(v({1,2})+v({1})−v({2})),Sh⁡2(v)=12(v({1,2})+v({2})−v({1})).(29)\begin{aligned} \operatorname{Sh}_{1}(v) &= \frac{1}{2}\left(v(\{1,2\})+v(\{1\})-v(\{2\})\right),\\ \operatorname{Sh}_{2}(v) &= \frac{1}{2}\left(v(\{1,2\})+v(\{2\})-v(\{1\})\right). \tag*{(29)} \end{aligned}

Applying these formulas to dMdM, the martingale part of dVdV, yields

Sh⁡Δ(dM)=12(ΔBdF+ΔdF⊥),Sh⁡Λ(dM)=12(ΛBdσ+Λdσ⊥).(30)\begin{aligned} \operatorname{Sh}_{\Delta}(dM) &= \frac{1}{2}\left(\Delta^{B}dF+\Delta dF^{\perp}\right),\\ \operatorname{Sh}_{\Lambda}(dM) &= \frac{1}{2}\left(\Lambda^{B}d\sigma+\Lambda d\sigma^{\perp}\right). \tag*{(30)} \end{aligned}

This can alternatively be expressed in the following form:

Sh⁡Δ(dM)=ΔBdF−12((ΔB−Δ)dF+(ΛB−Λ)dσ),Sh⁡Λ(dM)=ΛBdσ−12((ΔB−Δ)dF+(ΛB−Λ)dσ).(31)\begin{aligned} \operatorname{Sh}_{\Delta}(dM) &= \Delta^{B}dF-\frac{1}{2}\left((\Delta^{B}-\Delta)dF+(\Lambda^{B}-\Lambda)d\sigma\right),\\ \operatorname{Sh}_{\Lambda}(dM) &= \Lambda^{B}d\sigma-\frac{1}{2}\left((\Delta^{B}-\Delta)dF+(\Lambda^{B}-\Lambda)d\sigma\right). \tag*{(31)} \end{aligned}

In other words, Shapley value allocates half of the common risk components to each delta and vega risks. Note that when ρ=0\rho=0,

Sh⁡Δ(dM)=ΔdF,Sh⁡Λ(dM)=Λdσ,(32)\begin{aligned} \operatorname{Sh}_{\Delta}(dM) &= \Delta dF,\\ \operatorname{Sh}_{\Lambda}(dM) &= \Lambda d\sigma, \tag*{(32)} \end{aligned}

as expected.

General case

We can now extend this discussion to the general case of an arbitrary number of risk factors. We consider a portfolio of derivatives whose price dynamics is captured by a multifactor diffusion process:

dXt,j=Dj(t,Xt)dWt,j,dWt,jdWt,k=ρjkdt.(33)\begin{aligned} dX_{t,j} &= D_j(t,X_t)dW_{t,j},\\ dW_{t,j}dW_{t,k} &= \rho_{jk}dt. \tag*{(33)} \end{aligned}

The state vector Xt∈RnX_t \in\mathbb{R}^{n} may consist of both market observable factors such as prices, yields, exchange rates, etc, as well as of unobservable factors such as parameters capturing stochastic volatility, convenience yields, etc. For convenience, we work under the measure under which the pricing process is a martingale. The portfolio value process is given by Vt=V(t,Xt)V_t=V(t,X_t), where V(x)V(x) is a valuation function.

Traditional way to calculate the sensitivity of the portfolio value to XiX_i is via the infinitesimal scenario:

Xi→Xi+dXi,Xj→Xj,for j≠i.(34)\begin{aligned} X_i &\to X_i+dX_i,\\ X_j &\to X_j,\quad\text{for }j\ne i. \tag*{(34)} \end{aligned}

This leads to the usual sensitivities,

Δi=∇iV,(35)\Delta_i=\nabla_i V, \tag*{(35)}

where ∇i\nabla_i denotes partial derivative.

Alternatively, and more appropriately, in order to account for the correlation effects among the risk factors, we consider a scenario under which XiX_i moves by dXidX_i:

Xi→Xi+dXi,(36)X_i \to X_i + dX_i, \tag*{(36)}

accompanied by the induced moves of the remaining state variables:

Xj→Xj+ρjiDj(X)Di(X)dXi+dWj⊥.(37)X_j \to X_j + \rho_{ji}\frac{D_j(X)}{D_i(X)}dX_i + dW_j^{\perp}. \tag*{(37)}

Here, each dWj⊥dW_j^{\perp} is uncorrelated with dWidW_i. Hence the correlation adjusted sensitivity (generalized Bartlett’s delta) is given by

Di=∇iV+∑j≠iρijDj(X)Di(X)∇jV,(38)\mathcal{D}_i = \nabla_i V + \sum_{j\ne i}\rho_{ij}\frac{D_j(X)}{D_i(X)}\nabla_j V, \tag*{(38)}

or

Di=Δi+∑j≠iρijDj(X)Di(X)Δj.(39)\mathcal{D}_i = \Delta_i + \sum_{j\ne i}\rho_{ij}\frac{D_j(X)}{D_i(X)}\Delta_j. \tag*{(39)}

We can thus write

dV=V˙dt+12Var⁡(dX)+DidXi+∑j≠iΔjdXj⊥,(40)dV = \dot{V}dt + \frac{1}{2}\operatorname{Var}(dX) + \mathcal{D}_i dX_i + \sum_{j\ne i}\Delta_j dX_j^{\perp}, \tag*{(40)}

where V˙\dot{V} denotes time derivative. This equation expresses the impact of a move in the state variable XiX_i on the value of the asset in terms of the correlation adjusted sensitivity to the variable XiX_i, and the classic sensitivities multiplying the moves of the remaining variables with the components explained by the change in XiX_i removed.

Calculating Shapley value of the martingale part dMdM of dVdV is straightforward, and it yields

Sh⁡i(dV)=1n(DidXi+∑j≠iΔjdXj⊥).(41)\operatorname{Sh}_i(dV) = \frac{1}{n}\left(\mathcal{D}_i dX_i + \sum_{j\ne i}\Delta_j dX_j^{\perp}\right). \tag*{(41)}

The variance part of dVdV can be attributed using equation (11). As a result we find that the P&L attribution to the ii-th factor is given by

Sh⁡i(dV)=1n(DidXi+∑j≠iΔjdXj⊥)+12Cov⁡(dXi,dX).(42)\operatorname{Sh}_i(dV) = \frac{1}{n}\left(\mathcal{D}_i dX_i + \sum_{j\ne i}\Delta_j dX_j^{\perp}\right) + \frac{1}{2}\operatorname{Cov}(dX_i,dX). \tag*{(42)}

Notice that, strictly speaking, time decay is also part of the P&L attribution process. We have excluded it from the discussion, as it is not a risk factor and its attribution is straightforward.

In practice, the infinitesimal quantities dXidX_i are replaced by the (daily) changes δXi\delta X_i, and the equation above takes the form of the following approximation:

Sh⁡i(P&L)≈1n(DiδXi+∑j≠iΔjδXj⊥)+12Cov⁡(δXi,δX).(43)\operatorname{Sh}_i(\mathrm{P\&L}) \approx\frac{1}{n}\left(\mathcal{D}_i \delta X_i + \sum_{j\ne i}\Delta_j \delta X_j^{\perp}\right) + \frac{1}{2}\operatorname{Cov}(\delta X_i,\delta X). \tag*{(43)}

The difference P&L−∑iSh⁡i(P&L)\mathrm{P\&L} - \sum_i \operatorname{Sh}_i(\mathrm{P\&L}) is the unexplained P&L.

Attribution of VaR and ES

In this section we discuss how Shapley value can be used to allocate enterprise risk metrics such as VaR and ES to the risk factors of the portfolio. This can be accomplished regardless of the methodology used for VaR and ES estimation.

Elliptic distribution of losses

Value at Risk and Expected Shortfall are risk metrics typically used at the enterprise risk management level. Their primary goal is to determine the amount of cash reserves required to be held against potential portfolio losses at a given confidence level defined as q≡1−αq \equiv1-\alpha (say, 95%), where α\alpha is the significance level. Among the approaches to VaR and ES, the approach based on a parametric distribution of losses is the most straightforward, albeit not the most commonly used. Its main advantage is the conceptual simplicity and ease of estimation: only a well defined set of parameters requires estimation. On the negative side, this approach implicitly assumes stationarity of portfolio returns and often leads to unrealistic tail dependence.

A large class of examples of relevant parametric distributions is provided by the family of multifactor elliptical distributions [2]. Recall that a random vector X∈RnX \in\mathbb{R}^{n} has an elliptical distribution with parameters μ∈Rn\mu\in\mathbb{R}^{n} and positive definite Σ∈Mat⁡n(R)\Sigma\in\operatorname{Mat}_{n}(\mathbb{R}), if its density function is of the form

cndet⁡(Σ)−1/2h((x−μ)TΣ−1(x−μ)),c_{n}\det(\Sigma)^{-1/2}h\left((x-\mu)^{T}\Sigma^{-1}(x-\mu)\right),

with a nonnegative function hh on R\mathbb{R} and a normalizing constant cnc_{n}. Assuming that the distribution has moments of order one and two (which we are doing), E(X)=μE(X)=\mu, and Cov⁡(X)=κnΣ\operatorname{Cov}(X)=\kappa_{n}\Sigma, where κn\kappa_{n} is a constant. Important examples of an elliptical distribution are the multivariate Gaussian distribution, multivariate Student t distribution, and Laplace distribution.

Let XX represent the portfolio’s daily P&L. Under an elliptical distribution, VaR⁡q\operatorname{VaR}_{q} and ES⁡q\operatorname{ES}_{q} are linear in the standard deviation σ(X)=Var⁡(X)\sigma(X)=\sqrt{\operatorname{Var}(X)},

VaR⁡q(X)=μ(X)−Aqσ(X),(44)\operatorname{VaR}_{q}(X)=\mu(X)-A_{q}\sigma(X), \tag*{(44)}

and

ES⁡q(X)=μ(X)−Bqσ(X).(45)\operatorname{ES}_{q}(X)=\mu(X)-B_{q}\sigma(X). \tag*{(45)}

Here, AqA_{q} and BqB_{q} are constants depending only on the distribution and confidence level, and μ(X)\mu(X) is linear in XX. For example, if XX is distributed according to the the multinomial Gaussian distribution N(μ,Σ)N(\mu,\Sigma) and σ(X)=Var⁡(X)\sigma(X)=\sqrt{\operatorname{Var}(X)}, then

VaR⁡q(X)=μTX−Φ−1(q)σ(X),(46)\operatorname{VaR}_{q}(X)=\mu^{T}X-\Phi^{-1}(q)\sigma(X), \tag*{(46)}

where Φ(z)\Phi(z) is the cumulative standard normal distribution, and

ES⁡q(X)=μTX−φ(Φ−1(q))1−qσ(X),(47)\operatorname{ES}_{q}(X)=\mu^{T}X-\frac{\varphi(\Phi^{-1}(q))}{1-q}\sigma(X), \tag*{(47)}

where φ(z)=(2π)−1/2exp⁡(−z2/2)\varphi(z)=(2\pi)^{-1/2}\exp(-z^{2}/2) is the density of Φ(z)\Phi(z). Analogous explicit expressions exist for the Student t distribution, Laplace distribution, and others.

In order to find the Shapley value of (the nonlinear cooperative games defined by) VaR⁡α\operatorname{VaR}_{\alpha} and ES⁡α\operatorname{ES}_{\alpha}, we note first that the Shapley value of (the linear cooperative game defined by) μ\mu is given by Sh⁡i(X)=μ(Xi)\operatorname{Sh}_i(X)=\mu(X_i). Using the additive property of Shapley value, and the approximation (13), we find that

Sh⁡i(VaR⁡q)≈μ(Xi)−Aqρ(Xi,X)σ(Xi),(48)\operatorname{Sh}_i(\operatorname{VaR}_q) \approx\mu(X_i)-A_q\rho(X_i,X)\sigma(X_i), \tag*{(48)}

and

Sh⁡i(ES⁡q)≈μ(Xi)−Bqρ(Xi,X)σ(Xi).(49)\operatorname{Sh}_i(\operatorname{ES}_q) \approx\mu(X_i)-B_q\rho(X_i,X)\sigma(X_i). \tag*{(49)}

Historical bootstrapping

Historical bootstrapping (and variations thereof) is the default methodology for VaR and ES estimation. In this approach VaR is estimated as the relevant quantile of the empirical distribution of the (daily) P&L of a portfolio, calculated over a specified lookback period of historical data. Expected shortfall is estimated as the arithmetic average of the losses exceeding the estimated VaR level. This approach has the advantage that it is blind to the stationarity and tail properties of the portfolio return distribution.

For example, in order to find the 95% VaR with a two year lookback period, we (i) calculate the P&L over the past 500 business days, (ii) rank it in the ascending order PL1≤⋯≤PL500PL_1 \le\cdots\le PL_{500} (the largest loss first) and (iii) set VaR⁡95%=PL25\operatorname{VaR}_{95\%}=PL_{25}. The estimated expected shortfall is ES⁡95%=125∑j=125PLj\operatorname{ES}_{95\%}=\frac{1}{25}\sum_{j=1}^{25}PL_j.

The so estimated VaR and ES can be easily attributed to the predefined risk factors using Shapley value. To this end, for each chosen risk factor jj we select a subset XjX_j of the set of assets XX that represents this risk factor. Note that the subportfolios XjX_j representing different risk factors are not required to be mutually disjoint. This allows us to assign a particular asset to several risk groups. For example, in case of a portfolio of equities, these subsets can represent the stock’s industry, size, liquidity, etc.

Using the paradigm of historical bootstrapping, we can estimate the values VaR⁡q(S)\operatorname{VaR}_q(S) and ES⁡q(S)\operatorname{ES}_q(S) of a coalition SS of the factors defined by the subportfolios XjX_j. To this end we define the following characteristic functions VaR⁡h\operatorname{VaR}^h and ES⁡h\operatorname{ES}^h: 2N→R2^{\mathcal{N}}\to\mathbb{R}:

VaR⁡h(S)=historical VaR estimated off the series Xt,S=∑j∈SXt,j, t=1,…,500,(50)\operatorname{VaR}^h(S)=\text{historical VaR estimated off the series }X_{t,S}=\sum_{j\in S}X_{t,j},\ t=1,\ldots,500, \tag*{(50)}

and

ES⁡h(S)=historical ES estimated off the series Xt,S=∑j∈SXt,j, t=1,…,500.(51)\operatorname{ES}^h(S)=\text{historical ES estimated off the series }X_{t,S}=\sum_{j\in S}X_{t,j},\ t=1,\ldots,500. \tag*{(51)}

This allows us to allocate the portfolio VaR or ES (estimated via historical bootstrapping) to each of the factors: (i) through an exact calculation, if feasible, or (ii) via Monte Carlo simulations, otherwise (in case of a large number of factors).

Numerical simulations

Table 2 a numerical test of the concepts discussed in this section. To this end, we reuse the simulated data of Section 3.2. This data is meant to simulate daily P&L of a portfolio of 25 assets observed over the period of two years. Its historical VaR and ES at the 95% level of confidence are given by -302.4 and -358.3, respectively, and the corresponding values estimated from multifactor Gaussian model are -295.2 and -370.2. Not surprisingly they are close to each other as they represent finite sample estimates of the same quantities.

HVGVHEGE
-10.5-10.9-12.5-13.7
-22.0-16.8-25.9-21.1
-33.7-30.8-39.6-38.6
-30.8-37.1-39.7-46.5
-14.3-11.3-15.7-14.2
-12.8-12.0-15.5-15.1
-19.3-27.8-27.3-34.9
-27.9-39.5-33.2-49.5
-14.9-6.8-15.4-8.5
-24.7-31.9-31.3-40.0
-10.8-8.5-11.6-10.7
-5.5-1.3-6.6-1.7
-3.5-6.3-4.4-7.9
-19.4-22.6-24.5-28.4
-18.1-16.3-24.6-20.4
-8.0-4.5-9.7-5.7
-4.1-2.6-6.7-3.3
0.02.90.93.7
-0.92.2-0.72.8
-8.3-8.0-9.7-10.1
-3.3-5.9-3.0-7.5
-3.3-0.6-3.9-0.7
0.75.21.26.5
-4.4-0.3-5.2-0.4
-3.3-4.2-3.5-5.3

Table 2. Shapley value attribution of VaR (the first two columns) and ES (the last two columns) using historical and Gaussian estimates

The meaning of the columns in Table 2 is as follows:

  1. HV represents Sh^i(VaRh)\widehat{\mathrm{Sh}}_i(\mathrm{VaR}^{h}),

  1. GV represents Aqρ(Xi,X)σ(Xi)A_{q\rho}(X_i, X)\sigma(X_i),

  1. HE represents Sh^i(ESh)\widehat{\mathrm{Sh}}_i(\mathrm{ES}^{h}),

  1. GE represents Bqρ(Xi,X)σ(Xi)B_{q\rho}(X_i, X)\sigma(X_i),

for i=1,…,25i = 1, \ldots, 25. As before, for the Monte Carlo computations of the Shapley value (columns one and three), we used used 100,000 simulations.

References

  1. [1]Bartlett, B.: Hedging under SABR model, Wilmott Magazine, July/August, 2 - 4 (2006).
  2. [2]Cambanis, S., Huang, S., and Simons, G.: On the theory of elliptically contoured distributions, J. Multiv. Anal., 11, 368 - 385 (1981).
  3. [3]Boonen, T. J., De Waegenaere, A., and Norde, A.: A generalization of the Aumann–Shapley value for risk capital allocation problems, Eur. J. Oper. Res., 282, 277 - 287 (2020).DOI
  4. [4]Castro, J., Gomez, D., and Tejada, J.: Polynomial calculation of the Shapley value based on sampling, Comp. & Oper. Res., 36, 1726 - 1730 (2009).
  5. [5]Colini-Baldeschi, R., Scarsini, M., and Vaccari, S.: Variance allocation and Shapley value, Methodol. Comp. Appl. Prob., 20, 919 - 933 (2018).
  6. [6]Drehmann, M., and Tarashev, N.: Measuring the systemic importance of interconnected banks, J. Finan. Intermediation, 22, 586 – 607 (2013).
  7. [7]Hagan, P. S., and Lesniewski, A.: Bartlett’s delta in the SABR model, May, 54 - 61 (2019).
  8. [8]Hagan, P. S., Kumar, D., Lesniewski, A., and Woodward, D. E: Managing smile risk, Wilmott Magazine, September, 84 - 108 (2002).
  9. [9]Hagan, P. S., Kumar, D., Lesniewski, A., and Woodward, D. E.: Universal smiles, Wilmott Magazine, July, 40 - 55 (2016).DOI
  10. [10]Liben-Nowell, D., Sharp, A., Wexler, T. and Woods, K.: Computing Shapley value in super-modular coalitional games, in: Gudmundsson J., Mestre J., and Viglas T. (eds): Computing and Combinatorics, Lecture Notes in Computer Science, 7434, Springer (2012).
  11. [11]Mann, I., and Shapley, L. S.: Values of large games, IV: Evaluating the electoral college by Montecarlo techniques, Rand Corporation (1960).
  12. [12]Maschler, M., Solan, E., and Zamir, S.: Game Theory, Cambridge University Press (2013).
  13. [13]Moehle, N., Boyd, S., and Ang, A.: Portfolio performance attribution via Shapley value, arXiv:2102.05799 (2021).arxiv.org/abs/2102.05799
  14. [14]Mussard, S., and Terraza, V.: The Shapley decomposition for portfolio risk, Appl. Econ. Lett., 15, 713 - 715 (2008).DOI
  15. [15]Ortmann, K. M.: The link between the Shapley value and the beta factor, Decisions Econ. Finan., 39, 311 - 325 (2016).DOI
  16. [16]Owen A. B, and Prieur, C.: On Shapley value for measuring importance of dependent inputs, arXiv:161002080 (2016).arxiv.org/abs/1610.02080
  17. [18]Tarashev, N., Borio, C., and Tsatsaronis, K.: Attributing systemic risk to individual institutions, BIS working paper (2010).
  18. [19]Wang, S.: A set of new methods and tools for enterprise risk capital management and portfolio optimization, Tech. rep. SCOR Reinsurance Company (2002).

Paper details

vlnc:W_ezscvqq7
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 7

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Economics, Econometrics and Finance

Not recorded for this paper

View statement
Looking for related papers…
Contents