Abstract
A substantial body of work in the last 15 years has shown that expectiles constitute an excellent candidate for becoming a standard tool in probabilistic and statistical modeling. Surprisingly, the question of how expectiles may be efficiently calculated has been left largely untouched. We fill this gap by, first, providing a general outlook on the computation of expectiles that relies on the knowledge of analytic expressions of the underlying distribution function and mean residual life function. We distinguish between discrete distributions, for which an exact calculation is always feasible, and continuous distributions, where a Newton–Raphson approximation algorithm can be implemented and a list of exceptional distributions whose expectiles are in analytic form can be given. When the distribution function and/or the mean residual life is difficult to compute, Monte-Carlo algorithms are introduced, based on an exact calculation of sample expectiles and on the use of control variates to improve computational efficiency. We discuss the relevance of our findings to statistical practice and provide numerical evidence of the performance of the considered methods.
Similar content being viewed by others
Notes
Available on GitHub at https://github.com/AntoineUC/Expectrem.
Available at https://www.kaggle.com/ranja7/vehicle-insurance-customer-data and on file with the authors.
Available in the R package Expectrem as the dataset commerzbank.
References
Artzner, P., Delbaen, F., Eber, J.-M., Heath, D.: Coherent measures of risk. Math. Financ. 9(3), 203–228 (1999)
Asmussen, S., Albrecher, H.: Ruin Probabilities. Advanced Series on Statistical Science and Applied Probability, vol. 14. World Scientific, Singapore (2010)
Bellini, F.: Isotonicity properties of generalized quantiles. Stat. Probab. Lett. 82(11), 2017–2024 (2012)
Bellini, F., Di Bernardino, E.: Risk management with expectiles. Eur. J. Finance 23(6), 487–506 (2017)
Bellini, F., Klar, B., Müller, A., Gianin, E.R.: Generalized quantiles as risk measures. Insurance Math. Econom. 54, 41–48 (2014)
Bellini, F., Klar, B., Müller, A.: Expectiles, omega ratios and stochastic ordering. Methodol. Comput. Appl. Probab. 20(3), 855–873 (2016)
Birnbaum, Z.W.: An inequality for Mill’s ratio. Ann. Math. Stat. 13(2), 245–246 (1942)
Black, F., Scholes, M.: The pricing of options and corporate liabilities. J. Polit. Econ. 81(3), 637–654 (1973)
Breckling, J., Chambers, R.: M-quantiles. Biometrika 75(4), 761–772 (1988)
Brockwell, P.J., Davis, R.A.: Time Series: Theory and Methods, 2nd edn. Springer, Berlin (1991)
Cascos, I., Ochoa, M.: Expectile depth: theory and computation for bivariate datasets. J. Multivar. Anal. 184, 104757 (2021)
Chambers, R., Dreassi, E., Salvati, N.: Disease mapping via negative binomial regression M-quantiles. Stat. Med. 33(27), 4805–4824 (2014)
Chambers, R., Salvati, N., Tzavidis, N.: Semiparametric small area estimation for binary outcomes with application to unemployment estimation for local authorities in the UK. J. R. Stat. Soc. Ser. A 179(2), 453–479 (2016)
Daouia, A., Girard, S., Stupfler, G.: Estimation of tail risk based on extreme expectiles. J. R. Stat. Soc. B 80(2), 263–292 (2018)
Daouia, A., Girard, S., Stupfler, G.: Tail expectile process and risk assessment. Bernoulli 26(1), 531–556 (2020)
Daouia, A., Girard, S., Stupfler, G.: ExpectHill estimation, extreme risk and heavy tails. J. Econom. 221(1), 97–117 (2021)
Daouia, A., Stupfler, G., Usseglio-Carleve, A.: Inference for extremal regression with dependent heavy-tailed data. Ann. Stat. 51(5), 2040–2066 (2023)
Dawber, J., Salvati, N., Fabrizi, E., Tzavidis, N.: Expectile regression for multicategory outcomes with application to small area estimation of labour force participation. J. R. Stat. Soc.: Ser. A 185(Supplement_2), 590–619 (2022)
de Haan, L., Ferreira, A.: Extreme Value Theory: An Introduction. Springer, New York (2006)
del Castillo, J., Soler, D.M., Serra, I.: Ercv: Fitting Tails by the Empirical Residual Coefficient of Variation. R package version 1.0.1 (2019). https://CRAN.R-project.org/package=ercv
Dickson, D.C.: Insurance Risk and Ruin. Cambridge University Press, Cambridge (2016)
Eberl, A., Klar, B.: Stochastic orders and measures of skewness and dispersion based on expectiles. Stat. Pap. 64(2), 509–527 (2023)
Francq, C., Zakoïan, J.-M.: Maximum likelihood estimation of pure GARCH and ARMA-GARCH processes. Bernoulli 10(4), 605–637 (2004)
Galanos, A., Kley, T.: Rugarch: Univariate GARCH Models. R package version 1.4-9 (2022). https://CRAN.R-project.org/package=rugarch
Girard, S., Stupfler, G., Usseglio-Carleve, A.: Extreme conditional expectile estimation in heavy-tailed heteroscedastic regression models. Ann. Stat. 49(6), 3358–3382 (2021)
Girard, S., Stupfler, G., Usseglio-Carleve, A.: Nonparametric extreme conditional expectile estimation. Scand. J. Stat. 49(1), 78–115 (2022a)
Girard, S., Stupfler, G., Usseglio-Carleve, A.: On automatic bias reduction for extreme expectile estimation. Stat. Comput. 32(4), 64 (2022b)
Gneiting, T.: Making and evaluating point forecasts. J. Am. Stat. Assoc. 106(494), 746–762 (2011)
Hankin, R.K.S., Clausen, A., Murdoch, D.: Gsl: Wrapper for the Gnu Scientific Library. R package version 2.1-8 (2023). https://cran.r-project.org/package=gsl
Heston, S.L.: A closed-form solution for options with stochastic volatility with applications to bond and currency options. Rev. Financ. Stud. 6(2), 327–343 (1993)
Holzmann, H., Klar, B.: Expectile asymptotics. Electron. J. Stat. 10(2), 2355–2371 (2016)
Jones, M.C.: Expectiles and M-quantiles are quantiles. Stat. Probab. Lett. 20(2), 149–153 (1994)
Koenker, R.: When are expectiles percentiles? Econom. Theor. 9(3), 526–527 (1993)
Koenker, R.: Quantile Regression. Cambridge University Press, Cambridge (2005)
Krätschmer, V., Zähle, H.: Statistical inference for expectile-based risk measures. Scand. J. Stat. 44(2), 425–454 (2017)
Kuan, C.-M., Yeh, J.-H., Hsu, Y.-C.: Assessing value at risk with care, the conditional autoregressive expectile models. J. Econom. 150(2), 261–270 (2009)
Laudagé, C., Turkalj, I.: Calculating Expectiles and Range Value-at-Risk using Quantum Computers. arXiv:2211.04456 (2022)
Mező, I., Baricz, Á.: On the generalization of the Lambert W function. Trans. Am. Math. Soc. 369(11), 7917–7934 (2017)
Newey, W.K., Powell, J.L.: Asymmetric least squares estimation and testing. Econometrica 55(4), 819–847 (1987)
Otto-Sobotka, F., Spiegel, E., Schnabel, S., Schulze Waltrup, L., Eilers, P., Kauermann, G., Kneib, T.: Expectreg: Expectile and Quantile Regression. R package version 0.52 (2022). https://CRAN.R-project.org/package=expectreg
Padoan, S., Stupfler, G.: ExtremeRisks: Extreme Risk Measures. R package version 0.0.4 (2020). https://cran.r-project.org/package=ExtremeRisks
Racine, J.S., Hayfield, T.: Np: Nonparametric Kernel Smoothing Methods for Mixed Data Types. R package version 0.60-17 (2023). https://CRAN.R-project.org/package=np
Sobotka, F., Kneib, T.: Geoadditive expectile regression. Comput. Stat. Data Anal. 56(4), 755–767 (2012)
Sobotka, F.O., Kneib, T., Salvati, N., Ranalli, M.G.: Adaptive semiparametric M-quantile regression. Econom. Stat. 11, 116–129 (2019)
Stupfler, G., Usseglio-Carleve, A.: Composite bias-reduced \(L^p-\)quantile-based estimators of extreme quantiles and expectiles. Can. J. Stat. 51(2), 704–742 (2023)
Taylor, J.W.: Estimating value at risk and expected shortfall using expectiles. J. Financ. Econom. 6(2), 231–252 (2008)
Wuertz, D., Chalabi, Y., Setz, T., Maechler, M., Boudt, C., Chausse, P., Miklovac, M., Boshnakov, G.N.: fGarch: Rmetrics—Autoregressive Conditional Heteroskedastic Modelling. R package version 4031.90 (2023). https://CRAN.R-project.org/package=fGarch
Yee, T., Moler, C.: VGAM: Vector Generalized Linear And Additive Models. R package version 1.1-9. (2023) https://cran.r-project.org/package=VGAM
Zhu, L., Dong, Y., Li, R.: Semiparametric estimation of conditional heteroscedasticity via single-index modeling. Stat. Sin. 23(3), 1235–1255 (2013)
Ziegel, J.F.: Coherence and elicitability. Math. Financ. 26(4), 901–918 (2016)
Zou, H.: Generalizing Koenker’s distribution. J. Stat. Plan. Inference 148, 123–127 (2014)
Acknowledgements
The authors acknowledge two anonymous reviewers for their helpful comments that led to an improved version of this article, as well as Mikael Escobar-Bach for a useful chat about bounds for hazard ratios. This research was supported by the French National Research Agency under the Grants ANR-19-CE40-0013 (ExtremReg project), ANR-17-EURE-0010 (EUR CHESS) and ANR-11-LABX-0020-01 (Centre Henri Lebesgue). A. Daouia and G. Stupfler acknowledge financial support from the TSE-HEC ACPR Chair “Regulation and systemic risks”, and G. Stupfler acknowledges further support from the Chair Stress Test, RISK Management and Financial Steering of the Foundation Ecole Polytechnique.
Author information
Authors and Affiliations
Contributions
All authors contributed equally to the work, and read and approved the final manuscript.
Corresponding author
Ethics declarations
Conflict of interest
The authors declare that they have no conflict of interest.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
Proofs of the theoretical results
Proof of Theorem 2.1
Clearly
where \(\psi (x)=\int _{{\mathbb {R}}} t \mathbb {1}_{\{ t>x \}} \, \mu (\textrm{d}t)\) and \(\overline{F}(x)=1-F(x)\) with \(F(x)=\mu ((-\infty ,x])\). Consequently, for any \(i\in I\) and \(x\in [a_i,a_{i+1})\),
where \(m=\int _{{\mathbb {R}}} x \, \mu (\textrm{d}x)=\sum _{k\in I} p_k a_k\). In other words, the function \(g_{\tau }\) is continuous, piecewise linear and decreasing, and tends to \(+\infty \) (resp. \(-\infty \)) as \(x\rightarrow -\infty \) (resp. \(x\rightarrow +\infty \)). It follows that there is a unique index \(i=i(\tau )\) such that the two inequalities
hold. With this index i, the expectile \(\xi _{\tau }\) is the unique root of the linear function \(g_{\tau }\) on the interval \([a_i,a_{i+1})\), namely:
Now \(\psi (a_i)=\sum _{k>i} p_k a_k\) and \(\overline{F}(a_i)=\sum _{k>i} p_k\), so that, for any i,
The above pair of inequalities is therefore equivalent to
as announced. The two identities involving \(\xi _{\tau }\) follow immediately from (A1). \(\square \)
Proof of Theorem 2.2
Define a continuous function \(G=G_{\tau }\) by setting
Recall that \(g_{\tau }:x\mapsto \frac{2\tau -1}{1-\tau }\varphi (x) + m-x\) is convex. It is also clear that the derivative \(g_{\tau }'\) is negative. In particular, \(g_{\tau }\) is decreasing, and since \(\xi _{\tau }\) is the unique solution of the equation \(g_{\tau }(x)=0\), one has \(g_{\tau }(x_0)>0\) for any \(x_0<\xi _{\tau }\). Then
using the convexity property of \(g_{\tau }\). It follows that for any \(x_0<\xi _{\tau }\), \(x_0<G(x_0)\le \xi _{\tau }\), and therefore, for any starting point \(x_0<\xi _{\tau }\), the Newton–Raphson sequence of iterates \((x_n)\) is nondecreasing and bounded and hence convergent. The limit must be a root of \(g_{\tau }\) by taking limits in the equation \(x_{n+1}=G(x_n)\), meaning that \((x_n)\) converges to \(\xi _{\tau }\).
A Taylor expansion of \(g_{\tau }\) on the interval \([x_n,\xi _{\tau }]\subset [x_0,\xi _{\tau }]\) (on which \(g_{\tau }\) is twice continuously differentiable) with remainder in integral form entails
This is readily rewritten as
Then
for any \(k<n\). The proof is complete. \(\square \)
Proof of Theorem 2.3
With the notation of the proof of Theorem 2.2, the function G is twice continuously differentiable in a neighborhood of \(\xi _{\tau }\) with
In particular, \(G'(\xi _{\tau })=0\) and
Use a Taylor expansion to get, for any n,
for some \(c_n\in \left]x_n,\xi _{\tau }\right[\). It follows from Theorem 2.2 that \(c_n\rightarrow \xi _{\tau }\) and therefore, using the continuity of \(G''\),
as required. \(\square \)
Proof of Theorem 2.4
From the last chain of inequalities in the proof of Theorem 2.2, we get, for any n,
Take then \(c>0\) sufficiently large and write
The function \(\eta \) is continuous on \([c,\infty )\) and converges to 0 at infinity. By the representation theorem for regularly varying functions (see Theorem B.1.6 p.365 in de Haan and Ferreira 2006), \(\overline{F}\) is regularly varying with index \(-1/\gamma \), and then the von Mises condition ensures that f is regularly varying with index \(-1/\gamma -1\). By the uniform convergence theorem for regularly varying functions (see Theorem B.1.4 p.363 in de Haan and Ferreira 2006),
Moreover (see Bellini et al. 2014; Daouia et al. 2018)
Conclude from these two convergences that
The proof is complete. \(\square \)
Proof of Theorem 3.1
We first prove the asymptotic normality statement, and for this, it is enough to show that \(\widehat{\Sigma }_{12,n}/\widehat{\sigma }_n^2\rightarrow {\Sigma _{12}/\sigma ^2=\Sigma _{12}/\Sigma _{22}}\) in probability. By the law of large numbers, \(\widehat{\sigma }_n^2\rightarrow \sigma ^2\) in probability, so it suffices to show that \(\widehat{\Sigma }_{12,n}\rightarrow \Sigma _{12}\) in probability. Finally, since \(\widehat{\xi }_{\tau ,n}\rightarrow \xi _{\tau }\) in probability (for example by Theorem 2 in Holzmann and Klar 2016), it is enough to prove that \(\widehat{\overline{F}}_n(\widehat{\xi }_{\tau ,n})\rightarrow \overline{F}(\xi _{\tau })\) and \(\widehat{\varphi }_n^{(2)}(\widehat{\xi }_{\tau ,n})\rightarrow \varphi ^{(2)}(\xi _{\tau })\) in probability.
Fix \(\varepsilon >0\). Then \(\widehat{\xi }_{\tau ,n}\in [\xi _{\tau }-\varepsilon ,\xi _{\tau }+\varepsilon ]\) with arbitrarily high probability as \(n\rightarrow \infty \). Then clearly
with arbitrarily high probability as \(n\rightarrow \infty \). The upper bound converges in probability to \(\mu ([\xi _{\tau }-\varepsilon ,\xi _{\tau }+\varepsilon ])\), by the law of large numbers, which is arbitrarily small as \(\varepsilon \downarrow 0\). Conclude that \(\widehat{\overline{F}}_n(\widehat{\xi }_{\tau ,n}) = (\widehat{\overline{F}}_n(\widehat{\xi }_{\tau ,n})-\widehat{\overline{F}}_n(\xi _{\tau }))+\widehat{\overline{F}}_n(\xi _{\tau })\rightarrow \overline{F}(\xi _{\tau })\) in probability, by the law of large numbers again. We now prove that \(\widehat{\varphi }_n(\widehat{\xi }_{\tau ,n})=\widehat{\varphi }_n^{(1)}(\widehat{\xi }_{\tau ,n})\rightarrow \varphi ^{(1)}(\xi _{\tau })=\varphi (\xi _{\tau })\) before turning to the convergence of \(\widehat{\varphi }_n^{(2)}(\widehat{\xi }_{\tau ,n})\). Write
to obtain
with arbitrarily high probability as \(n\rightarrow \infty \). The upper bound is arbitrarily small as \(\varepsilon \downarrow 0\), so that the law of large numbers yields \(\widehat{\varphi }_n^{(1)}(\widehat{\xi }_{\tau ,n})\rightarrow \varphi ^{(1)}(\xi _{\tau })\) in probability. Finally
so that, with arbitrarily high probability as \(n\rightarrow \infty \),
Conclude in a similar fashion that \(\widehat{\varphi }_n^{(2)}(\widehat{\xi }_{\tau ,n})\rightarrow \varphi ^{(2)}(\xi _{\tau })\) in probability, as required. The fact that \(\widetilde{\xi }_{\tau ,n}\) has the lowest asymptotic variance among all asymptotically unbiased linear combinations of \(\widehat{\xi }_{\tau ,n}\) and \(\overline{X}_n-m\) is then obvious since
has this property.
It remains to prove the assertions about the variance reduction factor \(1-R(\tau ,\mu )\). This function is clearly zero at \(\tau =1/2\). Note also that
As a result, and since \(\tau \mapsto \xi _{\tau }\) is continuously differentiable on \(I=(\tau _1,\tau _2)\) (see Proposition 1(iii) in Holzmann and Klar 2016), the function \(\tau \mapsto 1-R(\tau ,\mu )\) is continuously differentiable on this interval. It remains to prove the statements about monotonicity. Set
so that \(R(\tau ,\mu )=(u(\tau ))^2/(\sigma ^2 v(\tau ))\) and therefore
Writing \(u(\tau )=(1-\tau ) {\mathbb {E}}((X-\xi _{\tau })^2 \mathbb {1}_{\{ X<\xi _{\tau } \}}) + \tau {\mathbb {E}}((X-\xi _{\tau })^2 \mathbb {1}_{\{ X>\xi _{\tau } \}})\) yields in particular that \(u(\tau )>0\) for any \(\tau \), meaning that the partial derivative \(\frac{\partial R}{\partial \tau }(\tau ,\mu )\) has the same sign as \(2 u'(\tau ) v(\tau ) - v'(\tau ) u(\tau )\). Now
Straightforward calculations yield
This quantity is positive on (0, 1/2) and negative on (1/2, 1) because \(\tau \mapsto \xi _{\tau }\) is strictly increasing (see Proposition 1(ii) in Holzmann and Klar 2016) and \(\xi _{1/2}=m\). The proof is complete. \(\square \)
Proof of Theorem 3.2
Let , where c is to be found so as to minimize the relative asymptotic variance of . Write
Applying Proposition 3.1 entails that the desired value of c satisfies
Apply Proposition 1(i) in Daouia et al. (2020) to obtain
leading to a choice of c minimizing the relative asymptotic variance of as
The statement on the asymptotic normality of with this choice of c is then immediate. Finding the maximum of \(\lambda \mapsto C(\gamma ,\lambda )\) is done by noting that this function is decreasing past \(1/\gamma -1\), and
Maximizing this function over \((0,1/\gamma -1)\) is straightforward and leads to the value \(\lambda ^{\star }\) specified in the statement of Theorem 3.2. The last asymptotic normality result follows by plugging \(\lambda ^{\star }\) into \(C(\gamma ,\lambda )\). \(\square \)
Detailed calculations related to the examples
Distribution supported on a set with three elements (Example 2.2)
Let \(\mu \) be the probability distribution on a set \(\{a,b,c\}\) with \(a<b<c\) characterized by \(\mu (\{b\})=p\) and \(\mu (\{c\})=q\), with \(p,q>0\) and \(p+q<1\). Then, from Corollary 2.1,
and
In particular, for the distribution \(\mu \) on \(\{0,1,2\}\) with \(\mu (\{1\})=p\) and \(\mu (\{2\})=q\),
Taking \(p=p_1(1-p_2)+p_2(1-p_1)\) and \(q=p_1 p_2\), for \(p_1,p_2\in (0,1)\), yields the expectile of the sum of two independent random variables having Bernoulli distributions with parameters \(p_1\) and \(p_2\):
Uniform distribution on \(\{1,\ldots ,n\}\) (Example 2.3)
Fix \(n\ge 2\). For the uniform distribution on \(\{1,\ldots ,n\}\), solving the inequalities of Corollary 2.2 is equivalent to finding the unique index \(i\in \{1,\ldots ,n-1\}\) such that
This is equivalent to finding the unique solution (which we already know to exist, by Corollary 2.2) to the inequalities \(P_{\tau }(i+1)<0\le P_{\tau }(i)\) for \(i\in \{1,\ldots ,n-1\}\), where \(P_{\tau }\) is the polynomial
This polynomial has discriminant \(4\tau (1-\tau )(n+1)(n-1)+1>0\) and then (for \(\tau \ne 1/2\)) two real roots \(x_{\tau ,-}\) and \(x_{\tau ,+}\) defined as
A straightforward calculation yields \(P_{\tau }(1)=\tau n(n-1)>0\) and \(P_{\tau }(n)=-(1-\tau )n(n-1)<0\). It follows that when \(\tau >1/2\) (resp. \(\tau <1/2\)), only the lowest (resp. largest) of the two roots \(x_{\tau ,-}\) and \(x_{\tau ,+}\) belongs to the interval [1, n]. Conclude that, in both cases, \(P_{\tau }(i+1)<0\le P_{\tau }(i) \Leftrightarrow i\le x_{\tau ,-}<i+1 \Leftrightarrow i=\lfloor x_{\tau ,-} \rfloor \). With this index i,
Consequently
with
Geometric distribution (Example 2.4)
For the geometric distribution with success probability \(p\in (0,1)\), namely, \(\mu (\{k\})=p(1-p)^{k-1}\) for any positive integer k, the inequalities of Theorem 2.1 read as
Solving these inequalities is equivalent to finding the index \(i\ge 1\) such that \(h_{\tau }(i+1)<0\le h_{\tau }(i)\), where
Straightforward calculations reveal that the unique root \(x_{\tau }\) of \(h_{\tau }\) over \([1,+\infty )\) satisfies the equation
This is a transcendental equation (unless \(\tau \ne 1/2\), for which \(x_{1/2}=1/p\)). Nevertheless, since by construction
for any \(p\in (0,1)\), and
for any \(\tau ,p\in (0,1)\), one may express \(x_{\tau }\) using the main branch of Lambert’s W function, that is
where, for \(x>0\), W(x) is the unique (positive) solution to the equation \(w e^w=x\). The main branch of the Lambert function is available numerically in R using (for instance) the gsl package (Hankin et al. 2023), acting as a wrapper for the GNU Scientific Library.
Note further that when \(\tau >1/2\), the function \(h_{\tau }\) is obviously decreasing, so the inequalities \(h_{\tau }(i+1)<0\le h_{\tau }(i)\) are equivalent to \(i\le x_{\tau }<i+1\), i.e. \(i=\lfloor x_{\tau } \rfloor \). When \(\tau <1/2\), it is readily shown that \(h_{\tau }'\) is decreasing and \(h_{\tau }'(1)=-(1-2\tau ) (1-p)\log (1-p)-(1-\tau )p<0\) for any \(p\in (0,1)\), so again \(h_{\tau }\) is decreasing and \(h_{\tau }(i+1)<0\le h_{\tau }(i) \Leftrightarrow i=\lfloor x_{\tau } \rfloor \). Conclude that
Cardano and Ferrari formulae (relevant to Sect. 2.3.1)
To solve a real cubic polynomial equation of the form \(x^3+bx^2+cx+d=0\), Cardano’s method consists in letting \(p=c-b^2/3\) and \(q=d+b(2b^2-9c)/27\), and in computing the discriminant \(\Delta =-4p^3-27q^2\) of the so-called depressed cubic \(X^3+pX+q=0\). If \(\Delta \le 0\), then the unique real root of the equation is given by
If on the contrary \(\Delta >0\), then necessarily \(p<0\) and there are 3 real solutions, given by Viète’s formula:
\(k \in \{ 0,1,2 \}\). To solve a real quartic polynomial equation \(x^4+bx^3+cx^2+dx+e=0\), Ferrari’s method first finds a root \(\lambda \) to the cubic equation \(8 \lambda ^3 -4c \lambda ^2+(2bd-8e) \lambda -b^2 e+4ce-d^2=0\). The four (possibly complex) solutions to the quartic equation are then
where \(\varepsilon _1, \varepsilon _2 \in \{ -1,1 \}\).
Student distribution with \(\nu =4\) degrees of freedom (Example 2.13)
Consider the Student distribution with \(\nu \) degrees of freedom, having probability density function
Here \(\Gamma \) is Euler’s Gamma function. When \(\nu =4\), the probability density function simplifies to
The change of variables \(t=2\tan (\theta )\) combined with the trigonometric identities \(\cos (3\phi )=4\cos ^3(\phi )-3\cos (\phi )\), \(\sin (3\phi )=3\sin (\phi )-4\sin ^3(\phi )\) and \(\sin (\arctan \theta )=\theta /\sqrt{1+\theta ^2}\) then yield, after straightforward calculations, the following closed form for the survival function:
Further straightforward calculations based on the change of variables \(u=t^2\) then provide
Since the Student distribution is centered, \(m=0\) and Eq. (3) is
This is a biquadratic equation, leading to \( \xi _{\tau }^2 = -2+1/\sqrt{\tau (1-\tau )} \) because \(\xi _{\tau }^2\ge 0\), and then
In general, the distribution function and mean residual life function of the Student distribution involve the hypergeometric function. It is not hard to see that, while the distribution function and mean residual life function can in fact be written in closed form when \(\nu \) is an even integer, resulting in a polynomial equation characterizing \(\xi _{\tau }\), only the cases \(\nu \in \{2,4,6\}\) result in an equation of degree 4 or lower.
Fisher distribution with (4, 4) degrees of freedom (Example 2.14)
The Fisher distribution with degrees of freedom \(\nu _1>0\) and \(\nu _2>0\) has density function
\(x>0\), where B is the Beta function. In the specific case \(\nu _1=\nu _2=4\), one finds \(\varphi (x)=(3x+2)/(x+1)^2\) for \(x>0\), and \(m=2\). Equation (3) is thus equivalent to the cubic equation
The discriminant of this equation is \(\Delta =108 \tau ^2 (2 \tau -1)/(1-\tau )^3\). If \(\tau \le 1/2\), then \(\Delta \le 0\) and the unique solution is
If now \(\tau > 1/2\), then \(\Delta > 0\) and the 3 possible solutions are
\(k \in \{ 0,1,2 \}\). Since \(\tau > 1/2\), \(\arccos ( \sqrt{(1-\tau )/\tau } )\in [0,\pi /2]\), and therefore (taking the constraint \(\xi _{\tau } \ge 0\) into account) \(k=0\) is the only admissible solution, namely
Pareto distribution with extreme value index 1/4 (Example 2.15)
The Pareto distribution with extreme value index \(\gamma >0\) has survival function \(\overline{F}(x)=x^{-1/\gamma }\) for \(x>1\). This distribution has a finite first moment when \(\gamma <1\), and since \(\varphi (x)=\gamma x^{1-1/\gamma }/(1-\gamma )\) for \(x>1\) and \(m=1/(1-\gamma )\), Eq. (3) leads to
When \(\gamma =1/4\), this is the quartic equation \(\xi _{\tau }^4+b \xi _{\tau }^3 + c \xi _{\tau }^2 + d \xi _{\tau }+ e=0\), where \(b=-4/3\), \(c=0\), \(d=0\) and \(e=(1-2 \tau )/(3(1-\tau ))\). Ferrari’s method leads to finding \(\lambda =\lambda _{\tau }\) which is a root of the cubic equation
The discriminant of this equation is
for all \(\tau \in (0,1)\), from which the unique solution of the equation involving \(\lambda \) is
Ferrari’s method yields four possible solutions. The only real-valued solution greater than 1 is obtained with \(\varepsilon _1=\varepsilon _2=1\), leading to the solution
Catalog of expectile functions of continuous distributions
This section provides a catalog of expectile functions of continuous distributions, obtained either via numerical means or, in exceptional cases, in closed or analytic form. Tables 3, 4, 5 and 6 give reference values for the expectiles of the standard Gaussian, log-normal, Student and chi-squared distributions, respectively. Figures 10, 11, 12 and 13 provide graphical representations of the corresponding expectile functions over (0, 1). Table 7 gathers closed-form expressions for expectiles of certain bounded continuous distributions. Table 8 gives analytic-form expressions for expectiles of some unbounded continuous distributions. Tables 9 and 10 list closed-form expressions for expectiles of the Hall–Weiss distribution and the Pareto distribution, respectively, with particular parameters.
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Daouia, A., Stupfler, G. & Usseglio-Carleve, A. An expectile computation cookbook. Stat Comput 34, 103 (2024). https://doi.org/10.1007/s11222-024-10403-x
Received:
Accepted:
Published:
DOI: https://doi.org/10.1007/s11222-024-10403-x