This page gathers my current research papers that are either under review, submitted for publication, or in the final stages of preparation. These works reflect ongoing research across my main areas of interest, including risk sharing and collaborative insurance, extreme value theory, actuarial science, machine learning, and financial econometrics.
Abstract
We study extreme trajectories of heavy-tailed mixed causal–noncausal autoregressive processes Φ(L)Ψ(L−1)yt = εt, with a focus on bubble-like episodes and their turning points. Using the spectral tail-process representation of the associated two-sided linear process, we show that, conditional on an extreme event, the normalized trajectory is governed by a discrete epicenter variable N, which identifies the innovation driving the episode. We then introduce the pure noncausal and pure causal components of the spectral tail process. Their behavior reveals a sharp regime switch around the epicenter and yields a simple turningpoint logic: the peak location is determined by N, up to a coefficient-induced shift, and can be detected through the disappearance of one pure component on one side of the boundary. We also derive one-step extreme forecasting results and asymptotic predictive distributions for gross-return ratios under extreme conditioning. To make the theory operational, we propose pure noncausal and pure causal residual plots as diagnostics for fitted MAR models during extreme episodes. The methodology is illustrated by simulations and by an empirical analysis of the detrended weekly VIX series, showing that the proposed diagnostics provide interpretable information on the build-up and unwinding phases of extreme movements.
Central limit theorems for squared increment sums of fractional Brownian fields based on a Delaunay triangulation in 2D
Chenavier, N. and C.Y., Robert
Abstract
An isotropic fractional Brownian field (with Hurst parameter H<1/2) is observed in a family of points in the unit square C=(−1/2,1/2]2% . These points are assumed to come from a realization of a homogeneous Poisson point process with intensity N. We consider normalized increments (resp. pairs of increments) along the edges of the Delaunay triangulation generated by the Poisson point process (resp. pairs of edges within triangles). Central limit theorems are established for the respective centered squared increment sums as N→∞.
Limit theorems for squared increment sums of the maximum of two isotropic fractional Brownian fields under fixed-domain asymptotics
Chenavier, N. and C.Y., Robert
Abstract
We study squared increment sums of the pointwise maximum of two independent and identically distributed isotropic fractional Brownian fields over a fixed two-dimensional domain. The fields are observed at the points of a homogeneous Poisson point process with intensity \(N\), and increments are computed along the edges of the associated Delaunay triangulation. In contrast with the case of a single fractional Brownian field, where centered squared increment sums satisfy a central limit theorem after the usual normalization, the pointwise maximum exhibits a different asymptotic regime. The dominant contribution comes from Delaunay edges located in a shrinking neighborhood of the random interface where the two fractional Brownian fields exchange the role of the maximizer. For Hurst parameter \(H<1/2\), we prove that the properly normalized squared increment sum converges in probability to a deterministic constant times the local time at zero of the difference between the two fields. This shows that the asymptotic behavior is governed by the geometry of the random contact set rather than by Gaussian fluctuation effects. The result provides a key ingredient for fixed-domain asymptotic inference in Brown--Resnick type models based on randomly located observations.
Asymptotic properties of maximum composite likelihood estimators for max-stable Brown-Resnick random fields over a fixed-domain
Chenavier, N. and C.Y., Robert
Abstract
Likelihood-based inference for max-stable random fields is challenging, since finite-dimensional densities are either unavailable in closed form or computationally intractable in moderate to high dimension. Composite likelihood methods, based on low-dimensional marginal densities, therefore provide a natural alternative. In this paper, we study maximum composite likelihood estimation for spatial Brown--Resnick random fields generated by isotropic fractional Brownian fields. We work under fixed-domain asymptotics: a single realization of the max-stable field is observed on an increasingly dense random set of sites, given by a homogeneous Poisson point process. Pairwise and triplewise composite likelihoods are constructed by retaining, respectively, the edges and the triangles of the associated Poisson--Delaunay triangulation. Our main results establish the consistency of the resulting maximum composite likelihood estimators of the scale and smoothness parameters, when the other parameter is known. Their asymptotic behaviour is non-standard: the estimators converge at rates depending on the smoothness parameter and their centered limits are non-Gaussian. More precisely, the limiting fluctuations are driven by aggregated local times associated with the canonical tessellation of the Brown--Resnick field. These results reveal a fundamental departure from the classical composite likelihood theory based on increasing domains or independent replications, and show that Gaussian uncertainty quantification may be misleading in fixed-domain inference for max-stable spatial extremes.
Forecasting with Markovian max-stable fields in space and time: An application to wind gust speeds
Cotsakis , R., E., Koch and C.Y., Robert
Abstract
Hourly maxima of 3-second wind gust speeds are important indicators of wind-storm severity, and their short-lead probabilistic forecasting is relevant for civil protection, infrastructure management, and insurance applications. We develop a parsimonious space-time extreme-value model for such data, based on a max-autoregressive max-stable field with an advection component. The model combines a causal Markovian temporal structure with a spatial Brown–Resnick innovation field, yielding interpretable parameters for temporal persistence, spatial dependence, and storm propagation. This structure also leads to a tractable ensemble forecasting strategy. We study the theoretical properties of the model and establish the strong consistency and asymptotic normality of the associated space-time pairwise likelihood estimator under increasing-domain asymptotics. The methodology is assessed in a retrospective
short-lead forecasting experiment using ERA5 reanalysis data for hourly maxima of 3-second wind gust speeds over Northwestern France during the Lothar storm in December 1999. The fitted model captures the asymmetric space-time correlation structure induced by advection and improves forecasting scores relative to a spacetime Brown–Resnick benchmark.
Tail index regression for heavy-tailed panel data with latent factors
Dutto, F., T., Giroux, E., Koch, O.D., Zerbib and C.Y., Robert
Abstract
Extreme value applications with panel data require methods that can accommodate both cross-sectional heterogeneity and common time variation, while inference remains difficult because tail observations are sparse. This paper studies tail index regression for heavy-tailed panel data when the tail behavior depends on observable covariates and latent common factors. We propose a peaks-over-threshold pseudo-likelihood method in which the tail index is specified through individual-specific regression coefficients and factor loadings. The exceedance thresholds are updated jointly with the model parameters. The procedure is designed for large panels and exploits the generalized Pareto approximation for threshold exceedances together with a latent factor structure that captures residual cross-sectional dependence and unobserved heterogeneity. We develop an asymptotic theory under joint divergence of the cross-sectional and time dimensions. In particular, we establish consistency of the slope and loading estimators, consistency of the latent factor estimates, and, under stronger conditions, uniform consistency and asymptotic normality. An application to U.S. stock returns illustrates the empirical content of the method. The estimated latent tail factor is strongly related to market volatility and equicorrelation. These results provide a tractable econometric framework for studying tail risk comovements in large panels.
Tail cointegration: a positive independent component approach
Gouriéroux, C., and C.Y., Robert
Abstract
This paper introduces tail cointegration as an extreme-value analogue of classical cointegration for multivariate non-negative heavy-tailed risks. Tail cointegrating vectors are defined as admissible linear combinations whose upper tail is asymptotically negligible relative to the common marginal tail. We study this notion in a positive independent component model with independent regularly varying sources and strictly ordered tail indices. In this model, tail cointegration is equivalent to eliminating the exposure to the unique heaviest-tailed latent source while preserving non-negativity of the remaining source exposures. This yields a conic characterization of the tail-cointegration set. We establish identification of the mixing directions from the simplicial support induced by positivity and propose a minimum-volume simplex estimator after radial projection. The reconstructed sources are used to estimate
the latent tail indices by Hill estimators and to recover their ordering by tail heaviness. We derive consistency and asymptotic distribution results for the reconstructed-source Hill estimators and establish consistency and convergence rates for a normalized estimated tailcointegrating vector. Simulations illustrate the finite-sample performance of the geometric estimator, the tail-ordering step, and the resulting tail neutralization.
Hill Random Forests
Maillart, A., and C.Y., Robert
Abstract
Tail index regression is a central problem in extreme value statistics when the heaviness of the tail of a response variable depends on covariates. A major difficulty is that global high-threshold rules may induce a selection bias in the covariate space: exceedances tend to concentrate in regions where the tail is heaviest, making the estimation of the whole tail index function unreliable. This paper introduces a random-forest methodology for conditional tail index estimation based on covariate-adaptive thresholding. In the theoretical analysis, we use an oracle conditional threshold ensuring that exceedances remain asymptotically balanced over the covariate space. Above this threshold, the log-excesses are asymptotically exponential with a covariate-dependent mean equal to the Hill index function. This motivates a new class of Hill regression trees built with a gamma deviance splitting criterion rather than a squared-error criterion. Individual trees are aggregated through a harmonic mean, or equivalently through an arithmetic average of the corresponding tail-index estimators. We establish consistency in integrated deviance for the resulting infinite Hill random forest under regular variation, uniform second-order conditions and suitable tree-growth assumptions. We also discuss practical implementation with estimated conditional thresholds and compare the proposed method with standard regression forests applied to log-excesses. Numerical experiments illustrate the importance of conditional thresholding and the benefits of using a likelihood-based splitting rule adapted to exponential log-excesses.
The balance property: from insurance pricing to risk-sharing
Denuit, M., C.Y., Robert and M.V.W. Wüthrich
Abstract
The balance property is an in-sample global calibration constraint requiring the total fitted actuarial price to be equal to the total observed loss. For generalized linear models (GLMs) fitted by maximum likelihood estimation (MLE), this property holds automatically under the canonical link choice, but it generally fails under non-canonical link choices. Existing balance correction methods include quasi-MLE, simple post-processing of the intercept, and constrained MLE. This note proposes a fourth approach, motivated by the linear regression approximation of the conditional mean risk-sharing (CMRS) rule in large heterogeneous pools. Starting from the correctly specified MLE, the proposed correction allocates the aggregate balance defect according to the fitted conditional variances. We compare the four methods through two complementary asymptotic criteria. First, a delta method expansion gives the first-order mean squared error (MSE) of the fitted pure premium. Under this criterion, the proposed correction shares the same MLE component as the simple post-processing and the constrained MLE corrections, but differs in the way it allocates the balance defect. Second, we derive its average Kullback–Leibler (KL) divergence and show that the varianceweighted correction is locally optimal among additive first-order balance corrections acting directly on the fitted mean function. Thus, the proposed method complements constrained MLE: the latter is optimal within the parametric GLM class, whereas the former is optimal in a larger class of mean function corrections and has a direct risk-sharing interpretation. The paper finally highlights the link between insurance pricing models possessing the balance property and risk-sharing rules. This view reconciles two seemingly disconnected core actuarial topics, ex-ante premium calculation and ex-post loss allocation to participants in a peer-to-peer insurance pool.
Linear risk sharing in community-based insurnce: ruin reduction in the compound Poisson model
Flores-Contro, J.M., M., Denuit, and C.Y., Robert
Abstract
This paper studies proportional risk sharing at claim occurrence time in community-based insurance. Each participant is modeled by an individual Cram´er–Lundberg surplus process, and, whenever a claim is reported within the pool, its cost is redistributed according to a fixed allocation matrix. We compare the infinite-time ruin probability of each participant under stand-alone operation and under pool participation. Our main result shows that pooling reduces, for every participant, the infinite-time ruin probability when claim severities belong to a common scale family, the allocation rule satisfies full allocation and actuarial fairness, and each transfer remains bounded by an individual capacity condition. The proof relies on a convex-order comparison between the losses borne inside the pool and the corresponding stand-alone losses. We also clarify the role of these assumptions by showing that, outside this fram ework, pooling need not be beneficial for all participants. Numerical illustrations with Exponential and LogNormal severities support the theoretical findings and highlight how the design of proportional sharing rules affects solvency. The paper thus provides simple and interpretable sufficient conditions under which transparent linear risk-sharing arrangements improve individual solvency in community-based insurance.
Conditional mean risk sharing with settlement delays in peer-to-peer disability insurance
Denuit, M., and C.Y., Robert
Abstract
This paper develops a dynamic conditional mean risk-sharing mechanism for peer-to-peer disability insurance when claim amounts are not observed at occurrence time. The model is motivated by cooperative arrangements such as Dutch bread funds, where participants mutually protect each other against the economic consequences of disability. We consider a two-state model in which participants move between active and disabled states, and where the ultimate cost of a disability claim is determined by the duration of the disability spell. Since this duration is only known at recovery, the conditional mean risk-sharing allocation cannot be implemented immediately when the claim is reported. The proposed mechanism requires active participants to contribute an initial reserve when a disability spell starts. This reserve is then dynamically updated as information about the claim development becomes available, and it converges at settlement to the terminal conditional mean risk-sharing allocation. We show that the dynamic reserve is a conditional expectation of the corresponding thinned individual loss under the information used by the pool. This representation yields martingale, conditional full-allocation and actuarial fairness properties. Most importantly, the cumulative contribution paid by each participant is smaller in convex order than the corresponding stand-alone incurred-and-reserved disability cost. Hence every risk-averse participant is willing to join the pool. Numerical illustrations show how reserves and final adjustments depend on disability duration and heterogeneous recovery rates.
Funded risk-sharing schemes with proportional shortfall allocation
Denuit, M., and C.Y., Robert
Abstract
This paper proposes a new pre-funded risk-sharing mechanism among economic agents. Individual contributions are deterministic and selected ex ante, while ex-post payoffs depend on realized individual losses and aggregate surplus or deficit at the level of the pool. Full compensation is awarded to claiming participants and surplus is distributed as long as the aggregate losses remain smaller than the accumulated contributions, whereas compensations are reduced proportionally in case of deficit. The paper establishes the existence of actuarially-fair contributions, ensuring that each agent’s contribution equals the corresponding expected payoff. The risk-reduction properties of the proposed mechanism are analyzed with respect to the convex order. Particular attention is devoted to the case where total contributions are equal to the expected aggregate loss. In this case, limiting values of individual contributions are derived as well as the asymptotic distribution of the number of participants who are only partially compensated. Refined results are obtained in the homogeneous case, allowing for a better understanding of the proposed risk-sharing mechanism. Numerical illustrations based on Monte-Carlo methods demonstrate how the theoretical results can be applied in practice.
Quantile-based fully-funded risk-sharing schemes
Denuit, M., and C.Y., Robert
Abstract
This paper studies a fully-funded risk-sharing mechanism for a pool of agents facing heterogeneous stochastic losses. Participants jointly agree on a total funding level and compensations are awarded through a quantile-based rule that endogenously adjusts to realized aggregate losses. The mechanism operates under two regimes: in surplus states, losses are shifted upward through a common quantile increment to provide compensations; in deficit states, an endogenous deductible restores feasibility. We show that the allocation defines a well-posed nonlinear operator characterized by monotonicity and convexity properties of marginal quantile functions. Existence and uniqueness of the endogenous shift parameters and of fair contribution vectors are established under mild distributional assumptions allowing for atoms at zero. A central result reveals a funding phase transition at the critical level equal to the expected aggregate loss. At this level, the mechanism becomes asymptotically mean-preserving as the pool size grows. In deficit and surplus regimes, stabilization results and first-order approximations are derived to characterize the large-scale behavior of individual allocations. In homogeneous large pools, semi-explicit distributional representations are obtained and convergence in law of individual net allocations is established. Finally, under convexity conditions on quantile functions, it is shown that participation improves welfare in second-order stochastic dominance, or increasing-concave order. These results position fully-funded quantile-based sharing as a structurally robust and scalable resource allocation mechanism under aggregate risk constraints.
Reverse-time scores and convex-order monotonicity for additive processes with Lévy clocks
Dendoncker, V., M., Denuit, and C.Y., Robert
Abstract
We study reverse-time conditional means, scores and drift quantities in additive noise models driven by deterministic L´evy clocks. Starting from an integrable signal Y and an independent centered additive process (Zt)t∈[0,T ], we consider the forward noising model Xt = Y + Zt, 0 ≤ t ≤ T, and its reverse-time version Xs = XT−s. The main examples are Gaussian noise with deterministic variance clock, compensated compound Poisson noise with deterministic intensity clock, and symmetric α-stable noise with stable clock, for α ∈ (1, 2). We first derive Bartlett–Tweedie identities for additive noise models and characterize, within the infinitely divisible class, the cases in which the denoising map x → E[Y | Y + ε = x] is affine. This occurs precisely when the signal and the noise are compatible with the same convolution semigroup. We then prove that the reverse conditional mean and the denoising error normalized by the accumulated noise clock are martingales in reverse time. As a consequence, both families are increasing in the convex order along reverse time. The actual reverse drift, or reverse local mean drift in the jump and stable cases, is obtained by multiplying this normalized score by the instantaneous clock rate. It is therefore not a martingale in general, but it defines a peacock when the instantaneous rate is non-decreasing in reverse time. Explicit linear formulas are obtained for semigroup-compatible initial laws. The Cauchy case is treated separately, since the lack of integrability prevents an ordinary convex-order analysis of the reverse score drift.
Convex order monotonicity of conditional mean predictors in latent mixture models
Dendoncker, V., M., Denuit, and C.Y., Robert
Abstract
This paper studies the monotonicity, in the convex order, of conditional mean predictors based on aggregate statistics in latent mixture models. For partial sums Sj = X1 +· · ·+Xj , we give general conditions under which E[Xi | Sj ] is dominated by E[Xi | Sj−1] in the convex order, i < j. Such an ordering means that enlarging the conditioning sum reduces the dispersion, and hence the informativeness, of the corresponding conditional mean predictor. The first set of results is based on a projection identity for conditional expectations and includes the classical case where an independent noise term is added to the signal. We then show that the same convex-order reduction follows from a Markov dependence structure. This provides a tractable route for latent mixture models: when the variables are conditionally independent given a common factor and belong to natural exponential families with a common canonical parameter, the sufficiency of partial sums yields the required Markov property. A complementary analysis characterizes situations in which the conditional mean predictor is affine or proportional to the conditioning sum. In particular, for convolution semigroups associated with infinitely divisible distributions, we identify the latent scaling structures leading to proportional predictors. The results apply to a range of standard models, including Poisson, binomial, negative binomial, Gamma, normal, inverse Gaussian, Tweedie, compound Poisson and tempered stable distributions.