Ecodep Calendar, click the event to get the Zoom link.
Talks recordeded on the Ecodep Youtube Channel.
- 5/5/2021, 14:00 Kamila Kare (SAMM, Paris 1, Panthéon-Sorbonne) Data Driven Model Selection for Same-Realization Predictions in Autoregressive Processes. Slides.
Abstract. This paper is about the one-step ahead prediction of the future of observations drawn from an infinite-order autoregressive AR(∞) process.It aims to design penalties (completely data driven) ensuring that the selected model verifies the efficiency property but in the non asymptotic framework. We present an oracle inequality with a leading constant equal to one. Moreover, we also show that the excess risk of the selected estimator enjoys the best bias-variance trade-off over the considered collection. To achieve these results, we needed to overcome the dependence difficulties by following a classical approach which consists in restricting to a set where the empirical covariance matrix is equivalent to the theoretical one. We show that this event happens with probability larger than 1-c0/n3 with c0>0. The proposed data driven criteria are based on the minimization of the penalized criterion akin to the Mallows's Cp. Monte Carlo experiments are performed to highlight the obtained results.
Reference. K Kare (2021). Data Driven Model Selection for Same-Realization Predictions in Autoregressive Processes. Hal Preprint.
- 4/28/2021, 14:00 Alexander Kreiss (KU Leuven) Non-Parametric Modelling of Interactions Among Vertices in Dynamic Networks. Slides.
Abstract. We will consider dynamic networks in which the vertices (the actors) can interact with each other along the edges of the network. We assume that over the observation period [0,T] the number of vertices remains fixed while the edges between them may change randomly over time. The occurrence of interactions between the actors is modelled by specifying a Cox-Type model which allows for additional, time-varying covariates. Our interest lies in non-parametrically estimating the (possibly) time-varying effect of the covariates on the interactions. To this end, we introduce a kernel-based local likelihood estimator and study its asymptotic (as the network grows) performance. Moreover, we introduce two test statistics which evaluate the fit of the non-parametric compared to parametric models. From a theoretical point of view a particular challenge when handling this type of data is that neighboring actors in the network influence each other and cannot be treated as independent. We introduce therefore weak dependence measures on dynamic networks based on correlation, mixing and temporal m-dependence. The results are illustrated on bike sharing data.
This is partially joint work with Enno Mammen (Heidelberg) and Wolfgang Polonik (UC Davis).
References. A Kreiß, E Mammen, W Polonik (2019) Nonparametric inference for continuous-time event counting and link-based dynamic network models. https://doi.org/10.1214/19-EJS1588.
A Kreiß (2019) Correlation bounds, mixing and m-dependence under random time-varying network distances with an application to Cox-Processes. https://arxiv.org/abs/1906.03179.
A Kreiß, E Mammen, W Polonik (2021) Testing For a Parametric Baseline-Intensity in Dynamic Interaction Networks. https://arxiv.org/abs/2103.14668.
- 4/7/2021, 14:00 Diu Tran (University of Jyväskylä, Helsinki) Statistical inference for Vasicek-type model driven by Hermite processes. Slides.
Abstract Let Z denote a Hermite process of order q >= 1 and self-similarity parameter H ∈ (1/2, 1). This process is H-self-similar, has stationary increments and exhibits long-range dependence. When q = 1, it corresponds to the well-known fractional Brownian motion, whereas it is not Gaussian as soon as q >= 2. In the talk, we deal with a Vasicek-type model driven by Z, of the form dXt = a(b − Xt)dt + dZt. This model includes the fractional Vasicek model and Hermite-driven Ornstein-Uhlenbeck process. Here, a > 0 and b ∈ R are considered as unknown drift parameters. We provide estimators for a and b based on continuous-time observations. For all possible values of H and q, we prove strong consistency and we analyze the asymptotic fluctuations. This is a first step to estimate parameters of a stochastic equation driven by a Hermite process. Joint work with Prof. Ivan Nourdin from University of Luxembourg.
Reference. I Nourdin, D Tran (2019): Statistical inference for Vasicek-type model driven by Hermite processes. Stoch. Proc. Appl., 129, no. 10, pp. 3774-3791. ArXiv.
- 3/31/2021, 15:00 Frederic Barraquand (IMB Bordeaux) Inferring species interactions using Granger causality and convergent cross mapping. Slides.
Abstract. How to reliably infer interactions between species from time series of their population densities is a long-standing goal of statistical ecology. Usually this inference is done using multivariate (linear) autoregressive models, defining interactions through Granger causality: x causes y whenever x helps predicting future y values. However, the entangled nature of nonlinear ecological systems has suggested an alternative causal inference method based on attractor reconstruction, convergent cross mapping, which is increasingly popular in ecology. Here, we compare the two methods. They uncover interactions with surprisingly similar performance for predator-prey cycles, 2-species chaotic or stochastic competition, as well as 10- and 20-species networks. Thus, contrary to intuition, linear Granger causality remains useful to infer interactions in highly nonlinear ecological networks. We conclude on inevitable similarities between Granger-causal methods and convergent cross mapping due to interaction definitions, and provide suggestions to improve many-species interaction inference.
Reference. F Barraquand, C Picoche, M Detto, F Hartig (2019). Inferring species interactions using Granger causality and convergent cross mapping. https://arxiv.org/abs/1909.00731.
- 3/24/2021, 14:00 Benjamin Poignard (Riken AIP, Osaka) An introduction to sparsity: modelling, properties and applications. Slides.
Abstract. The application domains of sparse modelling have been substantially widened by the availability of high-dimensional data. In particular, high-dimensional statistical modelling is concerned with the significantly large number of parameters to estimate. To tackle the over-fitting issue, penalised/regularized estimation methods have been gaining much attention. In this talk, I will introduce the concept of sparsity together with the standard penalisation methods for sparse modelling and the implications in terms of statistical properties. To illustrate the relevance of sparse modelling, I will present some applications to models that typically suffer from the so-called curse of dimensionality.
References. B Poignard, J-D Fermanian (2021). High-dimensional penalized arch processes. Econometric Reviews Volume 40, 2021 - Issue 1.
B Poignard, M Asai (2020). A Penalised OLS Framework for High-Dimensional Multivariate Stochastic Volatility Models. Papers In Economics & Business, Discussion Paper 20-02.
- 3/10/2021, 16:00 Julien Randon-Furling (Paris 1, Panthéon Sorbonne) Convex Hulls of Random Walks. Slides.
Abstract. This talk will cover a range of results on the convex hull of random walks in the plane and in higher dimension: expected perimeter length in the planar case, expected number of faces on the boundary, expected d-dimensional volume, and other geometric properties of such random convex polytopes. Applications in ecology include estimations of animals' home ranges and minimal habitat sizes in conservation parks.
References. J Randon-Furling, D Zaporozhets (2020). Convex hulls of several multidimensional Gaussian random walkls. arXiv:2007.02768.
J Randon-Furling, F Wespi (2017). Facets on the convex hull of d-dimensional Brownian and Lévy motion.. Physical Review E.
SN Majumdar, A Comtet, J Randon-Furling (2010). Random convex hulls and extreme value statistics. Journal of Statistical Physics.
- 1/27/2021, 15:00 Benjamin Bobbia (CYU & Univ. Franche-Comté) Extreme quantile regression: A coupling approach and Wasserstein distance. Slides.
Abstract. In this work, we develop two coupling approaches for extreme quantile regression. We consider i.i.d copies of Y in R and X in R^d and we want an estimation of the conditional quantile of Y given X = x of order 1- alpha for a very small alpha > 0.
We introduce the proportional tail model, strongly inspired by the heteroscedastic extremes developed by Einmahl, de Haan and Zhou. The main assumption is that the tail distribution of Y is asymptotically proportional to the conditional tail of Y given X = x. We propose and study estimators of both model parameters and conditional quantile, which are studied by coupling methods.
References. B Bobbia, C Dombry, D Varron (2020). The coupling method in extreme value theory. https://arxiv.org/pdf/1912.03155
B Bobbia, C Dombry, D Varron (2020). Extreme quantile regression in a proportional tail framework. https://arxiv.org/pdf/2002.01740
- 12/2/2020, 16:00 Rolando Rebolledo (University of Valparaiso) Open-system approach to ecological networks. Abstract, Talk, and Slides.
Reference. R Rebolledo, SA Navarrete, S Kéfi, S Rojas, PA Marquet. An Open-System Approach to Complex Biological Networks. SIAM Journal on Applied Mathematics, 79(2):619–640, 2019.
- 11/25/2020, 16:00, Félix Cheysson (Agro-Paristech Paris) Properties of Hawkes processes. Talk, Slides.
Abstract. Hawkes processes are a family of stochastic processes for which the occurrence of any event increases the probability of further events occurring. When count data are only observed in discrete time, we propose a spectral approach for the estimation of Hawkes processes, by means of Whittle's parameter estimation method. To get asymptotic properties for the estimator, we prove alpha-mixing properties for the series of counts, using the Galton-Watson properties of the cluster representation of Hawkes processes. Simulated datasets and an application to the incidence of measles in France illustrate the performances of the estimation, notably of the Hawkes excitation function, even when the time between observations is large.
- 11/18/2020, 16:00, Marc Lavielle (Inria & CMAP, Polytechnique) Modelling the COVID 19 pandemic requires a model... but also data! Talk, Slides.
Abstract. I will present in this talk some models for different Covid-19 data. I will first propose a SIR-type model for the data provided by the Johns-Hopkins University for several countries: these are the daily numbers of confirmed cases and deaths. The same model is used for all countries but the parameters of the model change from one country to another to reflect differences in dynamics. In particular, the model incorporates a time-dependent transmission rate, whose variations are thought to be related to the public health measures taken by the country in question.
I will then present a model for French hospital data provided by Santé Publique France: daily numbers of hospitalization, admissions in intensive care units, deaths and hospital discharges.
The proposed models may seem relatively simple, but it must be understood that they do not pretend to describe the spread of the pandemic in a precise and detailed way. Their role is to adjust the available data and provide reliable forecasts: their complexity is therefore adjusted to the amount of information available in the data. Indeed, very few parameters are needed to properly describe the outcome of interest and the prediction proves to be stable over time. Two interactive web applications are available to visualize the data and the adjusted models:
http://shiny.webpopix.org/covidix/app1/ for JHU data,
http://shiny.webpopix.org/covidix/app3/ for SPF data.