Presentations

2024-06-29

1. 

   Adaptive prediction for Functional Times Series

   Hassan Maissoro

   ISNPS 2024, Braga, Portugal, 25-29 June, 2024

   Abs PDF Slides

   An adaptive procedure for curve prediction for a stationary functional time series is proposed. The sample paths of the functional times series are assumed to be irregular and are observed with error at discrete times in the domain. Our linear predictor is based on the best linear unbiased predictor (BLUP) and on the adaptive nonparametric mean and autocovariance functions estimators. That is, the bandwidth parameters of these estimators are chosen adaptively with respect to the local regularity of the sample paths. The benefit of such a procedure will be a reduction in risk prediction compared to existing procedures.

2024-05-30

1. 

   Adaptive prediction for Functional Times Series

   Hassan Maissoro

   55es Journées de la Statistique de la SFdS, Bordeaux, France, 27-30 May, 2024

   Abs PDF Slides

   An adaptive procedure for curve prediction for a stationary functional time series is proposed. The sample paths of the functional times series are assumed to be irregular and are observed with error at discrete times in the domain. Our linear predictor is based on the best linear unbiased predictor (BLUP) and on the adaptive nonparametric mean and autocovariance functions estimators. That is, the bandwidth parameters of these estimators are chosen adaptively with respect to the local regularity of the sample paths. The benefit of such a procedure will be a reduction in risk prediction compared to existing procedures.

2024-03-15

1. 

   Adaptive estimation for Weakly Dependent Functional Times Series

   Hassan Maissoro

   Functional Data Analysis Workshop, Lille, France, 14-15 March, 2024

   Abs arXiv HTML PDF Slides

   We study the local regularity of weakly dependent functional time series, under L^p-m-approximability assumptions. The sample paths are observed with error at possibly random, design points. Non-asymptotic concentration bounds of the regularity estimators are derived. As an application, we build nonparametric mean and autocovariance functions estimators that adapt to the regularity of the sample paths and the design which can be sparse or dense. We also derive the asymptotic normality of the adaptive mean function estimator which allows for honest inference for irregular mean functions. An extensive simulation study and a real data application illustrate the good performance of the new estimators.

2023-09-13

1. 

   Learning the Smoothness of Weakly Dependent Functional Times Series

   Hassan Maissoro

   ENSAE-ENSAI Days, Paris, France, 12-13 September, 2023

   Abs HTML PDF Slides

   We consider functional time series where the sample paths are observed with error at possibly random discrete points in the domain. We reconsider the local regularity estimator proposed by Golovkine et al. (2022) in the context of weakly dependent curves, under the assumption of L^p-m-approximability. In this new framework, we derive non asymptotic exponential bounds for the concentration of the regularity estimators. This will further allow to diagnose a change of regularity along the sample paths, to build optimal estimator of mean and (auto)covariance functions, \emphetc. An extensive simulation study illustrate the good performance of our estimators with finite time series. The simulation experiments also provide guidance for the choice of the hyperparameters involved in our estimation method.

2023-07-07

1. 

   Learning the Smoothness of Weakly Dependent Functional Times Series

   Hassan Maissoro

   54es Journées de la Statistique de la SFdS, Bruxelles, Belgique, 03-07 July, 2023

   Abs HTML PDF Slides

   We consider functional time series where the sample paths are observed with error at possibly random discrete in the domain. We reconsider the local regularity estimator proposed by Golovkine et al. (2022) in the context of weakly dependent curves, under the assumption of L^p-m-approximability. In this new framework, we derive non asymptotic exponential bounds for the concentration of the regularity estimators. This will further allow to diagnose a change of regularity along the sample paths, to build optimal estimator of mean and (auto)covariance functions, \emphetc. An extensive simulation study illustrate the good performance of our estimators with finite time series. The simulation experiments also provide guidance for the choice of the hyperparameters involved in our estimation method.

2023-04-04

1. 

   Modelling a collection of curves using Functional Data Analysis

   Hassan Maissoro, and Sunny Wang

   Breizh Data Day, Saint-Brieuc, France, 04 April, 2023

   HTML Slides

2022-11-08

1. 

   Learning the Smoothness of Weakly Dependent Functional Times Series

   Hassan Maissoro

   Journée Mathématiques et Entreprises, Paris, France, November 8th, 2022

   Abs HTML Poster

   We consider stationary functional time series where each observation is a trajectory, measured with error at discretely, possibly randomly, sampled domain points. We consider the estimator for the local regularity of the trajectories introduced by Golovkine et al. (2022) in the context of dependent observations. A non-asymptotic bound for the concentration of the local regularity estimator is derived for functional time series which are L^p-m-approximable in the sense of Hörmann and Kokoszka (2010). We also derive a non-asymptotic concentration bound for the Hölder constant estimator. Given the estimates of the local regularity and the Hölder constant, one can diagnose changes in regularity along the trajectory, build optimal recovery of the trajectories, \emphetc. Our estimates perform well in simulations. Real data sets illustrate the finite sample performance.

2022-09-04

1. 

   Learning the Smoothness of Weakly Dependent Functional Times Series

   Hassan Maissoro

   5th Workshop On Goodness-Of-Fit, Rennes, France, 2-4 September, 2022

   Abs HTML Poster Slides

   We consider stationary functional time series where each observation is a trajectory, measured with error at discretely, possibly randomly, sampled domain points. We consider the estimator for the local regularity of the trajectories introduced by Golovkine et al. (2022) in the context of dependent observations. A non-asymptotic bound for the concentration of the local regularity estimator is derived for functional time series which are L^p-m-approximable in the sense of Hörmann and Kokoszka (2010). We also derive a non-asymptotic concentration bound for the Hölder constant estimator. Given the estimates of the local regularity and the Hölder constant, one can diagnose changes in regularity along the trajectory, build optimal recovery of the trajectories, \emphetc. Our estimates perform well in simulations. Real data sets illustrate the finite sample performance.