Regularised forecasting via smooth-rough partitioning of the regression coefficients (Q2002584)

A sequence of \(n\) i.i.d. square-integrable random functions \(X_i(t)\in L^2[a,b]\) is observed on an equispaced grid \(t_0,t_1,\dots,t_T\), which generates the discretized curves \(\{X_i(t_0),X_i(t_1),\dots,X_i(t_T)\}\), \(i=1,\dots,n\), \(t_0=a\), \(t_T=b\). The aim of the paper is to predict the final point \(X_i(t_T)\) from the previous observations \(\{X_i(t_0),\dots,X_i(t_{T-1})\}\). The dependence of \(X_i(t_T)\) from the previous observations is modeled linearly in form of a Smooth-Rough-Partition model with an unknown change-point index \(q, 1\leq q\leq T\). There is an extensive description of the model, of the joint estimation procedure of parameters and simulation results are shown. Interesting applications by using real data as country fertility rate data, Mexico city pollution data, stock volatility series and sunspot number data are presented. Under a set of quite special assumptions on parameters and functions, a result on the asymptotic behavior of the estimator for the change-point index \(q\) is proved.