Efficient generalized cross-validation for state space models (Q3753355)

From MaRDI portal

!

This is the item page for this Wikibase entity, intended for internal use and editing purposes. Please use the normal view instead:

scientific article; zbMATH DE number 3990708
Language Label Description Also known as
default for all languages
No label defined
    English
    Efficient generalized cross-validation for state space models
    scientific article; zbMATH DE number 3990708

      Statements

      Efficient generalized cross-validation for state space models (English)
      0 references
      0 references
      0 references
      0 references
      1987
      0 references
      influence matrix
      0 references
      state space model
      0 references
      generalized cross validation criterion
      0 references
      point smoothing algorithm
      0 references
      Kalman filter
      0 references
      smoothing spline problems
      0 references
      variance components models for time series
      0 references
      stochastic polynomial spline models
      0 references
      adaptive regression
      0 references
      The initial model considered is \(y(i)=s(i)+e(i)\), \(i=1,...,n\), where s(i) is an unobserved Gaussian signal and the e(i) are independent \(N(0,\sigma^ 2)\) and independent of s(i). The s(i) are generated by the state space model \[ (*)\quad s(i)=h(i,\theta)'x(i),\quad x(i+1)=F(i,\theta)x(i)+u(i) \] where u(i) is a sequence of q-dimensional independent Gaussian random vectors. The estimate of s(i) is \(A(\theta)y,\) \(y'=(y(i),...,y(n))\) with \(A(\theta) = E\{sy'\}E\{yy'\}^{-1}\), \(s'=(s(i),...,s(n))\). The problem then is to determine \(\theta\) and this is to be done so as to minimise \[ V_ n(\theta) = \{\hat e(\theta)'\hat e(\theta)/n\}[1-\{A(\theta)\}/n]^{- 2}, \] which is called the generalized cross validation criterion. The \(\hat e(\theta) = \{I-A(\theta)\}y\). The initial problem considered is to compute \(tr\{A(\theta)\}\) in \(O(n)\) operations. This is effected via the fixed point smoothing algorithm of the Kalman filter apparatus corresponding to (*). Two further extensions are given, one to the case where the initial value, \(x(0),\) in (*) is incompletely specified and a further one where the initial model is of the form \(y(i)=s(i)+z(i)'\beta +e(i)\), \(i=1,...,n\) and z(i) is an observed vector. An interesting discussion is given of special cases of the general set-up including smoothing spline problems, variance components models for time series, stochastic polynomial spline models, adaptive regression models and fixed regressors corresponding to trading day effects.
      0 references

      Identifiers