ON THE UNBIASEDNESS PROPERTY OF AIC FOR EXACT OR APPROXIMATING LINEAR STOCHASTIC TIME SERIES MODELS (Q3738436)

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 3972030
Language Label Description Also known as
default for all languages
No label defined
    English
    ON THE UNBIASEDNESS PROPERTY OF AIC FOR EXACT OR APPROXIMATING LINEAR STOCHASTIC TIME SERIES MODELS
    scientific article; zbMATH DE number 3972030

      Statements

      ON THE UNBIASEDNESS PROPERTY OF AIC FOR EXACT OR APPROXIMATING LINEAR STOCHASTIC TIME SERIES MODELS (English)
      0 references
      0 references
      1985
      0 references
      autoregressive moving average model
      0 references
      Akaike information criterion
      0 references
      log maximum likelihood
      0 references
      invertible, Gaussian ARMA (p,q) model
      0 references
      stationary time series
      0 references
      ARMA(p,q)
      0 references
      logarithmic likelihood functions
      0 references
      spectral density
      0 references
      asymptotically unbiased
      0 references
      asymptotic bias
      0 references
      AIC
      0 references
      The author considers a sequence of observations x(1),...,x(N) of a stationary time series in the form \[ x(t)=\sum^{\infty}_{m=0}c(m)a(t-m),\quad c(0)=1 \] where a(t) \((t=0,\pm 1,...)\) is a sequence of independent random variables with mean 0, variance \(\sigma^ 2_ a\), fourth cumulant \(k_ 4\) and bounded absolut moments \({\mathcal E}| a(t)|^{\tau}\) of order \(\tau\). The ARMA(p,q) model is identified. \(L_ N[\eta]\) denotes the logarithmic likelihood functions for a model from the class of models parameterized by the vector \(\eta\) (for observations x(1),...,x(N)); \(E_ N[\eta]={\mathcal E}L_ N[\eta].\) For the case of coincidence of the spectral density f(\(\lambda)\) of the series under consideration with one of the densities of the model being fit f[\(\eta\) ](\(\lambda)\), it is proved for \(k_ 4=0\) that \(L_ N[{\hat \xi}^ N]-L_ N[{\hat \eta}^ N]-(\dim \zeta -\dim \eta)\) as the estimate of \(E_ N[{\hat \zeta}^ N]-E_ N[{\hat \eta}^ N]\) (\({\hat \zeta}\),\({\hat \eta}\) are estimates minimizing the likelihood function of the relevant class) is asymptotically unbiased; for \(k\neq 0\) the kind of bias is indicated. There is also an expression for \[ \lim_{N\to \infty}{\mathcal E}\{E_ N[{\hat \eta}^ N]-L_ N[{\hat \eta}^ N]\}. \] When f(\(\lambda)\) is distinct from any of the densities of the ARMA(p,q) model the asymptotic bias is of order \({\mathcal O}[(\delta +)^{\max \{p,q\}}]\), \(\delta +\) is a fixed number. In this way the properties of AIC are rigorously analyzed.
      0 references

      Identifiers