To provide a parameter estimating method and device for simplifying the convergence of the values of parameters, and for shortening a calculating time at the same time, and a computer program for realizing a computer as a parameter estimating device, and a recording medium.
This data predicting device 1 accepts time series constituted of a plurality of data, and generates a white noise, and calculates a matrix ψ and Y necessary for Gauss-Newton method by defining the auto-correlation function of the white noise of lug 0 as dispersion S02, the auto-correlation function of the white noise of any lug other than 0 as 0, and the value of the inter-correlation function of the time series and the value of the white noise in the future of the time series as 0, and a matrix f(θ(h)) and are calculated from a temporary value θ(h) of the parameter θ when the time series are applied to an ARMA model, and the next temporary value θ(h+1) is calculated by the Gauss-Newton method, and the parameter θ is decided by repeated calculation, and prediction data are calculated by using the parameter θ.
Hideno Kono