<p>This paper introduces a modified Fisher scoring algorithm for maximum likelihood estimation of state-space model parameters, enhanced with a bootstrap-based approximation of the Fisher information matrix. The proposed method, referred to as the Boost Fisher Scoring (BF) algorithm, aims to improve convergence and the accuracy of standard errors, particularly in small samples or under model misspecification. A robust extension (BFout) is also developed for time series containing outliers, in which bootstrap resampling is performed over cleaned standardized residuals. Extensive simulation studies compare the performance of the proposed methods with classical Fisher scoring and nonparametric bootstrap, under various scenarios of sample size, variance, and autocorrelation. The results show that the BF and BFout algorithms offer improved numerical stability and competitive accuracy, with significantly lower computational cost than full bootstrap procedures. Applications to synthetic and real temperature forecast data demonstrate the practical value of the proposed methodology for robust calibration and inference in time series modeling.</p>