摘要
This paper addresses an important gap in time series analysis by establishingrigorous results on prediction and model selection for weakly stationary processes withconditionally heteroscedastic errors, using mean squared prediction error (MSPE) as theevaluation metric. A key contribution is the derivation of a negative moment bound for theminimum eigenvalue of the sample autocovariance matrix. This result enables an explicitasymptotic decomposition of the MSPE for least squares predictors into interpretable com-ponents reflecting model complexity, misspecification, and time-varying volatility. Lever-aging this decomposition, we develop a model selection criterion that consistently iden-tifies the MSPE-optimal subset autoregressive model, even in the presence of conditionalheteroscedasticity and model misspecification. Simulation studies support the theoreticaldevelopments and demonstrate the practical effectiveness of the proposed method.