To address the high-dimensional redundancy and nonlinear feature extraction challenges of financial multisource heterogeneous time series data, this paper proposes an automated dimensionality reduction framework that integrates a TCN-Autoencoder with dual screening of information value and the Pearson correlation coefficient. Design of a multiscale dilated convolution block combined with an autoencoder with residual connections to extract low-dimensional latent representations in unlabelled data. On this basis, the information value (IV) of each potential feature is calculated on the basis of the labelled samples to eliminate the weak distinguishing factor, the Pearson correlation coefficient is subsequently used to remove the redundant features, and the potential features with strong predictive ability and good complementarity are finally refined. The experiments prove that the method significantly reduces feature dimensionality and redundancy while enhancing the discriminative power and interpretability of the features, providing high-quality and robust input support for extreme financial risk prediction.