A Deep Neural Network-Driven Method for Outlier Identification of Regional Economic Statistical Data

Abstract

Regional economic statistical data serves as the fundamental basis for local governments to formulate industrial policies, fiscal plans and livelihood guarantee schemes. Taking deep neural networks as the core tool, this paper constructs an automatic outlier identification framework adaptable to multi-indicator panel datasets of regional economic statistics. The research adopts panel data covering 10 core economic statistical indicators from 31 provincial-level administrative regions across China during 2016– 2025 as experimental samples, and builds a hybrid neural network integrating fully connected deep autoencoders and Long Short-Term Memory (LSTM) networks. Through unsupervised learning, the model excavates internal correlations among economic indicators and their temporal evolution patterns, and takes reconstruction error as the criterion for anomaly judgment to realize batch outlier detection. Experimental results show that the proposed hybrid deep neural network model achieves a precision of 94.72%, a recall rate of 93.16% and an F1 composite score of 93.93%, all outperforming the comparative models. Five standardized data tables are designed in this paper to record sample dataset descriptions, model hyperparameter configurations, reconstruction error threshold divisions, performance comparisons of multiple identification models, and classified statistics of typical outlier cases respectively. The study verifies that deep neural networks can break the linear constraints of traditional methods, accurately capture complex nonlinear characteristics of regional economic data, and effectively reduce missed and false detections of statistical outliers. This method can be directly embedded into data verification systems of local statistical departments, providing intelligent technical support for quality control of economic statistical data.