A RAG-Augmented Hybrid Forecasting Framework for Policy-Driven Sales Anomaly Explanation: Evidence from the Chinese Automotive Market (2016–2022)

Accurately forecasting automotive sales and interpreting the policy mechanisms that drive market anomalies are critical challenges for industry stakeholders and policymakers alike. Existing machine learning approaches provide quantitative prediction but offer no qualitative explanation for detected deviations. This paper proposes a novel hybrid framework that integrates gradient-boosted tree and recurrent neural network forecasting with a Retrieval-Augmented Generation (RAG) pipeline to simultaneously forecast monthly Chinese passenger vehicle sales and generate grounded, document-cited policy explanations for anomalous periods. Using the SRNI-CAR dataset (39,496 model-month records, 2016–2022), we train XGBoost and Long Short-Term Memory (LSTM) models achieving 18.21\% and 18.31\% MAPE respectively on held-out 2021–2022 test data. Walk-forward cross-validation across four annual folds confirms consistent generalisation (mean MAPE 13.90\% ± 3.27\%, excluding the COVID-19 structural break of 2020). Anomalous months are detected via residual thresholding and queried against 1,780 Chinese-language policy articles indexed in ChromaDB using multilingual Sentence-BERT embeddings. A locally-hosted Qwen3.5:4b language model synthesises retrieved policy context into verifiable, source-cited explanations. Qualitative evaluation demonstrates that the RAG component accurately attributes the June 2019 subsidy reform and February 2020 COVID-19 market suppression to their corresponding policy documents. This work constitutes the first application of a LangChain-free RAG architecture to Chinese automotive policy corpus analysis, bridging the gap between quantitative sales forecasting and qualitative policy interpretation.