Mapping Hybrid and Ensemble Models for Financial Volatility Forecasting: A Bibliometric and LLM-Assisted Review

Abstract

Financial volatility forecasting has undergone a rapid methodological transition from parametric econometric models towards machine and deep learning, hybrid architectures, and ensemble techniques. Despite this expansion, the literature lacks a synthesis combining bibliometric mapping with a granular methodological taxonomy of hybrid and ensemble models for volatility forecasting. This paper addresses that gap by retrieving 690 publications from Scopus and Web of Science, mapping the bibliometric landscape with bibliometrix, and constructing a 121-paper core corpus through multi-stage filtering on citation impact, recency, and Bradford Zone 1 sources. The corpus is classified using a tendimensional taxonomy covering model category, hybrid subtype, base models, combination strategy, forecast target, input features, data frequency, market and asset class, evaluation framework, and methodological novelty. To assess scalable annotation, we implement a three-model LLM-assisted pipeline using Claude 4.6, Gemini 3, and GPT-5, validated against a domain-expert human audit on a stratified subsample. Bibliometric results show a marked acceleration after 2020 and convergence between financial econometrics and computational predictive modelling. Hybrid and ensemble architectures outperform single-model benchmarks, with sequential GARCH–DL cascades, stacking and shrinkage ensembles, and decomposition-based hybrids emerging as prominent designs; CEEMDAN and VMD provide the most transferable gains. LLM agreement is strongly dimensionspecific: lexically observable dimensions (data frequency, model category, forecast target) achieve moderate agreement, whereas inferential dimensions (evaluation framework, methodological novelty) remain unreliable under abstract-only classification. These findings position LLMs as useful but bounded research assistants for systematic reviews in quantitative finance, capable of scaling first-pass classification when embedded in transparent, multi-model, and human-validated workflows.