Pricing autonomy under global shocks: a parsimonious and empirically validated SVAR-transformer framework for Shanghai crude oil futures market maturity

This paper investigates the dynamic response of Shanghai crude oil futures (INE) to international benchmark price shocks and evaluates the evolution of market maturity from its inception to early 2025. Using weekly data from March 30, 2018, to April 30, 2025, we propose a hybrid modeling framework that integrates a Structural Vector Autoregression (SVAR) with a Transformer architecture. The SVAR component captures linear structural relationships and enables economic interpretation through impulse response and variance decomposition, while the Transformer—specifically its linear attention mechanism—models nonlinear dynamics and extreme event patterns that elude purely linear models. Rather than assuming that the neural component is superior by construction, the framework is evaluated against a conventional VAR/SVAR, a multilayer perceptron (MLP), and a linear-attention-only specification using chronological holdout evaluation, expanding-window validation, ablation tests, and alternative structural identification schemes. The full hybrid specification reduces the out-of-sample RMSE for INE returns by approximately 7.7% relative to VAR(2), while the improvement over the linear-attention-only specification is concentrated in high-volatility weeks. The mean nonlinear contribution ratio is approximately 0.184 in crisis weeks compared with 0.051 in normal weeks. Empirical results reveal a three-stage maturation process: an initial cultivation period (2018–2019) characterized by external-dependent mild fluctuations; an extreme shock period (2020–2022) marked by resonance-driven volatility; and a mature transition period (2023–2025) where internal and external shocks achieve more balanced contributions. The findings indicate that INE has moved toward greater market maturity, with improved but still incomplete pricing autonomy. The unconditional variance decomposition remains close to the conventional SVAR result, whereas the principal incremental contribution of the hybrid model lies in identifying state-dependent nonlinear adjustments during crisis periods. Results remain qualitatively robust under sign restrictions, a long-run restriction, and reverse recursive ordering.