The Quest for Adaptive Inference: Comparing FC-TVPVAR and LSTM-TVPVAR in High-Dimensional Volatility Scenarios

The increasing complexity of financial contagion, combined with successive global crises, has intensified non-linear cross-market connections and the autoregressive conditional heteroscedasticity characteristics of financial time series, thereby exposing the limitations of constant variance models. In response, this study introduces a novel hybrid LSTM-TVP-VAR framework by combining Long Short Term Memory (LSTM) residual correction model with Time Varying Parameters Vector Auto Regression (TVP-VAR) to explore the cross-market non-linear connections and hidden temporal patterns across the cryptocurrency, commodity, technology, and clean energy markets. Performance results, including the Mean Total Connectedness Index (mTCI) value of 60.55%, Herfindahl-Hirschman Index (HHI) value of 0.0832, in addition to the Coefficient of Variation (CV) scores of 42.16% and 18.75% indicate that LSTM-TVP-VAR outperforms the FC-TVP-VAR in detecting non-linear connections, evenly allocating the volatility transmissions, and decreasing the sensitivity of results to idiosyncratic shocks from individual assets. Finally, Diebold-Mariano (DM) results verify the statistical significance of the contributions of the model and bootstrap test results validate the elimination of inherent over-fitting.