When Cross-Market Memory Fails: Leakage-Audited Regime-Conditioned Graph State-Space Forecasting
Abstract
Cross-market return models can be competitive over an entire test period yet fail immediately after market-state transitions, and apparent gains can be exaggerated when market closes, macroeconomic vintages, and missing observations are not aligned to the information actually available at each decision time. We therefore treat information-set control as the primary methodological contribution and use Transition-Regulated Intermarket Graph Mamba (TRIG-Mamba) as a mechanism-controlled implementation for testing when memory and cross-market coupling should adapt. Four daily panels are compiled directly from public historical observations using exchange-calendar availability records and historical ALFRED/FRED vintages, and external validity is additionally tested on the independently published Exchange-Rate (ER-8) benchmark of eight daily currency series from 1990–2016. Under the primary fixed holdout, TRIG-Mamba obtains average MAE, RMSE, directional accuracy, and information coefficient of 0.617, 0.901, 55.60%, and 0.067. Relative to the pre-specified FinMamba comparator, the average MAE gain is 0.56% and the directional-accuracy gain is 0.29 percentage points, while worst-regime MAE decreases from 0.804 to 0.777. Expanded comparisons with XGBoost using momentum features, an attention-LSTM, and a GARCH-MIDAS-X specification do not change the average ordering. On ER-8 under the standard 60%/20%/20% split with a 168-day input window and 12-day horizon, TRIG-Mamba obtains RSE 0.0304 and correlation 0.9638, compared with 0.0309 and 0.9627 for FinMamba. Calendar-block tests show that the MAE advantage over FinMamba is small in the relatively stable 2017–2019 interval (0.17%), larger during 2020–2022 (1.95%), and remains positive in 2023–2024 (0.68%). On USS-30, the 10-bps long–short check improves annualized net return from 7.38% to 8.06% and Sharpe from 0.91 to 0.97, but the economic effect remains modest and statistically uncertain. The evidence supports a bounded claim: rigorous point-in-time information control and regime-conditioned graph-state transitions can reduce transition-local forecasting error in liquid daily markets, without establishing causal spillovers or universal economic superiority.