Time-Aligned Rolling-Correlation Graph Fusion for Financial Fair-Value Observability Proxies: A Controlled Transformer Study
Fair-value hierarchy decisions concern the observability and significance of valuation inputs, whereas securities-market time series provide only indirect proxies. We study whether time-aligned rolling-correlation graphs improve sequence-based prediction under an explicitly synthetic three-class observability task. A reproducible architecture pairs window-level Transformer representations with Pearson-correlation message passing, then uses an element-wise gate to combine own-instrument and peer context. We test stable and changed-correlation regimes over three independently generated populations and two model initializations each, including static, recent-only and randomly rewired graph controls. In the changed-correlation setting, the dynamic hybrid obtains macro-F1 0.671 ± 0.081, compared with 0.667 ± 0.083 for the static hybrid, 0.668 ± 0.070 for rewired edges, and 0.832 ± 0.020 for random forests. Removing the fusion gate reduces dynamic macro-F1 to 0.536 ± 0.072. The stable-regime comparison likewise shows no systematic advantage for rolling graphs. The evidence suggests that fusion can be useful, but small gains attributed to dynamic topology can disappear under appropriate negative controls. The archived source, data generation process, individual probabilities, and paired comparisons make these results independently testable. The proxy classes are not audited IFRS 13 fair-value levels.
Authors
- Quinn Cruz
Institutions
- University of Wisconsin–Madison (US)
Publication Details
- Journal
- International Journal of World Economic Research
- Published
- 2026-11-16
- DOI
- https://doi.org/10.54097/hm6fa751
- Primary Topic
- Stock Market Forecasting Methods
- Type
- article
- Field-Weighted Citation Impact
- 0.00