Xiang et al. (2023) — THGNN for Stock Price Direction
Authors: Shiwan Xiang, et al.
Venue: referenced in Fanshawe Masih Cameron 2025 THGNN Correlation Forecasting (paper cites it as Xiang et al. 2022 / 2023 — bibtex inconsistency in the source paper).
Contribution
The original Temporal-Heterogeneous GNN (THGNN) architecture: Transformer temporal encoder + GAT relational encoder, applied to next-day stock price direction classification on US equities. Reports 57.9% accuracy and superior Sharpe / Calmar / Information Ratio vs Transformer-only, GAT-only, GRU, LSTM, and several hybrid baselines.
The architecture is exactly what Fanshawe Masih Cameron 2025 THGNN Correlation Forecasting adopts and extends. Key extensions in the THGNN-correlation paper:
| Dimension | Xiang et al. (2023) | Fanshawe et al. (2025) |
|---|---|---|
| Task | next-day price direction | 10-day correlation residual |
| Transformer layers | 1 self-attention layer | 4 pre-norm layers |
| Transformer heads | (single) | 8 heads, d_k=16 |
| Sequence length | 20 days | 30 days |
| GAT layers | 1 | 3 |
| GAT heads | (single) | 4 heads |
| Edge attributes | sector codes | corr, \lvert\mathrm{corr}\rvert, sign, sector, subindustry, relation class |
| Output | binary up/down | Fisher-z residual per pair |
| Loss | cross-entropy | Huber + Gaussian-soft-binning histogram |
| Interpretability | attention weights | Gradient × Input + attention |
Importance to this wiki
- This is the architectural antecedent of the correlation-forecasting paper. Without Xiang et al. the joint temporal + relational structure would not have been established as a working pattern for finance.
- It also validates that Transformer + GAT stacks beat pure-Transformer, pure-GAT, GRU, LSTM on US-equity prediction — the comparative claim Fanshawe Masih Cameron 2025 THGNN Correlation Forecasting inherits.