LLM-Enhanced Stock Prediction on a Company Relation Graph
This model repository documents the modeling component of the project LLM-Enhanced Stock Prediction on a Company Relation Graph.
The project compares three models for company-quarter stock prediction:
- Text + Num MLP: a multilayer perceptron over frozen transcript embeddings and numerical market features.
- Graph + Num GCN: a two-layer GCN over an SIC2 company relation graph using numerical node features.
- Text + Graph + Num GCN: the main model, combining transcript embeddings, numerical features, and graph propagation.
Important Note
No trained checkpoint file is currently included in this repository because the local project directory does not contain saved .pt, .pth, .ckpt, or .safetensors weights. The repository therefore provides the model definitions, training/evaluation code, metrics, and reproducibility instructions rather than a ready-to-load checkpoint.
Results
| Model | Val Acc. | Val Macro-F1 | Val AUC | Test Acc. | Test Macro-F1 | Test AUC |
|---|---|---|---|---|---|---|
| Text + Num MLP | 0.583 | 0.520 | 0.516 | 0.579 | 0.504 | 0.489 |
| Graph + Num GCN | 0.575 | 0.471 | 0.476 | 0.696 | 0.413 | 0.449 |
| Text + Graph + Num GCN | 0.567 | 0.518 | 0.518 | 0.651 | 0.543 | 0.528 |
The graph-enhanced model improves test macro-F1 from 0.504 to 0.543 and test AUC from 0.489 to 0.528 relative to the Text + Num MLP baseline.
Reproduction
Use the companion dataset repository, then run:
python scripts/07_train_baselines.py
The GCN is implemented directly with torch.sparse.mm; torch_geometric is not required.
Companion Dataset
The intermediate dataset artifacts are hosted in the companion Hugging Face dataset repository.