TongjiFinLab/THGNN

THGNN is an academic reference implementation from Tongji University's Finance Lab. With 117 stars and 18 forks, it has achieved some visibility in the quantitative finance research niche. However, its velocity is effectively zero (0.0/hr) and it is nearly three years old, suggesting it is a 'code-for-paper' repository rather than a maintained library. In the world of quantitative finance, the 'moat' is rarely the model architecture itself, which is easily reproducible, but rather the proprietary data pipelines, feature engineering, and execution infrastructure—none of which are present here. The defensibility is low (3) because the approach—while theoretically sound in its use of heterogeneous graphs to model stock relationships—can be trivially cloned or superseded by more recent transformer-based time-series models (like Informer or Autoformer) or graph-based successors. Frontier labs (OpenAI, Anthropic) pose low risk as they focus on general-purpose reasoning rather than verticalized financial alpha generation. The primary threat is displacement by foundation models for time series (e.g., Lag-Llama, TimesNet) which are increasingly outperforming specialized architectural tweaks on noisy financial data.