JohnnyGanYu/VolatilityReturn

VolatilityReturn reads like a serious competition solution rather than a generic ML notebook dump. The README explains the target clearly: predict Ret5 and Ret60 for 30 high-volatility symbols across normal and extreme market regimes, optimize average Pearson IC, and package a model under platform constraints. The project structure suggests real iteration: Numba-accelerated feature engineering with 165 features, LightGBM local/global/extreme-regime models, GRU and Transformer sequence models, per-dataset ensemble weights, local evaluation, platform submission scripts, feedback-driven optimization, smart packaging, and documented v1-v7 evolution. I especially like that the README includes lessons learned, such as single-target sequence models overfitting and dual-target GRU/Transformer helping eR60. That kind of postmortem detail makes the repo more useful to other ML practitioners. The presence of pytest markers for unit, property, integration, and slow tests is also a good sign. The main limitation is reproducibility. The competition data and trained model files are understandably absent, but the repo would be easier to evaluate with a tiny synthetic dataset, a smoke-test command, expected output shapes, and a documented end-to-end “train small, predict small, evaluate small” path. Financial ML code also benefits from stronger leakage warnings, seed control, dependency pinning, and a clear explanation of what platform feedback was used for greedy weight optimization.