An independent factor-research stack built to answer one question: how many published sell-side equity factors survive a careful replication and a proper multiple-testing correction? Includes a live 29-panel factor dashboard.
A 60,000 × 40+ feature matrix from TA-Lib indicators, triple-barrier labelling, and time-series cross-validation — an end-to-end GBDT pipeline built to avoid label leakage.
Sequence models in PyTorch on the same feature system as the GBDT pipeline — a controlled comparison of tree ensembles versus recurrent architectures.
A two-factor-family research pipeline over ~15 years of daily data on 30+ Chinese commodity futures, with a full IC / ICIR / monotonicity / half-life evaluation suite.