README.md

metadata

license: unknown
task_categories:
  - tabular-classification
tags:
  - finance
  - stocks
  - qrt
pretty_name: QRT Data Challenge — stock return sign prediction

QRT Data Challenge — data & artifacts

Data for the QRT challenge: predict the sign of a stock's next-day return from its last 20 daily returns and relative volumes plus GICS classification. Code that consumes and produces these files: https://github.com/edereynaldesaintmichel/qrt_stock_prediction

data/ — raw challenge files

• x_train.csv / x_test.csv — one row per (ID, DATE, STOCK): RET_1..RET_20 (daily returns, RET_h is calendar day t−h), VOLUME_1..VOLUME_20 (relative volumes), and GICS columns SECTOR, INDUSTRY_GROUP, INDUSTRY, SUB_INDUSTRY.
• y_train.csv — target RET: 1 if the next-day return is positive.
• submission_random_benchmark.csv — random-benchmark submission example.
• benchmark/ — the organizers' benchmark notebook and its submission.

DATEs are shuffled and non-adjacent, but at a fixed DATE every stock's RET_h refers to the same calendar day — each (DATE, h) is an aligned cross-sectional observation ("pseudo-day").

artifacts/ — generated by the repo's pipeline

• features_{train,test}.parquet — ~200 cross-sectional features (src/features.py).
• panel_{ret,vol}_mktadj.npy (stocks × pseudo-days, market-adjusted), panel_mkt.npy, graph_stocks.npy — pseudo-day panels (src/build_graph.py).
• nbr_{pos,neg}.npy, w_{pos,neg}.npy — top-100 positively/negatively correlated neighbors per stock with empirical-Bayes-shrunk correlation weights.
• peer_{train,test}.parquet — correlation-weighted peer momentum scalars.
• oof_*.npy / test_*.npy — out-of-fold and test predictions per model (LightGBM, CatBoost, MLP v1/v2, LGB+peer), aligned with the csv row order.
• feature_importance.csv, submission.csv.

All artifacts are reproducible from data/ with the repo's scripts.