Build a cross-sectional return-prediction model

You are given a real-market-derived daily cross-sectional factor panel with anonymized identifiers (opaque datetime/instrument IDs, 360 factor columns and a forward-return label). Train a method that outputs a continuous score per instrument-day; it is scored by the cross-sectional rank agreement between your scores and forward returns on a hidden, later period. Real cross-sectional alpha is weak and noisy — what matters is generalization, not fitting the data you can see.

Hard Constraints

• Submit an algorithm (predict), not a precomputed table of scores — the grader re-runs your code on hidden data.
• Keep the exact signature: predict(train_df, valid_df, test_features_df) -> DataFrame[datetime, instrument, score].
• Output must cover every test (datetime, instrument) row, with finite scores, no missing rows and no duplicates — otherwise the submission scores 0.
• Do not hardcode or look up answers: datetime/instrument are opaque IDs (not real dates/tickers), the test period is sealed, and there is no network at any time.

What You Have

• Data (/app/data/): train_panel.parquet (earlier period), valid_panel.parquet (a later, non-overlapping period), and feature_catalog.csv (factor families: f001..f158 = family A, f159..f360 = family B). The 360 columns f001..f360 are 6 price/volume fields, each given as a 60-day history — six contiguous 60-column blocks, one field per block — so they form a 6-field x 60-day temporal panel. datetime is an opaque, chronologically-ordered day index (d000001, …); instrument is an opaque ID (s00001, …); the raw factors contain missing values. A later, non-overlapping test period is sealed in the grader — you never see it.
• The editable baseline /app/methods/main/ — this directory is what gets graded. It is a deliberately weak template (a family-A subset + a vanilla MLP). Improve it in place, or rewrite the algorithm entirely. No stronger reference is shipped — going weak→strong is the task.
• Your self-check surface (free, unlimited): python /app/selfcheck.py trains your current methods/main/solver.py on train and reports ICIR on valid. This is your only feedback loop — iterate against it as much as you want. It is a proxy: the sealed test period differs, so a high valid ICIR is necessary but not sufficient.

What You Submit

Edit /app/methods/main/solver.py to expose this exact signature:

def predict(train_df, valid_df, test_features_df) -> pd.DataFrame  # columns: datetime, instrument, score

You may add helper modules next to solver.py. The runtime provides numpy / pandas / scikit-learn / torch / pyyaml. There is no submit step and no per-attempt feedback — work and self-check for as long as your run window allows, then leave your best solver.py in place; it is graded once at the end on the hidden test period.

How It Is Judged

After your run, the grader copies your methods/main/ into a clean box, hands predict the hidden test-period features (no label), and scores the result. The grader re-runs your predict on a hidden panel of similar size, in a container with a 30-minute limit and more CPU than your own box.

• IC = the same-day cross-sectional Pearson correlation between your score and the forward-return label;
• ICIR = the mean daily IC over the test period divided by its standard deviation.

The metric is ICIR (higher is better).