Skip to content

Tutorial: Cross-sectional finance ranking (flagship)

Goal: uncertainty for a panel ranker (one row per date Ɨ asset) with walk-forward CV, rank-geometry decoupling, and context-level health gating.

Panel format

Long-format DataFrame: one row per (date, asset) with features, score, and target.

import pandas as pd
from deup.domains.finance import CrossSectionalDEUP

# panel_df columns: date, score, vol_20d, market_vol_21d, target_20d, ...
model = CrossSectionalDEUP(
    horizon=20,       # uses target_20d when present
    cv=5,
    embargo=1,
    random_state=0,
)
model.fit(panel_df)

Defaults wired in: DEUPRanker + PurgedWalkForward + rank residualization + HealthIndex + finance g-feature preset.

Predict

pred, unc = model.predict(test_panel, return_uncertainty=True)

Pass the same date column via the panel ā€” residualization is applied per date.

Calibrated intervals (held-out calibration split)

model.calibrate(cal_panel, alpha=0.1)
interval = model.predict_interval(test_panel)
interval.lower, interval.upper

Use a calibration panel the model did not see during fit.

Context-level health (Finding 2; Sanderink, 2026)

Do not rely on mean(g) per day at Nā‰ˆ50 ā€” see Aggregation reliability (Sanderink, 2026, Finding 1).

report = model.health_report(test_panel)
report.health    # per-date score in [0, 1]
report.gate      # True = trust / trade this date

Walk-forward on pre-computed residuals

If you already have walk-forward scores and realized rank losses in a panel (Step 0 done elsewhere), you can train only the error predictor g:

from deup.domains.finance_walkforward import walkforward_g_on_enriched

preds, diagnostics = walkforward_g_on_enriched(enriched_df, horizons=[20])

See the API reference for walkforward_g_on_enriched.

Next