README.md

---
license: cc-by-nc-4.0
task_categories:
  - tabular-classification
tags:
  - cmbs
  - cre
  - credit-risk
  - sec-edgar
  - lineage
pretty_name: CMBS Special-Servicing Transfer Early-Warning
---

# CMBS Special-Servicing Transfer Early-Warning

Leakage-safe CMBS special-servicing-transfer prediction dataset for the job: identify CMBS assets at elevated risk of transferring to special servicing within the next 12 months, using only point-in-time information available from SEC-filed CMBS loan reports.

This release is a flat supervised learning table for tabular classification, AutoML, TabPFN, XGBoost, and credit-risk benchmarking.

```python
import pandas as pd
train = pd.read_parquet("train.parquet")
test = pd.read_parquet("test.parquet")
```

## Listing Terms

Contact: cairn@cmdrvl.com

License: CC-BY-NC-4.0. This dataset is open source for non-commercial use only.

Commercial use: Snowflake Marketplace listing coming soon, contact cairn@cmdrvl.com.

## Files

| File | Rows | Positives | Positive rate | Notes |
| --- | ---: | ---: | ---: | --- |
| `train.parquet` | 497,552 | 10,825 | 2.18% | `reporting_period_end_date <= 2022-12-31` |
| `test.parquet` | 299,401 | 6,581 | 2.20% | `reporting_period_end_date >= 2023-07-01` |
| `all.parquet` | 796,953 | 17,406 | 2.18% | Combined file with `split` column |

The six-month embargo window from 2023-01-01 through 2023-06-30 is excluded from all published files. Rows whose full 12-month forward label window is not yet observable are also excluded rather than shipped as negatives.

## Grain

One row is one CMBS asset observation at:

- `deal_cik`
- `loannumber`
- `assetnumber`
- `reporting_period_end_date`

`observation_id` is a stable row id. `cik` is retained alongside `deal_cik` for compatibility with SEC filing workflows.

## Label And Split

Target column: `transfers_to_special_servicing_within_12m`

The target is `1` when the asset's first observed special-servicer transfer date occurs after the observation period end date and within the next 12 months. Rows on or after that asset-level first transfer date are dropped before modeling, so the feature table contains pre-transfer observations only.

Split policy:

- Train: reporting periods on or before 2022-12-31.
- Embargo: 2023-01-01 through 2023-06-30, excluded.
- Test: reporting periods on or after 2023-07-01.

The split is intentionally temporal, not random. It asks models to generalize from the 2020-2022 regime into the 2023+ reporting regime. The embargo is a buffer band discarded between train and test so the two sets do not touch at the boundary, reducing leakage from near-adjacent observations of the same asset.

## Leakage Verification

This release is gated by a machine-checkable leakage receipt before publication. The receipt checks:

- No feature rows with `period_end >= first_special_servicer_transfer_date`.
- No published rows in the six-month embargo window.
- No point-in-time appointed-servicer join where the source filing date is later than the panel filing date it is joined to.
- Exact split reproduction: train 497,552 with 10,825 positives, test 299,401 with 6,581 positives.

## Column Families

The table has 121 columns. Families are:

- Keys, provenance, split, and label: `observation_id`, SEC deal identifiers, loan number, asset number, period end, `source_filing_id`, `filing_date`, `source_url`, target, and `split`.
- Decoded CREFC categories: payment status, workout strategy, loan structure, payment type, interest accrual method, original interest rate type, payment frequency, servicing advance method, and modification.
- Raw CREFC parallels: raw values are retained only as explicit `*_code` columns next to decoded `*_label` columns.
- Balance and rate features: current/scheduled/actual balances, original amount, balance deltas, balance-to-original ratios, report-period rate, and current rate. Interest rates are normalized to a single percentage-point scale (e.g. `4.51` = 4.51%); balance-to-original ratios above 2x are treated as bad-denominator artifacts and flagged missing.
- Term features: seasoning and months to maturity.
- Property features: property count, valuation-weighted DSCR, NOI, NCF, occupancy, valuation, revenue, expense, rentable area, units/beds/rooms, weighted year built, dominant property type, and dominant property state. Occupancy is normalized to a single 0-100 percent scale. The DSCR columns keep a `_pct` suffix for name stability but hold coverage RATIOS (e.g. `1.96` = 1.96x), not percentages.
- Completeness signals: `is_stub` and `tape_completeness_score` (see Missingness Honesty).
- Payment and delinquency features: severity rank, delinquency indicator, delinquency streak, trailing 12-period delinquency count, and payment status change.
- Modification and workout features.
- Point-in-time servicer features: primary servicer, appointed master servicer, appointed special servicer, appointed trustee, party fields, source rule, and roster coverage flags.
- Missingness flags for numeric fields that are zero-filled.

## Missingness Honesty

Numeric financial fields are mostly clean but not perfect. Missingness flags show:

- Actual balance: 5.11%.
- Actual loan balance: 5.11%.
- Scheduled balance: 3.65%.
- Original loan amount: 2.37%.
- Report-period rate: 5.13%. This includes filed literal `0.0` rates on active loans, which are physically impossible and are now treated as missing rather than shipped as present data (a curation fix that raised the honest rate-missingness from 3.64%).
- Current property DSCR using NOI: 67.27% populated.
- Current property occupancy: 66.94% populated.
- Dominant property type: 91.19% populated.
- Dominant property state: 89.37% populated.

Unit and value curation (so a `_missing = 0` value is trustworthy):

- Interest rates are normalized to one percentage-point scale and filed zeros are flagged missing.
- Occupancy is normalized to one 0-100 percent scale.
- Balance-to-original ratios above 2x (mis-scaled original loan amounts) are flagged missing.
- Seasoning cannot be negative; an origination date filed after the reporting period is flagged missing instead of clamped to zero.

Low-information stub rows are flagged, not dropped:

- About 11.9% of rows have `payment_status_label = Not Reported` — near-empty-tape placeholder observations. Their positive rate is 1.32% versus 2.18% overall. Dropping them would break the published row-count claims, so they are retained and flagged with `is_stub = 1` and a low `tape_completeness_score` (fraction of six core tape signals present) so you can filter or down-weight them. To train on the higher-signal population, filter `is_stub = 0`.

Sparse fields are sparse by nature, not dirty:

- `workout_strategy_label` is `Not Reported` for 98.82% of rows because workout strategy is generally only populated for distressed or workout-context loans.
- `appointed_trustee_name` is `Not Reported` for 78.01% of rows because trustee roster extraction depends on available point-in-time counterparty filings.
- Point-in-time counterparty roster coverage is 44.11%. When a counterparty roster is not available, servicer features fall back to source-visible `primaryservicername` where available and carry `special_servicer_source_rule` / `servicer_pit_source` so the fallback is explicit.

## Right-Censoring Caveat

The newest reporting periods may not yet have a full future 12-month observation window. Those right-censored rows are excluded from the shipped benchmark instead of being labeled as negatives. Future refreshes can add those periods once their full label window is observable.

## Provenance

The differentiator is provenance to SEC filings. The pipeline starts from EDGAR asset-period filings, keeps a stable observation id, preserves source filing lineage in the model, decodes CREFC codes through versioned crosswalk models, and publishes only after leakage, split, null, dictionary, and lineage checks pass.

Every published row includes `source_filing_id`, `filing_date`, and `source_url`. `source_filing_id` is the ABS-EE filing/submission identifier carried by the SEC filing-derived source panel. Use `source_url` to trace an observation back to the SEC archive context for the source filing-derived loan report, not to a vendor spreadsheet or opaque hand-curated sample.

## Quickstart

```python
import pandas as pd

train = pd.read_parquet("train.parquet")
test = pd.read_parquet("test.parquet")

y_train = train.pop("transfers_to_special_servicing_within_12m")
X_train = train.drop(columns=[
    "observation_id",
    "cik",
    "deal_cik",
    "loannumber",
    "assetnumber",
    "reporting_period_end_date",
    "source_filing_id",
    "filing_date",
    "source_url",
    "split",
])

y_test = test.pop("transfers_to_special_servicing_within_12m")
X_test = test.drop(columns=[
    "observation_id",
    "cik",
    "deal_cik",
    "loannumber",
    "assetnumber",
    "reporting_period_end_date",
    "source_filing_id",
    "filing_date",
    "source_url",
    "split",
])
```