Backtested like a hedge fund. Published like a regulator.
Public CRE disclosure, refined into a distress read you can act on — and audit.
A hedge fund backtests a signal and keeps the result to itself, because it trades the signal — disclosure would hand away the edge. We're the inverse: the signal is the product, so proving it and selling it are the same act. Trust here can't be "trust us, it works" — it has to be the method, published in full and open to your audit.
A fund proves itself with realized P&L. We prove ourselves with method. Here's ours.
Every historical call is scored using only the data that existed that day — the DSCR, the maturity, the special-servicing status as reported then. No number from the future ever touches a past prediction.
The model is re-fit as the record grows, quarter by quarter, and only ever sees outcomes that had already happened. It is never trained on the very transfers it's later asked to have predicted.
A "3% risk" is measured to mean ~3% — on loans the model wasn't trained on. The probability is honest, not just a ranking, so the watch line means the same thing every quarter.
And here's the other side — the false alarms. A watch list is deliberately broad; it over-flags so it can catch trouble early. Of the watches that have resolved, about 50% went on to transfer (356 of 708); the rest fell back below the watch line without a transfer. Most watches — 3109 — are still open, so that hit rate will keep moving. We don’t claim the flag caused the recoveries; only that we raised it first.
Real loans, real dates. Each was on our watch list — point-in-time, no hindsight — well before it transferred to special servicing.
| Property | Trust | Flagged | Transferred | Lead |
|---|---|---|---|---|
| 1384 Broadway New York, NY |
CD 2017-CD3 | 2022-09-15 | 2026-03-15 | 42 mo early |
| 5001 East Commercenter Drive Bakersfield, CA |
JPMCC Commercial Mortgage Securities Trust 2017-JP6 | 2022-09-15 | 2025-12-15 | 39 mo early |
| 600 Grant Street Pittsburgh, PA |
Morgan Stanley Capital I Trust 2021-L6 | 2023-06-15 | 2026-05-15 | 35 mo early |
| 25350-25360 Magic Mountain Parkway Santa Clarita, CA |
BENCHMARK 2018-B2 | 2022-09-15 | 2025-06-15 | 33 mo early |
The receipt behind principle 03, not just the assertion: predicted transfer probability vs. the realized 2-quarter transfer rate, on walk-forward scored quarters the model wasn't trained on. A calibrated model keeps the last two columns close.
| Predicted band | Scored quarters | Mean predicted | Realized 2q transfer |
|---|---|---|---|
| 0–1% | 95,824 | 0.33% | 0.47% |
| 1–2% | 20,735 | 1.42% | 0.95% |
| 2–3% | 7,745 | 2.44% | 1.78% |
| 3–5% | 6,041 | 3.8% | 3.61% |
| 5–10% | 3,239 | 6.68% | 4.72% |
| 10%+ | 700 | 14.07% | 15.71% |
Open by design
- How every model is validated — the walk-forward, the no-lookahead rule
- The full scorecard, misses included — recall, lead time, calibration
- The features that drive each score (e.g. DSCR, maturity, special servicing)
- Named, dated receipts you can check against the public record
Proprietary — the refinery
- The full modeling pipeline and how ~30 raw public feeds are cracked into graded signals
- The exact weighting, thresholds, and the engineering that ties it together
- You can audit our gasoline's octane; the catalytic process is the business