Default risk in commercial real estate loans · Class of 2026
Loan-month rows
766,558
SEC filing observations
Unique loans
24,771
After entity resolution
Reporting months
105
Dec 2016 – Aug 2025
Features
205
Loan + property + macro
Obs-level default rate
3.29%
Loan-level ever-default
19.96%
Mild-delinq → default
56.6%
Median time to default
63 days
From first delinquency
Observation-level default rate by year
Observation-level default rate by property type
XGBoost classifier predicting whether a currently performing loan will default within the next N months. Time-aware train/validation/test split.
Test-set performance comparison across horizons
Lift: % of test defaults captured at each risk decile (3-month model)
Top predictive features (gain-based)
Loan fixed-effects estimates of next-3-month default risk on macro shocks. State-varying shocks also include reporting-month FE.
Estimated effect of macro shock on next-3M default probability (pp per 1pp shock)
Placebo lead test (future shock should not predict current default)
| Shock | Placebo (pp) | 95% CI | p-value |
|---|---|---|---|
| Fed funds 3M Δ | −0.680 | [−0.81, −0.55] | significant |
| 10Y Treasury 3M Δ | +0.135 | [+0.01, +0.26] | significant |
| 30Y mortgage 3M Δ | −0.288 | [−0.41, −0.17] | significant |
| State unemployment 3M Δ | −1.460 | [−1.61, −1.31] | significant |
| State real GDP 3M Δ | +1.034 | [+0.92, +1.15] | significant |
All five placebo tests are significant, suggesting residual time-series confounding. The macro associations are interpreted as suggestive panel evidence rather than clean causal estimates.
Cross-sectional IPTW: bulge-bracket vs niche originators
Naive default — BB
18.73%
Naive default — niche
21.37%
IPTW ATE
−2.81pp
95% CI [−3.89, −1.73]
5Y survival — BB
72.87%
Observation-level default rate by property state. Hover any state for details.
Highest-default states (observation-level)
| State | Default rate | Observations |
|---|
Simplified scoring tool inspired by the project's XGBoost feature importance. Adjust loan characteristics to see an indicative 3-month default probability.
Indicative only. Based on simplified weights derived from the capstone's feature importance and panel macro coefficients. Not a substitute for the full XGBoost model.