all numeric pca 95
All governed numeric features compressed to 95% training-fold explained variance.
MAE
57.036
RMSE
71.672
R²
0.716
Pipeline Overview
Evidence is transformed at application grain and evaluated as a simulation-only credit-risk diagnostic.
Credit-scoring dataset
Application-grain source evidence
Source tables
11
Prepared rows
1,000
Candidate features
126
Feature engineering & selection
10 final model inputs from governed candidates
Candidate features
126
Final model inputs
10
The persisted diagnostic model retains 10 raw features, the smallest set within one standard error of the lowest nested-CV MAE. Missing source values remain unavailable, receive fold-local imputation, and retain a missingness indicator.
| Final model feature | Business meaning | Data source | Permutation importance |
|---|---|---|---|
| max_dpd_12m | Governed engineered feature; description pending registry entry. | credit_facility | 47.3642 |
| requested_loan_to_revenue_ratio | Requested credit relative to annual business revenue. | loan_application + financial_statement | 34.9561 |
| requested_loan_to_operating_cashflow_ratio | Governed engineered feature; description pending registry entry. | derived application-grain feature | 2.0743 |
| debt_service_coverage_ratio | Governed engineered feature; description pending registry entry. | derived application-grain feature | 1.2558 |
| installment_to_operating_cashflow_ratio | Governed engineered feature; description pending registry entry. | derived application-grain feature | 0.9325 |
| installment_to_bank_inflow_ratio | Governed engineered feature; description pending registry entry. | credit_facility + bank_statement_month_sum | 0.9063 |
| positive_cashflow_month_ratio_6m | Governed engineered feature; description pending registry entry. | derived application-grain feature | 0.3774 |
| negative_cashflow_month_count_6m | Governed engineered feature; description pending registry entry. | derived application-grain feature | 0.3755 |
| debit_credit_ratio_6m | Governed engineered feature; description pending registry entry. | derived application-grain feature | 0.3443 |
| worst_collectability_12m | Governed engineered feature; description pending registry entry. | credit_facility | 0.2841 |
Comparable candidate branches
Group-safe regression evaluation with fold-local selection
one se 10 plus residual pca 95
One-SE minimal raw features plus PCA of the non-overlapping numeric residual block at 95% training-fold variance.
MAE
43.364
RMSE
55.854
R²
0.828
one se minimal 10 raw
Smallest 10-feature raw set within one standard error of the best nested-CV MAE.
MAE
43.420
RMSE
57.195
R²
0.819
top 20 permutation raw
Top 20 fold-local permutation-ranked raw features.
MAE
42.672
RMSE
55.759
R²
0.828
top 30 permutation raw
Top 30 fold-local permutation-ranked raw features.
MAE
42.579
RMSE
55.664
R²
0.829
top 45 permutation raw
Top 45 fold-local permutation-ranked raw features.
MAE
42.298
RMSE
54.613
R²
0.835
top 65 permutation raw
Top 65 fold-local permutation-ranked raw features.
MAE
42.666
RMSE
55.274
R²
0.831
Optional financial-statement resilience
A missing statement is evidence status, not an error or zero value
observed as recorded
MAE 37.82 · RMSE 51.68 · R² 0.853
financial statement unavailable
MAE 73.97 · RMSE 92.62 · R² 0.528
A score remains eligible with bank statements, bookkeeping, invoices, or a financial statement.
Model evaluation
Current persisted diagnostic credit-score model
catboost · 10 raw features
Validation MAE 42.83 · Test MAE 41.04 · Test R² 0.831
Synthetic FICO-style scores and internal SLIK-aligned tiers are diagnostic only.