Why not only OptimalBinning
Start with the honest point
Section titled “Start with the honest point”OptimalBinning already solves the static cutting problem very well.
RiskBands does not need to deny that. On the contrary: in the current
repository, the numeric supervised strategy uses optbinning.OptimalBinning in
the backend.
This means the correct message is not:
- “RiskBands completely replaces OptimalBinning”
The correct message is:
- “RiskBands reuses a strong static cut when it makes sense and adds the layers needed for credit selection under temporal stress”
What static search does not answer by itself
Section titled “What static search does not answer by itself”Static search is excellent for questions such as:
- where are the best aggregate cuts?
- how can static separation be maximized?
But credit teams also need to answer questions such as:
- does ordering between bins survive by vintage?
- where does the event rate cross over time?
- which bins lose coverage?
- are we accepting a candidate with hidden structural fragility?
- if two candidates are plausible, which one has the more defensible rationale?
What RiskBands adds on top
Section titled “What RiskBands adds on top”| Layer | Role |
|---|---|
| Temporal diagnostics | Look beyond aggregate separation |
| Structural penalties | Penalize fragile bins, low coverage, reversals, and volatility |
| Candidate comparison | Put static, temporal, and balanced candidates on the same scale |
| Auditable rationale | Explain why a candidate won in terms closer to the business |
| Credit-oriented score | Balance discrimination with temporal defensibility |
An honest comparison
Section titled “An honest comparison”If behavior by vintage remains coherent and the structure does not lose coverage, RiskBands can simply confirm the static winner.
If vintage reading reveals crossings, rare bins, or loss of coherence, maximizing only IV stops being a comfortable answer for credit.
Where Optuna fits, without exaggeration
Section titled “Where Optuna fits, without exaggeration”The project can also use Optuna as an optional search layer in supervised
flows.
This point matters, but it should not be confused with the main RiskBands thesis.
Optuna enters to explore configurations when that makes sense. The real differentiator remains the scale used to judge candidates:
- separation
- temporal stability
- coverage
- reversals
- structural fragility
- auditable objective score
Why this matters in PD
Section titled “Why this matters in PD”In credit, the practical decision is rarely:
- “what is the best set of static cuts from a mathematical point of view?”
It is usually closer to:
- “what is the best set of cuts that I can still defend when the portfolio changes over time?”
RiskBands was designed to occupy that space.