A gradient-boosted ensemble combines many weak signals — no single one is decisive — into a default-probability estimate, exactly what each tree here represents: bureau history, income, telecom tenure, utility-payment rate, bank balance, device value.
Per-tree vote v_i in [-1, +1], feature importance weight w_i:
raw = Σ(v_i · w_i) / Σ(w_i) (weighted ensemble output)
p_ok = 1 / (1 + e^(-3·raw)) (sigmoid → probability of "good")
score = 300 + 550 · p_ok (rescaled to a 300–850 range)
Decision engine: score < threshold → declined
threshold … threshold+70 → conditional (guarantor)
threshold+70 … threshold+150 → approved
> threshold+150 → approved, premium rate
- Applicant profile — a thin-file applicant has almost no bureau history (traditional trees vote low); a strong-file applicant has deep bureau history (traditional trees vote high).
- Alternative-data weight — how much the model trusts telecom/utility/bank/device signals. At 0% the four alt-data trees are ignored (neutral vote); at 100% their real signal counts fully — this is the lever that recovers thin-file applicants who would otherwise be declined.
- Approval risk threshold — the lender's risk tolerance; raising it declines more borderline applicants, lowering it approves more.
Real-world reference: a conservative traditional-only model may approve as few as ~25% of applicants, missing many thin-file borrowers — younger applicants, recent immigrants, gig workers — who would in fact repay reliably.