Back to docs

Understanding the score

What each number in the assessment tells you — and when to trust it.

The composite score

Every assessment returns a single composite score between 300 and 850, mapped to a letter grade (A+ to F) and the implied probability of default over the next 12 months (PD).

GradeWhat it means12-month PD
A+Exceptional — negligible default risk< 0.25%
A / A-Strong — investment grade equivalent0.25% – 1%
B+ / BHealthy — typical SME in good standing1% – 3%
B-Acceptable — watch for deterioration3% – 5%
C+ / CElevated — conditional approval advised5% – 10%
C-Weak — refer to committee / secure additional collateral10% – 15%
D+ / DPoor — decline or highly secured only15% – 25%
E / FDistressed — decline> 25%

How the composite is built

AgentCredit uses an ensemble of three independent models. The composite is a weighted average of their PD estimates; weights are learned from training data and adjust at scoring time based on what data you supplied.

  • Statistical scorecard — weight-of-evidence binned logistic regression, explainable at feature level
  • XGBoost ML model — gradient-boosted trees with SHAP explanations; a P&L-specialist head auto-routes when income-statement data is present
  • Altman Z-score — the 1968 Altman formula with 2007 industry calibration and a startup adjustment for firms < 3 years old

On top of that, three adjustments modulate the final PD: country risk (sovereign rating + OECD tier), industry risk (UK SIC section A–U), and data completeness (see below).

For the full technical breakdown — training data, algorithms, known limitations, performance numbers — see the model card.

Data completeness grading

UK micro-entities file balance-sheet-only accounts (no P&L) ~99% of the time. AgentCredit grades every assessment as full, partial, or limited, and applies a conservative PD uplift when data is incomplete:

  • Full — complete balance sheet + P&L + cashflow. No adjustment.
  • Partial — P&L missing or sparse. A modest conservative uplift is applied to PD and confidence intervals widen.
  • Limited — balance sheet only. A larger conservative uplift is applied; treat as directional, not definitive.

The UI flags this prominently. Interpret “Limited data” assessments as guidance, not ground truth.

Peer comparison

Every assessment benchmarks key liquidity, leverage, and profitability ratios against real UK peers. Buckets are keyed by UK SIC 2007 at two levels:

  • Division (2-digit, ~88 buckets) — preferred when enough peers are available
  • Section (letter A–U, 21 buckets) — coarser fallback

We return percentiles (P10, P25, P50, P75, P90) so you can see where the counterparty sits in the distribution. A company at the 25th percentile current ratio within its SIC division is below average but not necessarily distressed — always read the composite and the narrative together.

When to override the model

AgentCredit is a decision-support tool, not an automated approver. Consider overriding the composite when you have information the model doesn't:

  • Recent management changes, litigation, or press that the model can't see
  • Very recent filings or interim accounts more current than our training snapshot
  • Bank statement or Open Banking data (not yet wired in)
  • Your own default history with this counterparty

Admins on your account can amend the recorded grade with a mandatory written reason — the amendment is audited, visible to everyone on your team, and preserved in the record.