Score range and interpretation
What Avra's credit models return — PDs at five horizons from Avra GFM, and a 0–1000 score from a credit model built for your portfolio — and how to read them.
The outputs on this page come from two places. Avra GFM returns probabilities of default (PDs) at five horizons, and an embedding, for each Brazilian company. A credit model built for your portfolio on top of it learns your own definition of default and can return a 0–1000 score derived from its PD. This page shows what each returns, how to read it, and how to turn it into thresholds.
What you receive
Probabilities of default (PDs)
Avra GFM returns a PD for each company at five horizons — 30, 60, 90, 180, and 365 days — so each decision uses the window that matches its policy. You can read a PD directly: “This company has a 5% chance of defaulting within 365 days.” Use it for risk-based pricing, IFRS 9 staging, and capital allocation.
Credit score (0–1000)
A credit model built for your portfolio can return a score for decision engines and policy rules, derived from its PD. Higher scores mean lower risk. The score is monotonic, so a higher score always means a lower probability of default.
Embeddings
A 1024-dimensional vector per company, for your own models. Use it to find similar companies or to add relationship-aware features to your credit models. It also supports segmentation, monitoring, and anomaly detection.
Example responses
The result of an Avra GFM prediction holds the five PDs, each in a one-element list:
{
"pd_30d": [0.004],
"pd_60d": [0.006],
"pd_90d": [0.011],
"pd_180d": [0.024],
"pd_365d": [0.051]
}
| Field | Horizon | Typical use |
|---|---|---|
pd_30d |
30 days | Short-term liquidity, payment timing |
pd_60d |
60 days | Trade credit, early delinquency |
pd_90d |
90 days | Standard lending decisions |
pd_180d |
180 days | Medium-term portfolio planning |
pd_365d |
365 days | Annual loss forecasting, IFRS 9 staging |
A credit model built for your portfolio can return its score next to the PD it’s derived from, for example:
{
"credit_score": 842,
"probability_of_default": 0.051
}
Field names vary by model. Read them from the model version’s schema with GET /v1/api/models/{model_id}/versions/{alias_or_version_id}.
Reading the 0–1000 score
The score is a monotonic presentation of the PD behind it: every PD maps to one score, and a higher score always means a lower PD. The curve between them belongs to the model that produced it.
So a score is comparable only with scores from the same model version. When the version changes, the PD behind a given score can change too, and the cutoffs set on the old version need revalidating before they decide anything.
Defining delinquency
You define what a bad outcome is, and your historical accounts are labelled by your operational definition, for example:
| Definition | Description |
|---|---|
| MOB 6 > 30 | More than 30 days past due in the sixth month on book |
| FPD 90 | First payment default after 90 days |
| Charge-off | Debt written off as a loss |
A credit model for your portfolio learns that definition from your labels and GFM embeddings, so its PD is the probability of your outcome rather than a generic default.
Setting bands and thresholds
Thresholds belong to your portfolio, not to a fixed table. The PD behind a given score depends on the model, its version, and your definition of default. So derive risk bands from your own outcomes, and revalidate bands and thresholds when you move to a new model version. See Risk bands.
Wherever a decision needs a probability, such as pricing or provisioning, use the PD.
Why the probability matters
A 0–1000 score is convenient for decision rules. The probability behind it is what lets you price on expected loss, setting rates from the risk you’re actually taking. It also lets you manage the portfolio by calculating reserves and capital requirements, track how risk evolves over time, and see how it differs across customer types.
Where to go next
Risk bands
Group scores into bands with distinct, stable default rates, and map them to policy.
Credit
The data, the target, and the decision flow for credit.
Live predictions
Call a model for one company at the moment of underwriting.
Model versioning
Pin or follow a version, and revalidate thresholds when it changes.