---
title: "Provider Health Scoring Model"
description: "Build an explainable travel-provider health score using latency, success, freshness, price accuracy, coverage and handoff quality."
slug: "provider-health-scoring-model"
translationKey: "architecture-provider-health-scoring"
locale: "en"
type: "guide"
category: "architecture"
tags: ["provider","health-score","reliability","price-accuracy","observability"]
publishedAt: "2026-09-26"
updatedAt: "2026-09-26"
reviewedAt: "2026-09-26"
---

A provider health score should not exist just to answer “is this supplier good?” with one number. It should support **routing, fallback and operational decisions with consistent signals**.

## Production scenario

One supplier is fast but has poor price accuracy; another is slower but reliable; a third times out in one market. Latency alone is not sufficient.

## Score dimensions

Useful dimensions include availability, latency, freshness, price accuracy, unavailable-after-click, usable-offer coverage and—when appropriate—commercial feedback.

## Example model

Normalize each component and weight it by use case:

```text
Health =
  0.25 Reliability
+ 0.20 Latency
+ 0.20 PriceAccuracy
+ 0.15 Freshness
+ 0.10 Coverage
+ 0.10 HandoffQuality
```

These weights are illustrative, not universal.

## Avoid one global score

Performance can differ by market, endpoint and vertical. Consider provider × market, provider × endpoint and provider × vertical scores.

## Hysteresis

Do not disable a supplier because of one short spike. Use rolling windows, minimum samples and different thresholds for degradation and recovery.

## Routing use

The score can influence supplier eligibility, traffic weight, fallback order and circuit-breaker decisions.

It should not silently become a traveler-facing commercial ranking bias. Routing health and result ranking are separate concerns.

## Failure modes

Watch for low-sample volatility, one metric dominating the score, slow recovery, mixing commercial performance into technical health and unfair penalties in new markets.

## Observability

Expose component breakdown, not only the final number. Operators should understand why a provider scored 62.

## Alternatives

With only a few providers, simple red/amber/green states can be enough. Weighted scoring becomes useful when routing choices and quality differences are more complex.

## Production checklist

Normalize metrics, enforce minimum samples, use rolling windows and hysteresis, segment scores, expose components, audit manual overrides and keep routing health separate from ranking.