---
title: Confidence
nav: Confidence
description: How the confidence score is computed, the thresholds that gate a recommendation, and when the system abstains entirely.
group: concepts
order: 4
status: shipped
updated: 2026-09-21
---

Confidence is a decision control, not a badge. It determines whether a recommendation can be acted on at all.

The principle: **weak evidence should produce a request for better data, not a confident-looking guess.**

## The bands

| Score          | Level     | Decision    |
| -------------- | --------- | ----------- |
| below 0.40     | Very Low  | Abstain     |
| 0.40 – 0.59    | Low       | Abstain     |
| 0.60 – 0.74    | Medium    | Provisional |
| 0.75 – 0.89    | High      | Normal      |
| 0.90 and above | Very High | Normal      |

**Abstain** means no change is recommended. The screen shows what evidence is missing instead of a number you should not trust.

**Provisional** means the recommendation is shown, identified as provisional, and requires explicit acknowledgement. The intent is a smaller staged change rather than the full move.

**Normal** means it can proceed through ordinary review.

## How the score is computed

Five weighted components, each between 0 and 1:

| Component     | Weight | What it measures                                                                |
| ------------- | -----: | ------------------------------------------------------------------------------- |
| Data coverage |   0.35 | How many of the three economic inputs are present (PPP, price level, income).   |
| Freshness     |   0.20 | How recent the economic observations are.                                       |
| Store support |   0.20 | Whether the store actually supports pricing in this market.                     |
| Rounding rule |   0.10 | Whether a rounding increment is known for the market's currency.                |
| Stability     |   0.15 | How far the adjusted factor sits from 1.0 — a large move is a less certain one. |

The result is rounded to two decimal places.

Data coverage carries the most weight because it is the component most often at fault. A market with only PPP and no income or price-level data scores `1/3` on coverage, which alone costs it 0.23 of the total.

## The 5% floor

Separately from confidence, a recommendation whose change is smaller than 5% is not surfaced. Below that threshold the change is within the noise of rounding and exchange-rate movement, and acting on it costs more attention than it returns.

This is why a market can show a different parity target in the matrix without generating a recommendation.

## What makes a recommendation approvable

All four must hold:

- status is open;
- the product has not been withdrawn from the store;
- the confidence decision is not _abstain_;
- the absolute change is at least 5%.

The interface checks these before enabling the button, and the server enforces them again on submission. Approving a batch reports each skipped item individually with its reason — a reviewer who selected forty needs to know which ones did not go through.

## What is not implemented

Three parts of the confidence design are specified but not built:

- **Visible components.** Today the score is shown; its five parts are not broken out on screen.
- **Higher thresholds for protected or high-revenue markets.** The bands are currently global.
- **Calibration against observed outcomes.** The score has never been checked against what actually happened after a price change, because outcome measurement is not built. Treat it as a measure of evidence quality, not of predicted success.