Introducing Predictive Utility Scores: A Data Quality Metric That Actually Means Something

Ask a data buyer what quality means and you'll get a different answer every time. Complete? Fresh? Accurate? Well-structured? All of the above? The problem isn't that people disagree about quality — it's that the data industry has never converged on a single, comparable metric that buyers can use to evaluate datasets across sellers, formats, and domains.

Until now.

The Quality Score Problem

Most data marketplaces offer some form of quality indicator. A star rating. A completeness percentage. A freshness badge. These help, but they're insufficient for three reasons:

• They're not comparable. A 4-star dataset from Seller A and a 4-star dataset from Seller B might have wildly different actual quality — the metrics aren't normalized.
• They're incomplete. Completeness tells you nothing about licensing clarity. Freshness tells you nothing about schema reliability. You need the full picture.
• They're not automated. Most ratings rely on manual assessment or buyer reviews — which means new datasets have no scores, and scores drift as datasets degrade.

Introducing the Predictive Utility Score (PUS)

The Predictive Utility Score is a normalized composite metric (0-100) computed automatically for every dataset on Data Marketplace. It answers one question: How useful is this dataset likely to be for its declared purpose?

The PUS is built from four sub-scores, each measuring a distinct dimension of dataset fitness:

Schema Conformance (Weight: 25%)

Does the dataset match its declared structure? Our smart clearing agent verifies column names, types, cardinality, and required-field presence against the seller's schema declaration. A dataset that claims 12 columns of specific types and delivers 11 with mismatched types scores low. A dataset whose actual structure perfectly matches its declaration scores 100.

Data Quality (Weight: 30%)

How clean and complete is the actual data? This sub-score incorporates completeness (null rate), duplicate detection, distribution consistency across versions, freshness relative to the declared update cadence, and basic statistical validity. It's computed on the full dataset (or a statistically significant sample for large files) at listing time and re-evaluated on each version update.

Provenance Strength (Weight: 20%)

Can you trace where this data came from? Datasets with a registered SHA-256 commitment hash, a verified lineage chain (parent datasets linked), and a history of consistent clearing reports score high. Datasets uploaded without provenance commitments score zero on this dimension. Provenance isn't just academic — it's your audit trail when a regulator asks where your model's training data originated.

Rights Clarity (Weight: 25%)

How clear and complete are the usage rights? A dataset with a valid ODRL+DPV policy that passes structural validation, includes all required metadata (archetype, ticker, policy version, DPV profile with purpose and legal basis), and has been through the listing policy gates scores 100. A dataset with a vague or missing policy scores zero. You can't use data you're not confident you have the right to use.

Why These Weights?

Quality (30%) gets the highest weight because poor data produces poor outcomes regardless of how well-documented or well-licensed it is. Rights clarity (25%) and schema conformance (25%) share second place because a dataset you can't legally use or can't reliably parse is nearly worthless. Provenance (20%) rounds out the score — critical for enterprise buyers but less immediately blocking than the other three.

These weights are our starting defaults. Enterprise customers can adjust them to reflect their own priorities via the API.

Automated, Not Opinionated

The PUS is not a rating assigned by a human reviewer. It's computed deterministically from measurable properties of the dataset and its metadata. This means:

• Every dataset gets a score the moment it's listed — no cold-start problem.
• Scores update automatically when datasets are updated, re-cleared, or when policies change.
• Scores are reproducible — the same dataset with the same metadata always gets the same score.
• AI agents can filter by PUS in API queries, enabling autonomous quality-gated procurement.

Using PUS in Practice

For buyers: sort search results by PUS to surface the highest-quality datasets first. Set minimum thresholds in your agent's procurement rules. Compare datasets in the same domain on a level playing field.

For sellers: the PUS is your competitive signal. Improve your schema declarations, register provenance commitments, adopt a clear license archetype, and watch your score — and your position in search results — rise.

The PUS doesn't replace domain expertise. A dataset scoring 95 might still be wrong for your specific use case. But it eliminates the datasets that fail on objective, measurable dimensions — so your expertise is applied only where it matters.

Every dataset on Data Marketplace displays its PUS and sub-scores. Browse at datamarketplace.io.