Data quality, revisions, and point-in-time research

Give each observation more than one date

The economic date describes the period or event measured. The publication time describes when the source released information. The ingestion time describes when the office or provider received it. A revision adds another version and availability time. Conflating these fields can allow a study to use information before it was knowable, even when the calculation code is otherwise correct.

The St. Louis Fed's ALFRED documentation explains the purpose of economic-data vintages and describes how release dates are obtained. Vintage data help reconstruct earlier information sets, but a dated observation does not always establish precise intraday availability. For a time-sensitive model, verify the original release timestamp and the relevant delivery lag instead of assuming that a calendar date is sufficiently granular.

Specify what each field and identifier means

Create a data dictionary covering units, currency, timezone, adjustment status, return convention, source, and missing-value codes. Map instrument identifiers through mergers, symbol changes, delistings, and venue differences. A ticker is not a permanent identity. Historical universe membership should reflect the period being studied rather than only securities that survive in today's database.

In a hypothetical equity study, an analyst joins two files by ticker and accidentally combines an old issuer with a later company using the same symbol. The apparent long history is invalid. Rebuild the join using stable identifiers and dated mappings, then determine which outputs depend on the error. Keep the failed version for diagnosis, but do not continue presenting its return history as evidence.

Use checks matched to the failure mechanism

This original table connects common checks to the errors they can reveal. Passing these tests supports data integrity within their scope; it does not validate an investment strategy. Set tolerances from the field definition and use case rather than declaring every large move an error or every smooth series reliable.

Separate original evidence from revised analysis

AQR's data-library notes explain that updated historical series can change after vendor corrections. MSCI's June 2026 index policies also contain correction and historical-restatement provisions. These examples show why a fresh download may differ from an earlier research input. They do not imply that every provider revises data in the same way or that every discrepancy has the same significance.

Maintain an original-as-used record and a latest-corrected analysis, with a change description linking them. In a hypothetical macro model, an initially published indicator would have produced one classification while its revised value produces another. Preserve the original live research publication and label the revised calculation as a reconstruction. Actual investment executions remain historical account facts; a data revision must not rewrite their prices or existence.

Control transformations and correction impact

Store transformation definitions and versions alongside the raw input references. Reconcile prices and income before calculating returns. In a hypothetical total-return series, adding a dividend again to an already reinvested return double-counts income. A plausible-looking aggregate may hide that defect, so compare selected events and intermediate calculations rather than inspecting only annual performance.

When a discrepancy is found, identify affected instruments, dates, features, outputs, and committee decisions. Determine whether the change is a source correction, a methodological revision, or an internal processing error. Assign an owner and document the response. Prioritize by decision impact as well as numerical size: a small revision near a threshold can change an output, while a larger revision elsewhere may not alter the decision.

Make missing evidence a visible limitation

Where licenses permit, retain immutable snapshots or sufficient version references, receipt times, and integrity checks to support reconstruction. A file checksum establishes file identity, not economic correctness or timely availability. If historical vintages are unavailable, state what cannot be verified and consider a narrower study or a frozen forward observation. Do not invent publication lags merely to label a study point-in-time.

Failure cases include silently replacing old inputs, forward-filling through a genuine data outage without an approved rule, and treating agreement between vendors as independent confirmation when they share a source. Committee reporting should state how many material exceptions remain and what decisions they affect. Data controls support trustworthy evidence; they cannot remove uncertainty about future markets or turn software-quality checks into independent investment results.

Sources and further reading

1. Federal Reserve Bank of St. Louis: ALFRED Help and vintage-data definitions
2. AQR: About the AQR Data Library
3. MSCI: Index Policies, June 2026, including Corrections Policy

The linked sources provide conceptual background. The worked examples and planning assumptions on this page are this publication's own illustrations.

Continue the research