Methodology: How News Quantified Measures Market Reactions

TL;DR

Every number News Quantified publishes rests on five documented disciplines: a maintained event classification, point-in-time timestamps that are never retroactively improved, persistent entity resolution that survives corporate actions, stated return conventions per reaction window, and corrections that leave a dated trace instead of a silent edit. This hub documents each. If a published statistic cannot be traced back through this page, we should not have published it.

The five disciplines

1. Classification

Events are classified into the analyst-action classes this site publishes a measured cross-section for — upgrades, downgrades, initiations, maintains, price-target raises and price-target cuts — and are grouped on every ticker page under earnings, analyst actions and SEC filings. A class is published for a name only once it clears a stated minimum sample; below that floor it is measured but not shown, because a median over a handful of events is not a finding. The measured class list, each class with its own N, is event types; the canonical vocabulary is terminology.

2. Point-in-time timestamps

Timestamps record first public availability. They are never revised when a "better" time is later known, because a backtest keyed on improved timestamps sees information before the market did. This is a collection policy enforced since 2012, which is why it cannot be replicated by re-scraping archives today.

3. Entity resolution

Events link to persistent security identifiers, not ticker strings. Ticker reuse, symbol changes, mergers, and delistings do not orphan or misattach history. Delisted names remain in the dataset with full event records; excluding them would remove the left tail from every distribution.

4. Reaction measurement

Reactions are measured over defined windows: 5 minutes, 30 minutes and to the close on the day of the event, then a hold ladder at one, two, three, four and five trading days, and at one, three, six and twelve months. Every one of them uses the same return convention — the percent move against the price at that event's own news timestamp, at the stated horizon. Point-in-time, with no look-ahead: a horizon that has not yet elapsed is shown as pending with the date it matures on, never as a blank or a zero. Every published event study states its universe, N, date range, window, and convention in a standard methodology footer. If a study segments the sample, the segmentation rule is stated.

5. Corrections

Corrections leave a dated trace. Every record carries the time it was last changed, and a record we withdraw is flagged as withdrawn rather than dropped, so a correction is never a silent disappearance. Material changes are recorded in the dataset changelog, which states its own start date rather than back-dating entries it does not have. What we do not claim is the ability to reissue an arbitrary past delivery on demand; that would need a version store we do not keep.

What we publish and what we refuse to publish

We publish measured distributions: Ns, medians, percentiles, paths, and their methodology. We do not publish predictions, signals, trade recommendations, or any statistic whose sample construction we cannot document. Every event study carries the same standing disclaimer, reviewed once and never diluted per page.

Independent scrutiny

The dataset was examined in peer-reviewed research: Levi, Livnat, Zhang and Zhang (2016), available on SSRN. Separately, a leading systematic fund has licensed the data continuously for over a decade.

Frequently asked questions

What return convention do NQ event studies use?

The percent move against the price at the event's own news timestamp, measured at the stated horizon — point-in-time, with no look-ahead. Every study states its convention in the methodology footer; where a study deviates, the deviation is stated.

How does News Quantified handle survivorship bias?

Delisted and acquired companies remain in the dataset with full event history, and universes are constructed point-in-time, so distributions include the outcomes of companies that no longer exist.

Are corrections made to historical data?

Yes, and they are dated. Every record carries the time it was last changed, and withdrawn records are flagged rather than deleted, so a correction is visible as a correction. Material changes appear in the changelog. We do not claim to reproduce an arbitrary past delivery on demand.

Can I audit the methodology before licensing?

Partly, and in public. The measured class list with each class's N is event types; the same cross-section by sector is sectors; the canonical vocabulary is terminology; and every ticker page shows the individual events behind its numbers. Further documentation is provided to prospective licensees on request.

Historical reaction data. Not investment advice. Past reactions do not determine future reactions.