Methodology

The headline metric is a weighted sentiment index on a fixed −1…+1 scale, where −1 reads as consistently negative coverage and +1 as consistently positive. The index is reported per country and as a global aggregate, with breakdowns by topic and by financial-markets sub-segment.

The metric is calibrated to be intensity-aware: an article that is strongly negative contributes more than an article that is mildly negative, and the confidence of the model in its own judgment is folded into the weight. Deltas are reported in points (×100), never in percent: the index is signed and bounded, so a true percentage change is undefined near zero.

Coverage spans approximately 11,000 active news sources across 193 countries and territories and 20,000+ publicly listed companies. The source list is curated and pruned continuously; it combines mainstream national press, business and financial outlets, regional press, a long tail of country-specific publications discovered through five independent discovery passes, and international wire services.

Articles are collected through public news feeds. Each article carries its publication date, the source outlet, and our derived country attribution. We do not crawl paywalled bodies; the analysis runs on the publicly available metadata, title and lede where one is provided.

Each source carries two static attributes: a credibility score(an editorial rating of the outlet’s journalistic standards) and the press-freedom factorof the outlet’s home country (derived from the V-Dem freedom-of-expression index). Both enter the headline weighting (see below).

Each article is scored by a multilingual machine-learning model, fine-tuned for news rather than for product reviews or short-form social media. The model returns a probability distribution over five sentiment intensities (from strongly negative to strongly positive), from which we derive two outputs:

• Expected polarityon a −1…+1 scale, computed as the expected value of the five-point distribution. This is the per-article number that aggregates into the headline metric.
• Entropy-normalised confidenceon a 0…1 scale. An article on which the model is near-uniform across the five intensities (i.e. uncertain) carries a low confidence; an article on which the model concentrates mass on one intensity carries a high confidence.

A small fraction of multilingual short headlines on which the base encoder is unreliable is escalated to a large multilingual language model whose verdict is re-scored onto the same scale. Both outputs are written into the same store with a marker recording which model produced each row; the headline aggregation reads them identically. The article’s confidence still contributes proportionally to its weight.

On the headline surfaces, every article’s contribution is weighted by credibility × press-freedom × confidence. Credibility down-weights outlets whose journalistic standards are weaker; press-freedom factor down-weights outlets operating in environments where coverage is constrained; confidence down-weights articles on which the model itself is uncertain. A low-confidence article is not discarded — it simply counts for less than a high-confidence one.

The press-freedom factor is a bounded curve over the V-Dem index: a free-press outlet is at full weight, and a captured-media outlet is down-weighted but never silenced (a floor of 0.10guarantees representation). The adjustment is concentrated on the small number of countries whose own coverage is dominated by their own low-freedom domestic outlets; coverage from free-press environments is essentially unaffected.

This is a deliberate, transparent stance and the only methodological choice that materially differentiates the headline metric for state-aligned versus free-press countries. It is documented here because it should not surprise a customer reading the numbers.

Many articles concern more than one country (trade disputes, summits, regional crises). The system records every country a story concerns, with a mention-weighted score and an is-primary flag for the principal subject country. When a per-country aggregate is requested, the article counts once toward each country it attributes to; the headline index is not double-counted across countries because the aggregation operates per article-country pair, not per article.

Country attribution is multilingual and recognises both explicit mentions and standard wire-service datelines (e.g. “BAGHDAD (Reuters)” resolves to Iraq even when the country is not named in the lede).

Each article is classified into the international IPTC Media Topicstaxonomy — the standard used by Reuters, the Associated Press and major newsrooms. The taxonomy distinguishes seventeen primary topics including politics, economy and business, conflict and war, crime and justice, health, education, science and technology, arts and culture, lifestyle, environment, weather and disasters.

On the dashboard these seventeen IPTC categories are collapsed into six readable buckets — Politics, Economy, Society, Science & Technology, Environment, and Lifestyle — for the per-country topic snapshot. Sport is recognised by the topic model but is excluded from every headline aggregation: short-term fan-feeling is not a societal or economic signal, and including it would noise out the index. This is a single, explicit policy enforced in the backend; it cannot be opted out on a per-page basis.

Markets news is identified by a dedicated multilingual markets classifierthat combines language-agnostic strong rules (named indices, central banks, FX pairs) with a zero-shot transformer for the long tail of non-English markets coverage (Korean, Arabic, German, Spanish and others). The classifier is applied to economy-topic articles and marks each as “in markets” or not.

The markets sub-index uses the samecredibility × press-freedom × confidence weighting as the headline metric, restricted to the markets slice. This guarantees that the markets number reconciles with the headline number on the same data, and that the two cannot drift apart through alternative weightings.

On the markets monitor each region is also overlaid with its main equity index as a percent change (using listed tracking ETFs as regional proxies). The overlay is updated hourly and is for context only; it is not part of the sentiment metric.

The Economy Monitor reports what economic story the news is telling, across nine macro narratives: inflation, employment, growth, consumer, international trade, state finances, interest rates, energy and housing. Classification is done by a local large language model trained for this taxonomy; per article, the model returns the narrative and a separate economic-sentiment rating.

Macro sentiment is a deliberately separate scale from the credibility-weighted headline metric — it is the language model’s read on the economic situation described in the article. Mixing it with the headline number would conflate two distinct measurements. The Economy Monitor and the Sentiment headline can move in different directions for the same period; that is a feature, not a contradiction.

For a subset of countries each narrative time series can be overlaid against the matching official statistic (consumer price inflation, unemployment rate, policy rate, oil prices, house prices). The official series come from the Federal Reserve Bank of St. Louis FRED database, which mirrors OECD-harmonised series for the major economies. The overlay is informational and does not enter the narrative score.

The Framing surface is deliberately the only headline- adjacent view that does not apply the press-freedom weighting. Framing compares how a country talks about itself in its own media against how the rest of the world reports on the same country. Down-weighting state-aligned outlets there would hide the signal it is intended to detect.

The per-article perception layer additionally records a per-country stance (from strongly critical to strongly supportive of the subject country) and a frame label (is this a conflict story, an economic story, a governance story, a social story, and so on). Stance and frame are emitted by the same multilingual language model that produces the headline verdict and are recorded per article-country pair.

Live coverage updates continuously. The long-range history is assembled from two complementary instruments. First, a set of model-consistent news anchors— major long-running outlets whose archives are re-scored end-to-end by the same live model on the same scale (the New York Times since 1999, the Guardian since 1999, Reuters since 1999, Le Monde since 2000, the Straits Times since 2010, the Times of India from 2010). Second, the GDELT machine-coded news tone series (since 2000), which is dense in time but is a different instrument on a different scale.

We combine these into a single canonical monthly per-country historical sentiment series using a constrained ensemble that fits the historical anchors and GDELT against recent live data, then projects the fit backwards. The resulting line is the official historical reference for the dashboard. It is out-of-sample validated against the most recent year not used in the fit, and the validation r-squared and confidence interval are available on request.

Where comparable history is not yet sufficient (typically for long horizons or for very thin per-topic country slices), the dashboard reports “—” rather than a noisy number.

The Signals page surfaces anomalies and velocity events for each country and topic. A signal fires when one of two detectors crosses a threshold:

• Sentiment z-score. Comparison of the current seven-day weighted sentiment against a robust ninety-day trailing baseline. The baseline centre uses a smoothed median; the dispersion uses the raw daily series. Single-day bursts cannot dominate a baseline by construction.
• Velocity. Week-over-week change in the same weighted index. Sustained one-week shifts above a calibrated threshold fire as velocity signals.

Only the stronger of the two detectors fires per country×topic×day, and a deduplication key prevents the same scope re-firing the same day. Scopes with insufficient article volume or insufficient baseline coverage are skipped rather than reported with a noisy number.

The Global Sentiment Monitor is a measurement of what the news is saying, not a measurement of underlying conditions on the ground. Sentiment can lead, lag or contradict economic and political reality. Used well, the metric is most powerful as an early warning — surfacing what reporters and editors are choosing to write about, and how, before that consensus consolidates into the official statistics.

The metric does not eliminate the editorial and selection biases of the underlying media; it characterises them through the press-freedom weighting and the framing surface. Headline numbers for countries with sparse English-language coverage are estimated on a smaller sample and are more volatile by construction.

The Monitor is a research and analytical product. It is not a financial recommendation, a credit signal, an investment-grade rating, or a substitute for direct reporting on a story of interest. The numbers should always be read alongside the underlying articles, which are linked from every surface that shows an aggregate.