GDELT Project

What it is

GDELT is an open academic project that reads world news in 65 machine-translated languages and publishes what it found as bulk files and free APIs. It is not a financial data product and does not claim to be — no prices, no tickers, no securities anywhere in it.

What brings a market-tool builder here is narrower than the project's own description suggests. Two things are genuinely usable. The DOC 2.0 API turns any query into a tone or volume time series: mode=timelinetone for "Nvidia" over seven days returned 132 hourly average-tone points, with no key and no account. And the Global Knowledge Graph files carry per-article tone and theme coding, which is where a sentiment feature for a model actually comes from.

The half GDELT is famous for — the Event Database of who-did-what-to-whom back to 1979 — is the wrong half here.

Pricing

Free, without qualification. The terms of use grant "unlimited and unrestricted use for any academic, commercial, or governmental use of any kind without fee", asking only for a citation and a link when you redistribute. That is not an SPDX licence, but it is permissive enough for commercial use.

The one cost that can appear is Google's. The BigQuery copy is hosted at Google's expense and queried at yours, with the first 1 TiB of query bytes free each month and $6.25 per TiB after that in US regions. On tables this wide, a careless SELECT * over a year of the Knowledge Graph is how that allowance disappears in an afternoon.

Data & coverage

The 15-minute cadence is real and current. At 15:30 UTC on 14 September 2026 the update feed named three files for that slot — events, mentions and the Knowledge Graph — and the master list holds 1,195,197 files running back to February 2015, when the 2.0 format began.

One 15-minute Knowledge Graph file is 6.5 MB zipped and held 1,501 article records across 27 tab-separated columns. Per record you get a tone vector of seven numbers, a GCAM affect vector of several hundred to a couple of thousand dimensions (median 984 per record, 2,355 distinct across the file), and a theme list with a usable economic vocabulary — ECON_STOCKMARKET, ECON_IPO, ECON_BANKRUPTCY, ECON_INFLATION, ECON_OILPRICE. Market coverage is not a thin slice of it: theme:ECON_STOCKMARKET ran at 3.9% of all monitored volume over the week to 14 September 2026. The TV 2.0 API returns per-station series, Bloomberg and CNBC among them, but it is not current: a Bloomberg series runs from June 2009 only to 11 October 2024, and any window in 2025 or 2026 comes back as an empty {}. Treat television as an archive, not a feed.

Integrations

Plain HTTP GET returning JSON, CSV, RSS or HTML, plus the zipped CSVs and the BigQuery tables. Nothing needs authenticating.

GDELT publishes no client library in any language, and the two packages a reader will find on PyPI are both third-party and named nowhere in its documentation: gdelt (0.1.14, Linwood Creekmore III, GPL-3.0, last uploaded November 2023) and gdeltdoc (1.12.0, Alex Whitehead-Smith, April 2025), which wraps the DOC API only. Both work, neither is maintained by the project, and the GPL-3.0 on the first is worth noticing before it goes into a product.

Limitations

• No entity resolution, and this is the whole problem. Company names arrive as lowercased strings with no ticker, ISIN, CUSIP or LEI attached. In one file softbank and softbank group were separate entries, apple app and apple health were separate "organisations", and the same column held data centers, united states, m t bank stadium and how to rob a bank. The name-to-ticker mapping, the collisions on Apple, Delta, Gap and Visa, and the subsidiary rollup are all yours to build and maintain.
• The Event Database will barely carry a company signal. Of 1,438 events in one file, 106 had a business-typed actor, and those resolved to just 23 distinct names — overwhelmingly role nouns such as AIRLINE, BANK, BOSS, BUSINESS, COMPANIES, COMPANY, CORPORATION and EMPLOYER. CAMEO does code the occasional named firm through its multinational-corporation prefix — BOEING (MNCUSABUS) and COMMERZBANK (MNCDEUBUS) both appear — but two identifiable companies per 15-minute file is not a signal you can build on.
• The coding is machine-extracted and noisy. The first ECON_STOCKMARKET record in the file opened here was a securities class-action solicitation on a press-release wire. Wire spam, syndicated duplicates and aggregator reposts sit in the corpus at full weight, and deduplication is not done for you.
• The GEO 2.0 API is dead and still documented. Its own announcement post's example URL, api.gdeltproject.org/api/v2/geo/geo?query=trump, returned HTTP 404 on 14 September 2026. Other pages are stale rather than broken — the data page still promises a 2.0 upgrade to the Analysis Service "later this year", years after the fact.
• Modes disagree with each other. theme:ECON_STOCKMARKET returned a populated volume timeline and, as the identical query in mode=artlist, an empty {}.
• One request every five seconds, answered as HTTP 429. With no key that ceiling is per address, so anything at scale means the bulk files or BigQuery rather than the API.
• Tone is not financial sentiment. It is general-purpose affect scored over whole documents in a corpus that is mostly not about markets. Whether it predicts anything is your research problem, and the honest prior is that it mostly does not.

Alternatives

If what you want is headlines already tagged with the tickers they concern, this is the wrong tool. Marketaux, NewsFilter.io, Benzinga News API, StockNewsAPI and the free TickerTick all resolve entities to instruments, which is the expensive part. GDELT earns its place only when you want breadth no financial vendor sells — non-English coverage, macro and geopolitical themes, television airtime — and will build the mapping layer yourself. The rest of the market data category covers the priced side.