Agricultural news signals are claims pulled out of agricultural reporting in Kenya, Tanzania, Uganda and Rwanda and kept as rows. A row gives the commodity, the country and, if the report gives one, the county. It also carries the signal type, direction and horizon, a confidence score, the report title, the link to the report and a verbatim quote that supports the claim. So the news signals dataset records what people are saying, not what was measured.
That distinction decides every way the dataset gets used. A price in the market table is an observation: a market, a day, a unit. A signal is a sentence someone printed, cut down to fields. Its value: it arrives early and points to where you can check it. Its limit: nobody has verified it, and the dataset does not pretend otherwise.
What one row holds, field by field
Commodity is what the claim is about: coffee, cow milk, diesel, tea, and so on. Country is always there. County appears only when the report named one, so a national claim has no county and should not be forced into one. Signal type tells you what kind of claim it is: a price level, a price direction, a weather shock or a government programme, for example.
For price calls, direction is up or down. Horizon says how far the claim reaches: now, weeks to months, or a season or more. Confidence is a score attached to the extraction, and the rows counted below are the ones rated medium or high trust. The source link and the verbatim quote close the loop. If a row says coffee is rising, the quote gives the exact words and the link takes you to the whole report.
The quote field is the part worth defending. A label like direction up is the dataset's reading of a sentence. The sentence itself lets you disagree with that reading in seconds. A signal dataset that hides the quote asks you to trust a classifier you cannot inspect. That is a poor bargain for anyone who has to defend a number to a credit committee or a programme board.
Take the quote as the unit of audit. One recent coffee row carries these words: the average coffee auction price rose to USD7.21 per kilogramme in 2025 from USD4.90 in 2024, citing national statistics in the report. The row records that the report said so. It does not certify the figures. A careful user would fetch the auction results directly before repeating them.
The latest month, counted
Between 11 September 2026 and 11 October 2026, the dataset holds 2,027 signals rated medium or high trust. By country: Kenya 1,260, Uganda 421, Rwanda 176 and Tanzania 170. Those four add to 2,027 (1,260 plus 421 plus 176 plus 170), so the split is complete. Kenya carries roughly 62 percent of the rows (1,260 divided by 2,027). That is arithmetic on those counts, not a statement about how much farming news each country produces.
Keep that caution in mind. A country share here reflects how much agricultural reporting was picked up and rated trustworthy. It does not reflect the size of any country's agriculture. Tanzania's 170 rows are not evidence that little is happening in Tanzanian farming. They show how much of it reached this dataset in that window at that trust level. Read the counts as coverage of reporting.
Split the month by horizon: 1,010 now, 608 a season or more and 409 weeks to months. Those three add up to 2,027. About half the rows describe the present. That matches reporting's habit of describing what has just happened. Only a minority look a season or more ahead. Those rows are where a quote deserves the most scepticism.
What kinds of claims dominate
Eight signal types are listed for the month: government programme 360, production 294, price level 266, price direction 187, policy 153, weather shock 120, supply direction 116 and input cost 115. Their total is 1,611 (360 plus 294 plus 266 plus 187 plus 153 plus 120 plus 116 plus 115). That leaves 416 of the 2,027 rows outside these eight. This is only a partial list. The other categories are not itemised here, so don't read these as shares of the whole month.
Even so, if you expected a price feed, the biggest listed category is government programme, not price. Agricultural reporting is largely about announcements and output. The dataset reflects that honestly.
Price level and price direction add up to 453 rows (266 plus 187). A market analyst reaches for that slice first. But the two types are different. A price level claim puts something at a figure. A price direction claim says it is moving. Neither one is a sale price. A level quoted in print can be an asking price, an average or a figure cited from elsewhere.
Policy, weather shock, supply direction and input cost each run between 115 and 153 rows. These are the categories where a signal can flag something the price tables will only show later. Examples: a weather shock reported before a harvest estimate, or an input cost claim before a fertiliser series moves. The word is can. The dataset records the claim, not whether a later price confirmed it.
The price calls, and the one figure that looks odd
For the same month, price calls by commodity break into up and down like this: coffee 20 up and 25 down; cow milk 41 up and 2 down; diesel 9 up and 12 down; petrol 8 up and 11 down; tea 5 up and 9 down. These counts are claims, not sizes of moves.
The odd one is cow milk: 41 up against 2 down. No other commodity on the list is that lopsided. Coffee is nearly balanced, while fuel and tea lean mildly down. A lopsided count can be read at least two ways. Milk prices may have been rising in the reporting. Or a handful of widely repeated stories, each rewritten by several outlets, may have turned few underlying events into many rows. The dataset gives you a way to tell which: sort the 43 milk rows (41 plus 2) by source link and count distinct reports.
Run that check before you use the count at all. Duplicate coverage is the usual way a claim-counting dataset overstates agreement. Counts of signals measure attention. They only become evidence about prices once you compare them with an observed series, such as the market prices by month, to see whether the two agree.
Also note what's absent. The platform doesn't forecast prices. A direction field records what a report said; it isn't the platform's own call. The earlier post on commodity data from East Africa: market prices and auction results set out the observed price tables these claims should be read against.
Why labelling direction is harder than it looks
Research precedent shows how much analytical work sits behind a direction label. One proposed methodology, described in a farmdoc paper on whether price spikes always stem from news, collects agricultural news, detects price jumps non-parametrically, and then separates exogenous, news-driven spikes from endogenous ones unrelated to news. For a data user, the lesson is that news and price movement are not automatically linked. Figuring out which spikes news explains is a separate step.
A second line of work, a University of Georgia study of agricultural supply news as macroeconomic shocks, reports volatility-adjusted impulse responses linking commodity prices, a reduction in domestic real GDP, and effects on equity markets. That's a different setting from East African reporting. It shows news can be a measurable shock when classified carefully.
Neither paper describes this dataset or tests East African signals. What they offer is a standard to hold the rows against. The direction field records the claim in the report. It doesn't decide whether that claim was news-driven or already priced in. If you need that distinction, build it yourself by joining signals to observed prices and testing the timing.
Where the grain stops: counties and countries
County is optional in a row, and that's the first gap. The second gap: don't assume finer geography from the source side. The AICCRA account of the Kenya hub says volunteer rain gauges collect data in eight counties, downscaled seasonal models cover eight counties, and the Kenya Meteorological Department produces county forecasts at a resolution of up to 0.5 kilometres. Its stated next step is all 47 counties.
A signal that cites a county outside those eight inherits no finer rainfall grain from that programme. A news claim about a dry spell in a named county is just a sentence in print. Unless you check that a gauge network exists in that county, the claim isn't backed by one.
The same material shows another trap. The Kenya reach figure of 700,000 farmers appears under two platform names across two accounts from the same programme. The ILRI account of climate information services uses the name Kenya Agricultural Advisory Platform. The AICCRA story uses Kenya Agricultural Observatory Platform. Both give 700,000. A row with a source link and a verbatim quote resolves that ambiguity, because you can see which name the report used instead of guessing. Those figures count farmers reached, not rows in any dataset, and nothing here makes the hub a source of signals.
Coverage inherited from sources that name other countries
The dataset spans four countries, but much of the surrounding literature doesn't. The AICCRA story says AgData Hubs launched across six focus countries: Ethiopia, Ghana, Kenya, Mali, Senegal and Zambia. Tanzania, Uganda and Rwanda are not among them. The ILRI figure of 1.5 million farmers and livestock keepers reached, almost a third of them women, covers Ethiopia, Kenya and Senegal only. It can't be carried to the other three countries.
Household surveys have their own limits. The ten-facts paper on input use in sub-Saharan Africa, published in Food Policy in February 2017, uses LSMS-ISA data from over 22,000 households and 62,000 agricultural plots across six countries, including Tanzania and Uganda. Both waves are 2010/11. Rwanda is absent. For input use in the region, that survey is the nearest thing to ground truth. It predates the signals window by over a decade, so a signal about fertiliser use today cannot be checked against it directly.
The paper's point is that macro-level statistics cannot capture within-country heterogeneity and tend to have reliability problems. That supports row-level evidence. It is also a warning: a news claim gets less verification than a macro statistic. The EAC guidelines for agricultural statistics seek coherence across statistical sub-domains. Signals alone do not give you that.
What a signal answers, and what it cannot
A signal tells you what is claimed, for which commodity and place, how far ahead, and where the proof sits. Early on, you see attention to a topic shift. In the latest month, 120 rows are weather shocks and 115 are input cost claims. If those counts change sharply next month, that is a prompt to look at the observed series.
It can't give you a price. The observed table does that. Kenyan county prices from the official market system cover about 195 commodities in all 47 counties since May 2021, and fertilizer at market level in 29 counties. Coffee auction results list minimum, average and maximum price per grade. A news signal about coffee direction belongs next to those numbers, not in place of them.
Forecasting is not its job either. The platform gives you observations and arithmetic on observations. Price forecasts and yield predictions were built, tested against simply carrying today's figure forward, lost, and never shipped. A horizon field records what a report claimed about the future. Confidence scores the extraction. It is not the probability that the claim comes true.
Using the rows without fooling yourself
- Filter to the type you need, then read the quote column before the label column.
- Deduplicate by source link, because one story can seed many rows.
- Treat a missing county as national or unknown, never as a default county.
- Join to an observed series, such as the monthly market prices, and ask whether the signal led, lagged or was simply wrong.
- Keep Tanzania and Rwanda counts apart from Kenya's, since shares move on a handful of rows when the counts are 170 and 176.
The opposite trap is just as real. Since 1,260 of 2,027 rows are Kenyan, any regional average is a Kenyan average with three other countries attached. Report by country.
One more limit: the figures here cover a single window, 11 September to 11 October 2026. A single month is a count, not a trend. A change from the previous month would mean something only with that month's rows alongside. Used within those limits, the dataset is a fast, auditable index of what agricultural reporting says across four countries. Always check each claim against the link in the row.