← All research storiesFarming & food

PreprintFarming & food

Machine learning beat traditional methods at forecasting harvests across six continents, but bad years are still hard to predict

Researchers compared five kinds of models on crop data from more than 12,000 regions and found machine learning usually wins, yet forecasting year-to-year swings remains tough.

How far alongPreprintA manuscript shared publicly, often before or alongside journal review. Check its separate journal and review status (each story says what we know). Results may change.
How soon could it reach you?Small stepA small update to earlier work. Useful to researchers, but not a change you would notice yet.
Where it comes fromarXiv (preprint server)Nov 14, 2026Read the original →

What the scientists did

Knowing how much maize or wheat a region will harvest helps governments plan food supplies and warn of shortages. Scientists have built many kinds of forecasting models, but they are rarely compared fairly on the same data.

The team used CY-Bench, a shared test set of 63 country-and-crop datasets covering more than 12,000 regions on six continents. They compared five approaches: simple statistical baselines, process-based crop models that simulate how plants grow, standard machine learning using hand-chosen inputs, deep learning models that read data over time, and newer “tabular foundation models.” Each was scored on overall accuracy and separately on predicting differences between places, changes from year to year, and unusual years.

Words to know

Crop yield

How much crop is harvested from a given area of land.

  • Bushels of corn per acre
  • Tons of wheat per hectare

Features

Measurements that people choose and prepare to feed into a model. The abstract doesn’t list which ones were used; these are common examples.

  • Rainfall during the growing season (example)
  • Average temperature (example)
  • Satellite measures of how green crops are (example)

Anomaly

A year that is unusually good or bad compared with normal.

  • A drought year with a poor harvest
  • A bumper-crop year after ideal weather

What they found

Standard machine learning with hand-chosen features and tabular foundation models performed best overall. Data-driven models lowered forecast error in 72% of countries growing maize and 78% of countries growing wheat, with typical (median) error reductions of 8.8% for maize and 11.9% for wheat.

Every type of model found it much easier to tell which places produce more than to predict how a harvest will change from one year to the next or to catch unusual years. Machine learning did better than simple baselines on those harder tasks, but forecasting year-to-year swings remains challenging.

12,000+regions on six continents in the test
78%of wheat-growing countries where data-driven models cut error
11.9%median error reduction for wheat (8.8% for maize)

Why it matters to you

Food prices and food aid depend on knowing early when harvests might fall short. This study shows AI already helps, but the hardest and most important part, spotting a bad year ahead of time, still needs work. Think of it as an honest progress report rather than a breakthrough.

Where this could lead

Somewhat better harvest forecasts

Agencies could favor the best-performing model types to make crop forecasts more accurate.

1–3 years

Fairer model comparisons

The open benchmark, code and results dashboard let other researchers test new ideas on the same data.

Now

These are possibilities the research points toward, not promises. Most early findings take years of testing before they reach everyday life, and some never do.

Why scientists care

Many AI studies report big gains on small, hand-picked data sets. A large, standardized comparison shows where AI truly helps and points to what’s missing: better information about the weather and farming choices that make one year different from the next.

Keep in mind

This is a preprint. It has been accepted to the KDD 2027 data-science conference, but the final version will appear later. The test covers 63 country-and-crop datasets, so results may differ for other crops or smaller areas. The error reductions are medians, and in some countries data-driven models did not reduce error.

How far along is the science?

  • Peer-reviewedReviewed by independent experts for a journal before publication. Peer review is a quality check, not a grade of evidence strength: confidence still depends on study design, limitations, replication and the wider evidence.
  • PreprintA manuscript shared publicly, often before or alongside journal review. Check its separate journal and review status (each story says what we know). Results may change.
  • Conference talkPresented to other scientists at a meeting. Usually short and early; a full paper may come later.
  • Agency reportPublished by a science agency such as NASA, NOAA or NIH. Reviewed internally, but not by an outside journal.
  • Press releaseAn announcement from a university or company. Useful, but always check the research it describes.

How soon could it reach you?

  • Could reach you soonCould show up in products, services or advice soon. Our rough editorial guess is about 1 to 3 years, not a forecast from the research.
  • A few years outPromising, but needs more testing or engineering first. Our rough editorial guess is several years, with a lot of uncertainty.
  • Long-term scienceBasic research that builds knowledge. Real-world uses, if any, are far off.
  • Small stepA small update to earlier work. Useful to researchers, but not a change you would notice yet.

Every story links to its original source: the paper itself, or the conference program when no paper is out yet. Explanations are written with AI, usually from the paper’s abstract (each story says what it was written from), then checked against the source. We only use journals, recognized preprint servers, science agencies, scientific conferences and established science news outlets.