From papers to forecasts.
Five stages. The first and the last are fully public; the three in the middle are where our work lives.
Ingest
Title, abstract, field, date and authorship links of every paper. 163 million works from OpenAlex, every discipline, 1970–2025, for the flagship; 3.2 million arXiv papers for the earlier generations.
Recognise
Scientific concepts are found in the text with a curated vocabulary, and about 15,000 of them are kept as transferable methods after a vetting by hand; acronyms are resolved (“PINN” = physics-informed neural network) and homonyms flagged.
Map
For every year: which field uses which method and how much, which fields resemble each other, where researchers and papers sit across borders.
Learn
Several models are trained on decades of past crossings and vote together; each sees only outcomes that were already complete.
Forecast & verify
2.6 million possible crossings are ranked, and can be read through four views — from everything to classic methods entering fields where nobody uses them. The top ones are reviewed, published with a probability, a date and a fingerprint, and judged at the deadline.
What it listens to.
Four families of signals. We describe what each one captures; the exact quantities and how they are combined stay private.
Neighbourhood
Fields that share methods tend to share more. A method common among a field's closest neighbours, and missing from the field, is a gap.
Borders
Work that already sits between two fields — papers listed in both, researchers who publish in both — is often where a crossing starts.
Momentum
How fast the method is growing, how many new fields it entered recently, how young it is. Some methods are on the move; others have settled.
The field itself
How large the field is, how open it has been to newcomers in the past, how quickly it usually adopts what its neighbours use.
The test: replaying history without peeking.
Pick a year. The model is trained only on what was known then (cyan), forecasts the following three years (pink), and is scored on what really happened. Press play to walk through all eleven years. Shown here: v2 on arXiv, 2012–2022; the replay of v3 on all of science, 2008–2023, is on the track record.
Untouched crossings
Accuracy is reported separately for fields that had zero papers with the method at forecast time — the cases where following trends cannot help.
Baselines
Every score sits next to two simple strategies: picking the most popular methods, and picking at random. A model is only worth something if it beats both by a wide margin.
Preregistration
Development uses some years; the final confirmation uses a year never touched, with the success criteria written and fingerprinted before running it once.
Three generations, each one stronger.
Each generation reads more of science and listens to more signals, and each one passed the same test before being trusted. On the same untouched crossings of a year neither had seen, v2 got 44% of its top 100 right where v1 got 25%. v3 then left arXiv for all of science — fifty times more works — and still got 58 of its top 100 right in a sealed confirmation, beating the obvious forecast in every view.
| confirmation on 2023 · untouched crossings | v1 | v2 | change |
|---|---|---|---|
| top 100 fulfilled | 25% | 44% | +76% |
| top 1000 fulfilled | 13.5% | 17.8% | +32% |
| average precision | 0.062 | 0.103 | +67% |
A single preregistered run on a year never used before, with a two-year horizon. Untouched = the field had no paper with the method at the time.
The three generations side by side
The percentages are not measured on the same ground: v1 and v2 forecast on arXiv (physics, mathematics, computer science), v3 on all of science, with a different set of crossings. The fair comparison is against chance and against the obvious forecast, on the same crossings: that is where v3 moves furthest.
One model, four views
The model is never retrained for a question. What changes is which of its forecasts we look at: everything; no generative AI; classic methods; classic methods entering fields where nobody uses them. Each view passed a preregistered test on 2023.