Forecasting where
scarcity moves next
Vaticinus reads millions of research papers, trade flows, and supply records to work out where value moves next. It runs a deep-research engine alongside a forecasting model we train on our own leak-free data. Every call gets a date, a kill-criterion, and a score in public.
Calls we got right
Each call was made with only the information available before the question resolved, when the model knew nothing past its cutoff, then scored against what actually happened. These are selected resolved questions; the full, unfiltered scores are in the benchmarks above.
what vati does
A new kind of foresight
Despite decades of progress in computer modelling, the best predictions in the messy world of human affairs still come from top human forecasters, who use reasoning and judgement.
Vati predicts like an expert human forecaster, but with the speed and scale of a machine. Every call gets graded in public.
Our preprint shows why the field has been measuring forecasting skill wrong. Across 25 leak-free ForecastBench rounds (2024 to 2026), frontier LLMs add almost no information beyond the market price, and more scale has not changed that. Human superforecasters clear the same bar by an order of magnitude.
Vati has an edge on exactly the topics where a purely data-driven approach falls short, the ones where judgement decides the answer, not data alone.
We’ve developed our system to make medium term predictions (1 week - 1 year out) about geopolitics, business, policy, technology, and culture.
Performance
Benchmarking against human forecasters
On ForecastBench, the public benchmark the field runs, the top models already beat the best human superforecasters on the dataset half. Vati is built to land #1 among bots there, on a blind, leak-free test. The numbers below are internal, leak-free backtests run on the same public data. The official boards publish their next round in 2026.
Where Vati stands today, on the record:
Metaculus · live public record
- 154 dated forecasts, live now on the public Metaculus board
- One account anyone can open and check: the bot vaticinus, across 13 tournaments
- Nothing resolved yet. The scores land as the questions close.
- Nuclear risk
- Taiwan
- US elections
- Space & biotech
ForecastBench · dataset half
- Brier 0.124 on the dataset questions, beating the superforecaster crowd
- Built to rank #1 among bots
- Up from 0.142 after a type-aware prior rebuild
- Geopolitics
- Elections
- Technology
- Energy
Leak-free holdout (indicative)
- Brier 0.110 vs 0.245 base on a date-gated test
- 6 of 7 leak-free questions called correctly
- Small sample (N=7), marked indicative, not yet validated
- Company earnings & revenues
- Equity returns
- Bond yields
KalshiBench · vs frontier LLMs
- Brier 0.229, level with Opus 4.5 (0.227)
- ~6× lower calibration error (ECE)
- Leak-free per-category k-fold
- Politics
- Economics
- Crypto
- Climate
Every call is dated, sealed, and scored. What powers them is the layer underneath: the supply side of the world, tracked over time.
The data layer
The data most forecasters never look at
We track the supply side of the world over time. The research that flags a capability before it ships. The trade flows that show who controls a chokepoint. The mineral and component data that reveal where a bottleneck is forming, and the export controls that move it. A deep-research engine and a model we train on our own leak-free data turn that into dated, scored calls.
Forward calls
The calls on the board right now
Every call has a date, the constraint behind it, and a kill-criterion fixed the moment we make it. Some resolve within a year. The long-horizon tier looks a decade out, at limits physics and demography have already set. Open a board to read all of them.
Six places where physics or demography has already set the trajectory, while capital still looks elsewhere.
The grid runs out of people before it runs out of steel.
Cheap launch solved the wrong constraint. Here are six physical walls it cannot move.
Orbital data centers hit a heat-rejection wall before a launch-cost wall.
The semiconductor chain stripped to the chokepoints nothing cheaply replaces.
The ceiling on AI-compute wafers isn't ASML or EUV. It's one ion-beam line in Oberkochen.
Product
Providing foresight to decision-makers
Emerging use cases for predictive intelligence
Vati is built for teams operating under uncertainty, helping them cut through the noise, anticipate what’s ahead, and act with confidence.
- Macroeconomic shifts affecting demand
- Competitive moves by industry leaders
- Major regulatory changes
- Relevant geopolitical developments
- Disruptive technologies and innovation
- Commodity supply disruptions
- Legal rulings
- Central bank decisions
- Company announcements
- Geopolitical events impacting markets
- Revenue
- EBITDA
- Growth trajectory
- Key external risks and opportunities
- Industry trends
- Economic growth and inflation outlook
- Central bank policy and rate decisions
- Policy and regulatory direction
- Consumer demand and sentiment shifts
- Geopolitical developments and risks
- Competitor product launches
- Customer behaviour shifts
- Demand and growth dynamics
- Industry consolidation and new entrants
- Disruptive technologies and innovation
- New industry regulations
- Trade restrictions and tariffs
- Litigation outcomes
- Supply chain disruptions
- Cybersecurity scenarios
- Election outcomes
- Conflict escalation or ceasefires
- Migration flows and demographic shifts
- Sanctions and international policy changes
- Energy and trade disruptions
An instrument for calling the future, honestly
A generalist engine that follows an industry's causal spine: frontier, capability, dependencies, supply, demand, capital, price. It keeps only the calls the market hasn't made yet.
About us
Vaticinus is an independent forecasting instrument, built in the open and proven by its own scored record.
We didn't raise a round and write a story about it. We built a machine that makes dated, falsifiable calls, sealed them before the clock started, and let time grade them. The track record is the whole pitch. And because the thing is built to catch itself being wrong, it already has, more than once, all of it on the record.
