Forecasts

Calibrated probabilities.
With the receipts.

We don't predict. We produce probability distributions with explicit confidence intervals, sourced from 14 institutional-grade data feeds, and tracked against resolved outcomes. If we're wrong, you'll know — and so will we.

FORECAST · #FX-2026-0414 · 90D HORIZONUPDATING · 14 SIGNALS · 6H AGO

Probability of a sovereign-debt restructuring event in Argentina before Q3 2026.

43%
95% CI: 31% – 56%
+4.2pts past 7d
Economic
41%
Political
58%
Social
46%
Security
31%
Health
18%
Environment
22%
PRIOR · 0.34GDELT +2σFRED · CDS spreadACLED · −IODA · stableDIVERGENCE 0.21

The baseline problem

Why most forecasting fails.

Instability and change are not random, and they are not single-domain phenomena. They are preceded by compounding signals across politics, economics, security, health, environment, and the digital layer.

Most expert forecasters are poorly calibrated.
They are overconfident, anchor too hard on single narratives, and fail to update when evidence changes. Tetlock’s superforecaster research and IARPA’s ACE project converge on the same finding.
Tetlock, Kahneman, IARPA ACE
Single-domain priors miss multi-domain shifts.
The political scientist who calls elections often misses the economic inflection that caused the political shift. The economist tracking GDP misses the social instability building underneath.
The cross-domain blind spot
Hits get remembered. Misses don’t.
Most human forecasters never track their own accuracy. Without a feedback loop on resolved outcomes, calibration cannot improve — and confidence becomes a story instead of a measurement.
Why feedback matters

The four-stage architecture

A pipeline that shows its work.

Every probability passes through four sequential stages. Each one does specific, defensible work — and each one is auditable.

STAGE 01
Base rate anchoring
Every forecast begins with a calibrated reference class — what historically happens to entities of this profile, in this category, over this timeframe. The Kahneman outside view, with time-decayed analogues.
PRIOR · 0.34
STAGE 02
Bayesian signal updating
Each incoming signal updates the prior with a likelihood ratio weighted by source reliability and recency (≈24h half-life). Correlated sources are corrected — ten articles of one event are not ten signals.
0%100%PRIOR+ACLED+GDELT−FREDPOSTERIOR · 0.43
STAGE 03
Mixture over scenarios
Signals cluster into six domain paths — economic, political, social, security, health, environmental — each with its own posterior. Output is a weighted distribution. Divergence widens the CI.
ECONPOLSOCSEC
STAGE 04
Crowd calibration
Where present, expert user predictions blend in — capped at 30% influence, weighted by historical accuracy. Crowd divergence from the model is itself an informative signal.
MODEL · 70%CROWD · 30% (capped)weighted by historical accuracy

The signal architecture

14 institutional-grade sources.

A Bayesian system is only as good as its inputs. Diverse source types reduce correlated error — when news is suppressed, satellite, seismic, economic and internet-disruption layers keep working.

VELOCITY · ANCHORLIVE
GDELT
Hundreds of thousands of news sources in 100+ languages, geocoded with conflict scores.
15 MIN CADENCE
PRECISION · ANCHORLIVE
ACLED
50+ local researchers, 1,200+ non-English sources, verified field reporting.
24–72H VERIFIED
ECONOMICLIVE
FRED
816,000+ economic time series from 108 institutional sources.
108 INSTITUTIONS
HEALTHLIVE
WHO / CDC
Disease surveillance and outbreak alerts across 194 member states.
194 STATES
GEOPHYSICALLIVE
USGS
Real-time seismic monitoring with casualty estimation models.
REAL-TIME
EARTH OBSLIVE
NASA FIRMS
Near-real-time fire detection at 375-meter resolution from MODIS / VIIRS.
375 m RESOLUTION
DIGITALLIVE
IODA / Cloudflare
Internet disruption — a leading indicator of protest escalation and state crackdown.
BGP + DNS + ACTIVE
HUMANITARIANLIVE
ReliefWeb / UNHCR / IPC
Humanitarian and food-security intelligence from 4,000+ contributors.
4,000+ ORGS
CYBERLIVE
CISA KEV / NVD
Known exploited vulnerabilities and vulnerability database for cybersecurity threat signals.
KEV + CVSS
CONSENSUSLIVE
Wire services ×6
Six global wires for cross-source consensus — itself a reliability signal.
6 WIRE FEEDS
SENTIMENTLIVE
Channel telemetry
Activity spikes across high-signal Telegram and forum channels, deduped against media.
AGGREGATED
FINANCIALLIVE
CDS · FX · rates
Sovereign credit spreads, currency volatility and rate-implied default odds.
INTRADAY
CLIMATELIVE
ECMWF · NOAA
Anomaly detection on temperature, precipitation, and resource-stress indices.
DAILY GRIDDED
PROCESSEDLIVE
Brunu signal layer
Reliability weighting, decorrelation, recency decay and reflexivity adjustment over the eleven raw feeds.
INTERNAL
0.100.150.200.250.30M0M3M6M9M12BRIERSCOREMONTHS RUNNINGBASELINE · 0.250.151CURRENTCALIBRATION OVER TIME ↓ LOWER IS BETTER

The feedback loop

The system tracks its own accuracy. Then updates how it forecasts.

Every resolved forecast feeds back into the next generation cycle. We recompute our Brier score and per-category accuracy continuously. When the model has been systematically overconfident, it receives an explicit instruction to dampen confidence. When it has been underconfident, the inverse.

This is the computational analog of what distinguishes Tetlock's superforecasters from ordinary experts: the willingness to track your own accuracy and update your approach accordingly. Most forecasters never do this. We can't not.

0.151
CURRENT BRIER
−39%
VS. NAIVE BASELINE
12,418
RESOLVED FORECASTS
7
CATEGORY MODELS

Uncertainty decomposition

Honest about what it doesn’t know.

Uncertainty has two distinct components, and good forecasting communicates both. Confidence intervals on Brunu aren't decoration — they reflect the actual epistemic state.

EPISTEMIC · REDUCIBLE

What we don’t know yet.

Sparse signals, conflicting sources, prior at extremes

This narrows as more signals arrive, sources align, and the scenario distribution tightens. Brunu surfaces it as a wider confidence interval — and shows you which signals would close the gap.

0%25%50%75%100%43%
ALEATORY · IRREDUCIBLE

What is genuinely random.

A single actor's decision on a specific day

Some futures cannot be narrowed further regardless of intelligence quality. Brunu keeps this floor visible. A 43% probability with a 7-point CI is a different statement than 43% with a 25-point CI — and the system says which.

0%25%50%75%100%43%

Two things almost no one else does

Reflexive weighting. Actor fingerprints.

At a certain density, coverage stops describing reality and starts shaping it. And the same structural signal means different things depending on who is making the next decision.

REFLEXIVE SIGNAL WEIGHTING

When coverage becomes the cause.

At high media density, news ceases merely signaling underlying conditions and starts causing them. Heavy coverage of a fragile bank accelerates a run. Brunu downweights media as a structural indicator past the threshold and upweights it as a behavioral cascade predictor. The system can’t be fooled by its own inputs.

SIGNAL WEIGHTMEDIA DENSITY →structural ↓cascade ↑crossover
ACTOR BEHAVIORAL FINGERPRINTING

Probabilities are not actor-agnostic.

We build probabilistic behavioral models for key actors — states, institutions, movements, market participants — based on historical decisions, stated objectives, risk tolerance and sensitivity to environmental change. Scenarios that require an actor to behave wildly off-pattern get lower weights. It’s pattern-of-life, applied at scale.

riskspeedconsol.opencoop.rhet.ACTOR · BCRA · 2024–2026

Track record

If we’re wrong, you’ll know.

Every resolved forecast is public. Every Brier score is public. Every recalibration instruction is logged. The track record is the product.

OVERALL BRIER
0.151
Versus 0.25 random baseline. Lower is better.
RESOLVED FORECASTS
12,418
Each one tracked, scored and fed back into the next cycle.
CATEGORIES TRACKED
7
Political, economic, social, security, health, environment, digital.
SIGNAL CADENCE
15
Minutes. Velocity layer ingests, posteriors update continuously.

Built for everyone who has to make a call

However you read it.

FOR THE GENERAL READER
Most tools tell you what happened. Brunu tells you what the signals are pointing toward — and shows its work.
FOR THE QUANT
14 institutional sources. Four-stage Bayesian architecture. Self-calibrating against its own track record.
FOR THE SKEPTIC
We don’t predict. We produce probability distributions with explicit CIs, sourced from verifiable data, tracked against resolved outcomes.
FOR THE DESK
The methodology used by quantitative political scientists and intel analysts, delivered through a consumer interface, continuously recalibrated.

Browse calibrated probabilities in the open.

Open a forecast. See the prior, the signals that moved it, the scenario distribution, the CI, and the resolution criteria. Subscribe to a category. Make your own predictions.

Open BrunuDownload for iOS