Inside stable regimes, distributional methods aggregate cleanly. At the boundary where the regime itself is shifting, those methods go silent. A small cohort of human forecasters produce their most accurate signals precisely at that boundary. We extract three concurrent signals from the microstructure of an operator-staked forecaster panel of 900+ contributors. One tracks the public consensus benchmark. Two are private.
Peer-reviewed methodology in the Journal of Financial Markets and eBioMedicine. Replicated on independent DARPA programme data and, in March 2026, on a structurally different public forecasting venue. A forward test is validating the system in real time. A systematic prop trading mechanism prices the IP.
The crowd forecast. Aggregated from 900+ forecasters with 7+ year median tenure. Tracks the public consensus benchmark for each release. Peer-reviewed: JFM (2024), eBioMedicine (2023).
Separation between the crowd consensus and an ML-identified cohort whose accuracy advantage is regime-conditional. Lives in the microstructure. Peer-reviewed: eBioMedicine (2023).
Z-scored magnitude of Signal 2, weighted by cohort track record. When it spikes (z ≥ 1.65), the consensus typically revises toward the insight estimate. Documented in private whitepaper (2026).
Signal 1 is public. Every consensus-aggregation system in the world produces it. Signal 2 requires identifying who in the forecaster population generates disproportionate accuracy under conditions of elevated uncertainty—and measuring the separation between their estimate and the crowd's. Signal 3 scores that separation by statistical magnitude and historical reliability. The system detects when the crowd's aggregate wisdom is breaking down and something else is replacing it.
Signals 1 and 2 are documented in peer-reviewed publications in the Journal of Financial Markets (2024) and eBioMedicine (2023), both replicated on independent DARPA programme data. Signal 3 is documented in a private whitepaper available to qualified parties under NDA.
Research →A controlled forward test commenced March 2026 with paper trading across three independent $1M pods to validate signals out of sample. Weekly signal hindsight reporting is published via Substack with full transparency on hits, misses, and methodology.
Qualified institutional evaluators can request access to the full trade-level dataset and performance metrics during the test period.
Follow the forward test on Substack →The signal depends on the quality of the forecaster network. 900+ active forecasters curated from 36,000+ candidates over 17 years, with a 7+ year median tenure on the platform. If you have a track record of thinking carefully about uncertain outcomes, we'd like to hear from you.
Join the Almanis network →Past performance is based on backtested data and is not indicative of future results.