Hybrid Forecast Intelligence

900+ forecasters predict upcoming US macroeconomic releases and other variables of interest through a private operator-staked prediction market (Almanis) with a 7+ year median forecaster tenure. The platform generates two concurrent time series at minute-level resolution: the general market consensus and an ML-filtered insight signal that amplifies forecasters identified as disproportionately accurate in real time. The third output is a z-scored divergence measure between the two, quantifying how far the insight estimate has moved from the crowd.

Approximately 95% of the time, the insight signal confirms the general consensus. This is informative: it indicates that the crowd is well-calibrated and discourages premature repositioning. In aggregate across 174 election markets, the two signals perform similarly (AUC 0.87 vs 0.86; eBioMedicine, 2023). In the US macroeconomic domain, aggregate accuracy is comparable; the insight advantage concentrates in high-divergence episodes. The insight advantage concentrates in the roughly 5% of episodes where divergence exceeds |z| > 1.65. In these rare periods, the general market forecast error rises significantly (p = 0.011) and the insight signal carries a measurable accuracy advantage (p < 0.001).

When strong divergence emerges, the general consensus has frequently converged toward the insight estimate, typically within 3 to 14 days in the backtest period. The system captures this drift—it does not depend on how the underlying macro question eventually resolves. Alpha is generated in the window between divergence onset and consensus revision, well before the scheduled release.

The feed is delivered at two levels. At question level: macro indicator, forecast period, question wording, settlement source. At trade level: millisecond-timestamped general and insight point estimates, hybrid–general divergence, z-score, and binary alert indicators. Position sizing, instrument selection, leverage, and risk parameters are determined by the client.

CPI and NFP are the strongest-performing macro sleeves across the backtest. PCE performance is weaker, consistent with its role as a confirmation of already-priced CPI constituent data. Coverage expansion is available through partner co-sponsorship of new question series.

The insight signal methodology is published in the Journal of Financial Markets (Bossaerts et al., 2024) and the ML augmentation in eBioMedicine (Gruen et al., 2023), both replicated on independent DARPA NGS2 programme data. The backtest covers 2019–2022 (1,189 market days) across six macroeconomic regimes, with 4,710 accepted trades on 7 FX pairs and 51 ETFs. Backtest excludes GDP growth and employment categories (insufficient event frequency for statistical evaluation). A controlled forward test (Phase 3) commenced September 2025 with paper trading across three independent $1M pods to calibrate signals IP. Weekly public signal hindsight reporting via Substack.

Past performance is based on backtested data and is not indicative of future results.