How the swarm signal works

Rather than one black-box oracle, seven small independent signal-agents each cast a directional vote with a cited rationale, plus a catalyst event-risk modulator. The swarm's aggregate — and how much the agents agree — is the output. Unfamiliar terms are defined in the glossary.

Last reviewed 2026-07-14.

The seven agents + the event-risk modulator

Catalyst is not a directional vote: an upcoming high/medium-impact catalyst within the horizon only raises event risk, which lowers conviction. An agent with no data abstains — it never invents a number or a citation.

Weights (config sig-v4-2026-07-26)

Conviction

Conviction = swarm agreement × coverage × strength × (1 − event risk) — not a probability the call is correct. Strength scales from 0 to 1 as the absolute score |score| climbs across the band between leanTau (0.15) and strongTau (0.45); that band sets strength, not the gate. The lean gate itself is |score| ≥ leanTau (0.15): a directional lean requires that, plus at least 2 participating agents, coverage ≥ 0.35, and conviction ≥ 0.2; otherwise it is shown as “Low-confidence · Neutral.” Probability-of-correct language is reserved for the empirical calibration buckets (each shown with its N).

Accuracy, shown honestly

Every hit-rate is shown next to its base rate (max(p_up, 1−p_up), the majority-class no-skill baseline) and its N — never alone. The launch number is an out-of-sample walk-forward backtest, limited to price-derivable signals and to today's surviving company panel (survivorship bias); the durable claim is the live, out-of-sample track record the prediction ledger measures. Below the minimum shown N the live percentage is suppressed and shown as pending.