Forecasting Without Power: Agency Horizons, Control Collapse, and the Leverage Gap

We study a structural regime in delayed feedback systems where statistically significant prediction persists after actionable influence has collapsed. We formalize this phenomenon as Forecasting Without Power (FAWP) and introduce the agency horizon, defined operationally as the earliest delay at which an agent’s action becomes statistically indistinguishable from background variation in a noisy observation channel. Using controlled simulations of unstable linear dynamics with delayed observations, bounded control, and growing noise, we measure mutual information between actions and delayed outcomes (steering coupling) alongside predictive mutual information between current diagnostics and future states. Across a wide parameter sweep, we observe a sharp control cliff at a finite post-zero delay τ_h⁺, where steering coupling falls below a fixed threshold, while predictive coupling remains non-zero well beyond this horizon. This produces a persistent leverage gap in which prediction exceeds influence. We validate the effect using stratified estimators, shuffle and autocorrelation-preserving shift controls, and event-based checks, confirming that post-horizon prediction is not an artifact of leakage, shared state, or time-index drift. We further document a resonance-like peak in predictive coupling near the agency horizon, followed by sustained prediction in the post-cliff tail. These results distinguish statistical foreseeability from operational control and show that increased predictive confidence can arise precisely when intervention capacity has already vanished. FAWP provides a diagnostic lens for systems in which monitoring outpaces action, with implications for automated control, markets, and decision-making under latency.