Machine IQ: A Declared-Window Summary of Forecasting Skill Across Horizons
Dr.Jagadeesh Selvaraj
Abstract
The intelligence quotient is a position in a reference group, built from a raw score, a norm group, and a rescaling. It works for people because the reference group is stable, the test items are unseen, and the test is short. Each of these conditions fails for artificial systems: items can be in the training data, scores reach the top of the scale within a few model generations, questions are answered at once while useful work takes days, and the world in which the work is done keeps changing. This article presents a framework that replaces the fixed test with probability forecasts of events that resolve at different horizons, scored by a proper scoring rule. We plot skill against the logarithm of the horizon, and averaging that curve with the one weight that doesn't change when the unit of time changes gives a unit-free quantity, Φ. Placed in a named reference group, Φ becomes the Horizon Intelligent Quotient (HIQ), called the Machine IQ when the reference group is machines. A four-part profile (Reach, Efficiency, Adaptivity, Breadth), a ceiling set by the world's predictability, a model of error accumulation in long tasks, a treatment of changing worlds, a study-design calculus, rules for fair human–machine comparison, and a set of guarantees complete the framework. The Machine IQ is a newly defined construct. It is normed here only in one round of ForecastBench forecasts (141 forecasters, 105 questions, five horizons), and every illustrative number elsewhere is constructed. The main closed-form results are checked by simulation. On real ForecastBench forecasts, a flat pooled skill curve conceals domain curves that decay, rise, or stay near zero; several rules of thumb fail and are corrected; and Φ ranks forecasters almost as a pooled score does on a balanced grid of horizons and predicts a model’s later rank equally well, but stays stable when the mix of horizons changes, which a pooled score does not. The claim is deliberately narrow. On a balanced grid of horizons, Φ is equivalent in practice to a pooled score; its value lies in fixing the window and the weight in advance, which keeps scores stable and comparable when horizons are unevenly resolved, and in reporting the skill curve, the ceiling, and the uncertainty alongside it.