Statistical learning theory / Open question
How Fast Can Forecasts Become Calibrated?
Reliable probability forecasts should become calibrated even when outcomes are chosen adversarially. The unresolved exponent measures the unavoidable speed of that process.
Seed catalog · Showing the bundled record.
Problem statement
In online binary forecasting, the learner predicts while a non-anticipating adversary chooses . Define and . Determine the minimax growth of , closing the published gap . An improved exponent on either side is a concrete intermediate target.
Assumptions and scope
- Both players may randomize and condition on past predictions/outcomes, but the adversary does not see the current realized prediction before selecting the current outcome. The horizon is known.
- The learner may choose its finite prediction grid as a function of T. The sum is over prediction values actually used, not fixed coarse bins; use cumulative l1 calibration, not smooth calibration or distance-to-calibration.
- The objective is expected error, with expectation over both players' randomness. Computational efficiency is a separate strengthening; do not replace this minimax objective by regret under a proper scoring rule.
What would settle it?
Matching asymptotic upper and lower bounds, or a rigorously improved upper exponent below 0.662942288 or lower exponent above 0.54389, with any logarithmic factors and adversary restrictions explicit.
Selected for a theoretical or computational target. A fitted scaling curve or improved benchmark score alone does not establish the stated learning-theoretic guarantee.
Status and evidence
Dagan et al. break the classical 2/3 upper barrier and prove a 0.54389 lower exponent even against oblivious adversaries. Dai and Fishelson supply the sharper explicit upper exponent. Recalibration, calibeating and alternative calibration metrics have different solved bounds and are not this target. This is an editorial reformulation of a published open problem. The preliminary review through 2026-11-19 does not certify novelty or exclude an unlocated resolution.
Preliminary source check; not an expert-verified open status or an independent audit of proposed proofs. The compilation date is not a literature-review cutoff.
Sources and further reading
- Dai and Fishelson — Explicit Asymptotic Bounds for Sequential Calibration Beyond T^(2/3) ↗problem statement · source checked 2026-11-19
HTML Theorem 1.1 and Section 2.1 checked: expected cumulative l1 error O(T^0.662942288); the introduction asks how far the exponent can be reduced. This improves the upper bound without closing the minimax gap.
- Dagan et al. — Breaking the T^(2/3) Barrier for Sequential Calibration ↗partial theory · source checked 2026-11-19
PDF pp.1–3 checked: cumulative l1 calibration, simultaneous non-anticipating play, sub-2/3 upper bound and Omega(T^0.54389) lower bound even for an oblivious adversary. Other calibration metrics can have different rates.