An Elementary Predictor Obtaining $2\sqrt{T}+1$ Distance to Calibration

Probabilistic predictions of binary outcomes are said to be calibrated, if, informally, they are unbiased conditional on their own predictions. For predictors that are not perfectly calibrated, there are a variety of ways to measure calibration error. Perhaps the most popular measure is Expected Calibration Error (ECE), which measures the average bias of the predictions, weighted by the frequency of the predictions. ECE has a number of difficulties as a measure of calibration, not least of which is that it is discontinuous in the predictions. Motivated by this, Błasiok et al. (2023) propose a different measure: distance to calibration, which measures how far a predictor is in $\ell_{1}$ distance from the nearest perfectly calibrated predictor. In the online adversarial setting, it has been known since Foster and Vohra (1998) how to make predictions with ECE growing at a rate of $O(T^{2/3})$. Qiao and Valiant (2021) show that obtaining $O(\sqrt{T})$ rates for ECE is impossible. Recently, in a COLT 2024 paper, Qiao and Zheng (2024) showed that it was possible to make sequential predictions against an adversary guaranteeing expected distance to calibration growing at a rate of $O(\sqrt{T})$. Their algorithm is the solution to a minimax problem of size doubly-exponential in $T$. They leave as an open problem finding an explicit, efficient, deterministic algorithm for this problem. In this paper we resolve this problem, by giving an extremely simple such algorithm with an elementary analysis.

We study a sequential binary prediction setting: at every round $t$, a forecaster makes a prediction $p^{t}\in[0,1]$, after which an adversary reveals an outcome $y^{t}\in\{0,1\}$. Given a sequence of predictions $p^{1:T}$ and outcomes $y^{1:T}$, we measure expected calibration error (ECE) as follows:

Following Qiao and Zheng (2024), we define distance to calibration to be the minimum $\ell_{1}$ distance between a sequence of predictions produced by a forecaster and any perfectly calibrated sequence of predictions:

where $\mathcal{C}(y^{1:T})=\{q^{1:T}:\mathrm{ECE}(q^{1:T},y^{1:T})=0\}$ is the set of predictions that are perfectly calibrated against outcomes $y^{1:T}$. First we observe that distance to calibration is upper bounded by ECE.

The upper bound is not tight, however. The best known sequential prediction algorithm obtains ECE bounded by $O(T^{2/3})$ (Foster and Vohra, 1998), and it is known that there is no algorithm guaranteeing ECE below $O(T^{0.54389})$ (Qiao and Valiant, 2021; Dagan et al., 2024). Qiao and Zheng (2024) give an algorithm that is the solution to a game of size doubly-exponential in $T$ that obtains expected distance to calibration $O(\sqrt{T})$. Here we give an elementary analysis of a simple efficient deterministic algorithm (Algorithm 1) that obtains distance to calibration $2\sqrt{T}+1$.

Before describing the algorithm, we introduce some notation. We will make predictions that belong to a grid. Let $B_{m}=\{0,1/m,...,1\}$ denote a discretization of the prediction space with discretization parameter $m>0$, and let $p_{i}=i/m$. For a sequence of predictions $\tilde{p}^{1},...,\tilde{p}^{t}$ and outcomes $y^{1},...,y^{t}$, we define the bias conditional on a prediction $p$ as:

To understand our algorithm, it will be helpful to first state and analyze a hypothetical “lookahead” algorithm that we call “One-Step-Ahead”, which is closely related to the algorithm and analysis given by Gupta and Ramdas (2022) in a different model. One-Step-Ahead produces predictions $\tilde{p}^{1},...,\tilde{p}^{T}$ as follows. At round $t$, before observing $y^{t}$, the algorithm fixes two predictions $p_{i},p_{i+1}$ satisfying $\alpha_{\tilde{p}^{1:t-1}}(p_{i})\leq 0$ and $\alpha_{\tilde{p}^{1:t-1}}(p_{i+1})\geq 0$. Such a pair is guaranteed to exist, because by construction, it must be that for any history, $\alpha_{\tilde{p}^{1:t-1}}(0)\leq 0$ and $\alpha_{\tilde{p}^{1:t-1}}(1)\geq 0$. Note that a well known randomized algorithm obtaining diminishing ECE (and smooth calibration error) uses the same observation to carefully randomize between two such adjacent predictions (Foster, 1999; Foster and Hart, 2018). Upon observing the outcome $y^{t}$, the algorithm outputs prediction $\tilde{p}^{t}=\operatorname{\text{argmin}}_{p\in\{p_{i},p_{i+1}\}}|p-y^{t}|$. Naturally, we cannot implement this algorithm, as it chooses its prediction only after observing the outcome, but our analysis will rely on a key property this algorithm maintains—namely, that it always produces a sequence of predictions with ECE upper bounded by $m+1$, the number of elements in the discretized prediction space.

Algorithm 1 (Almost-One-Step-Ahead) maintains the same state $\alpha_{\tilde{p}^{1:t}}(p)$ as One-Step-Ahead (which it can compute at round $t$ after observing the outcome $y_{t-1}$). In particular, it does not keep track of the bias of its own predictions, but rather keeps track of the bias of the predictions that One-Step-Ahead would have made. Thus it can determine the pair $p_{i},p_{i+1}$ that One-Step-Ahead would commit to predict at round $t$. It cannot make the same prediction as One-Step-Ahead (as it must fix its prediction before the label is observed) — so instead it deterministically predicts $p^{t}=p_{i}$ (or $p^{t}=p_{i+1}$ — the choice can be arbitrary and does not affect the analysis). Since we have that $|p_{i}-p_{i+1}|\leq\frac{1}{m}$, it must be that for whichever choice One-Step-Ahead would have made, we have $|\tilde{p}^{t}-p^{t}|\leq\frac{1}{m}$. In other words, although Almost-One-Step-Ahead does not make the same predictions as One-Step-Ahead, it makes predictions that are within $\ell_{1}$ distance $T/m$ after $T$ rounds. The analysis then follows by the ECE bound of One-Step-Ahead, the triangle inequality, and choosing $m=\sqrt{T}$.