# Calibrated Forecasting

### Probability Forecasting

Can one predict the weather ignorantly yet accurately? To be concrete, suppose we are given the weather record of the past 100 days,

we cannot access any other information (e.g., temperature, satellite images, etc.), and we know nothing about meteorology. Can we tell how likely it will rain tomorrow?

This problem can be cast as a probability forecasting problem. The protocol of probability forecasting is as follows. For $t = 1, 2, \ldots$,

1. Predictor announces $p_t \in [ 0, 1 ]$.
2. Reality announces $y_t \in \{ 0, 1 \}$.

The goal of Predictor is to make accurate predictions, such that $p_t$ is a good estimate of the likelihood of the event $\{ y_t = 1 \}$. As the protocol is deterministic, we do not say “probability of the event $\{ y_t = 1 \}$”.

Thinking of ‘$0$’ as ‘dry’, and ‘$1$’ as ‘wet’, we recover the original weather forecasting problem.

### Calibration

What precisely are the requirements a good prediction strategy should satisfy? The requirements in general depend on the actual application scenario. In this post, we will only consider a minimal requirement—any good prediction strategy should be calibrated.

Roughly speaking, the notion of calibration requires the predictions to comply with the empirical frequencies, i.e., for any $\rho \in ( 0, 1 )$, it holds that

To make the notion precise, consider the following discretization argument. Fix some $\varepsilon \in ( 0, 1 )$. Let us decompose the interval $[ 0, 1 ]$ as

where $M := \lceil \varepsilon^{-1} \rceil$. We denote the center of $I_m$ by $c_m$. Define the empirical frequencies

if the divisor is non-zero, and $\rho_m (T) := c_m$ otherwise. We would like, for all $m$,

We do not really want to care about those $m$’s for which $\sum_{t = 1}^T \mathbb{I}_{ \{ p_t \in I_m \} } = o( T )$, as they are irrelevant in the long run. A compact formulation of the requirement is

Some simple algebraic manipulations lead to the following (slightly different) defnition.

Definition. A (possibly randomized) prediction strategy is $\varepsilon$-calibrated, if the resulting sequence of predictions satisfy

Surprisingly, even when one knows nothing about the mechanism generating the sequence $( y_t )_{t \in \mathbb{N}}$, calibration is achievable.

Theorem. Suppose that Reality, before announcing $y_t$, does not know the exact value of $p_t$. For any $\varepsilon > 0$, there exists an $\varepsilon$-calibrated probability forecasting strategy.

Some quick observations:

1. If $y_t$ can depend on the exact value of $p_t$, calibration is impossible. For example, Reality can always choose $y_t = 1$ whenever $p_t \leq 0.5$, and $y_t = 0$ otherwise.

2. The remark above also shows the necessity of a randomized forecasting strategy. If Forecaster uses a deterministic strategy, for which $p_t$ is a deterministic function of $p_1, y_1, p_2, y_2, \ldots, p_{t - 1}, y_{t - 1}$, then Reality can also compute exact value of $p_t$.

### An Achieving Strategy

A calibrated forecasting strategy can be derived by the idea of Blackwell approachability.

The strategy only chooses $p_t$ from the set $\{ c_m \}$; effectively, Forecaster only needs to choose $m_t \in \{ 1, 2, \ldots, M \}$ for all $t$. Define, for all $m$,

We already know that Forecaster must adopt a randomized strategy. Suppose that Forecaster chooses $m_t$ randomly according to some probability distribution $P_t$ on $\{ 1, 2, \ldots, M \}$. Define

Forecaster’s Strategy. For each $t$, Forecaster first computes some $P_t$ such that for any probability distribution $Q$ on $\{ 0, 1 \}$,

where $\mathrm{proj}$ denotes projection onto the $1$-norm ball of radius $\varepsilon$ in $\mathbb{R}^M$, and then announces $m_t$ randomly according to $P_t$.

Lemma. Such $P_t$ always exists.

Remark. In practice, one can compute $P_t$ by solving the saddle point problem:

Notice that the objective function is bilinear; there are a variety of existing algorithms that solve this problem.

The strategy looks mysterious at first glance. I will introduce the theory of Blackwell approachability in the next post.

1. Cesa-Bianchi, N. and Lugosi, G., 2006. Prediction, Learning, and Games.
2. Dawid, A.P., 1982. The well calibrated Bayesian. J. Am. Stat. Assoc.
3. Dawid, A. P., 1985. Comment: The impossibility of inductive inference. J. Am. Stat. Assoc.
4. Foster, D. P., 1999. A proof of calibration via Blackwell’s approachability theorem. Games Econ. Behav.
5. Mannor, S. and Stoltz, G., 2010. A geometric proof of calibration. Math. Oper. Res.
6. Oakes, D., 1985. Self-calibrating priors do not exist. J. Am. Stat. Assoc.