Order-sensitivity and equivariance of scoring functions (Q2414489)

According to the authors, the general decision-theoretical framework consists of the following four elements: \par $\bullet$ observed realization $\{y_1,\ldots,y_n\}$, $n\geqslant 1$, of a time series $\{Y_t\}_{t\in\mathbb{N}}$, taking values in an observation domain $O$ with a $\sigma$-algebra $\mathcal{O}$, \par $\bullet$ a family $\mathcal{F}$ of probability distributions on $(O,\mathcal{O})$, containing distributions of $Y_t$, \par $\bullet$ forecasts $\{x_1^{(i)},\ldots, x_n^{(i)}\}$, $i\in\{1,\ldots,m\}$, competing forecasters taking values in an action domain $A\subset\mathbb{R}^k$, $k\geqslant 1$, \par $\bullet$ a scoring function $S:A\times O\rightarrow\mathbb{R}$. \par The scoring function is assumed to have the negative meaning. If forecaster reports the quantity $x\in A$ and $y\in O$ appears, is assigned penalty $S(x,y)$. Different forecasts ($i\in\{1,\ldots,m\}$) are compared by values of \textit{realized scores} \[ \overline{S}_n^{(i)}=\sum\limits_{i=1}^{n}S(x_t^{(i)},y_t). \] A forecast is supposed to be better if it's \textit{realized score} $\overline{S}_n$ is lower. \par To give a guidance on the choice of a suitable scoring function, the authors of the paper consider various properties of the scoring functions and introduce new quality criteria for such functions.