Eq. (9), p. 35 — expected event wins differ by m(R(y_i,θ) − R(y_j,θ))/(n−1)
ProvedForecastComp.IELF.eq_9forecasting-competitionlimit-accuracyp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let , let the outcomes of events be independent with , and let be the number of event lotteries of I-ELF won by forecaster . For any reports and any two forecasters ,
and consequently, whenever ,
The more accurate forecaster wins more event lotteries in expectation, by an amount growing linearly in .
Formalization Note The identity holds for every real-valued and all real reports (it only uses that each ELF lottery sums to one); no bound on is assumed.
Preamble
import Mathlib import Definitions.Def_ForecastComp_IELF_Model
Formal statement
namespace ForecastComp.IELF
/-- Eq. (9), p. 35: under independent outcomes with marginals `θ`, the expected numbers of event
wins of forecasters `i` and `j` differ by `m (R(y_i,θ) − R(y_j,θ))/(n−1)`; in particular the
difference is at least `mΔ/(n−1)` when `R(y_i,θ) − R(y_j,θ) ≥ Δ`. -/
theorem eq_9 {m n : ℕ} (hn : 2 ≤ n) (R : ℝ → Bool → ℝ)
(θ : Fin m → ℝ) (hθ : ∀ k, θ k ∈ Set.Icc (0:ℝ) 1)
(y : Fin n → Fin m → ℝ) (i j : Fin n) (hij : i ≠ j) :
expWins R θ y i - expWins R θ y j =
(m:ℝ) * (avgScore R θ (y i) - avgScore R θ (y j)) / ((n:ℝ) - 1) ∧
∀ Δ : ℝ, Δ ≤ avgScore R θ (y i) - avgScore R θ (y j) →
(m:ℝ) * Δ / ((n:ℝ) - 1) ≤ expWins R θ y i - expWins R θ y j := by sorry
end ForecastComp.IELF
Source
Witkowski et al., arXiv:2101.01816v2, p. 35, eq. (9), proof of Theorem 7 (Appendix M)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.