Temporal-Prior Monte Carlo Inverse Estimation and Attribution-Guided Adaptive Reweighting for Sequential Elimination Ranking Systems

Abstract

: Sequential elimination ranking systems decide outcomes by combining expert scores with crowd votes, yet the crowd-vote signal is rarely recorded and must be estimated from observed eliminations. This paper presents a Monte Carlo inverse framework that recovers the latent signal under temporal-smoothing priors and historical constraints, so that the estimate stays consistent across consecutive rounds rather than fitted round by round. Backtesting over 293 elimination rounds from 34 seasons reaches a consistency rate of 89.76 percent, with the estimated vote of eliminated candidates holding a standard deviation of only 0.025 and 72.3 percent of samples concentrated in a narrow band. A rank-difference analysis then compares two scoring rules: one shows a clear crowd bias, eliminating 63.2 percent of high-ranked candidates in contested rounds, while an expert-override mechanism corrects 52.7 percent of disputed eliminations. Gradient-boosted attribution with Shapley values shows that expert scores are driven mainly by a single dominant feature at 80 percent contribution and fall with a candidate covariate at a correlation of -0.43. A dynamic reweighting protocol raises the top-candidate retention rate to 89 percent, a gain of 27 points, and lifts public alignment to 78 percent. Sensitivity analysis confirms stability.