Operational Weighting Changes Cooperative Forecast Valuation: A Residual-Fixed Sensitivity Analysis for Federated Demand Forecasting
Authors/Creators
Description
Predictive-score-based valuation often allocates value from residual or accuracy improvements while leaving the map to downstream operational value implicit. We study federated
demand forecasting among firms as a transferable-utility cooperative game whose characteristic function is an inventory-motivated weighted reduction in predictive residual scale.
Weights combine the normal-loss expected-cost coefficient in holding and shortage costs
with a √
Li + 1 protection-period scaling factor under an i.i.d. one-step error benchmark;
they supply a relative inventory-motivated weighting rule rather than a calibrated estimator of realized inventory cost, and are not claimed to coincide with realized path inventory
cost under a finite-horizon single-echelon diagnostic. Coalition models use exact federated
ridge on additive training statistics Ti = (Ai
, bi
, ni) with size-scaled regularization (including the intercept). The design caches residual maps and changes only operational weights,
so that, conditional on each seed’s fixed residual map, ranking and normalized-allocation
shifts are induced by weighting rather than refitting. Along a mean-normalized path from
equal weights to the largest tested effective weight imbalance (κeff ≈ 2.94), Spearman agreement with the unweighted residual ranking falls from 1 to 0.89–0.77 for n ∈ {4, 6, 8} (40
seeds), with full-rank-change rates up to 1.00 and exhaustive assignment dispersion (4! = 24,
6! = 720, 8! = 40320) showing that, within the tested DGP and weight multiset, fixed κeff
does not uniquely determine rank agreement. In an n=4 finite-horizon path-cost diagnostic
using average cost per hold-out period, residual-weighted and path-cost Shapley rankings are
negatively correlated (bootstrap CIs exclude zero). We report analytically computed first
rank-change thresholds along the mean-normalized path, Wilson intervals, and robustness
under fixed versus size-scaled ridge and moderate sample-size heterogeneity.
Files
zk_shapley_paper.pdf
Additional details
Related works
- Is supplement to
- Software: 10.5281/zenodo.23457791 (DOI)