Replication package for "Utility-Weighted Calibration under Order-Book Frictions: What Recalibration Can and Cannot Buy"

Replication package for the paper "Utility-Weighted Calibration under Order-Book Frictions: What Recalibration Can and Cannot Buy" (manuscript IREF-D-26-00685R2, International Review of Economics and Finance). The paper asks when recalibrating a predictive density improves a constrained economic decision under measured order-book frictions, and identifies three quantities that govern how much room a recalibration layer has to act: the decision's responsiveness, the forecast's out-of-sample information, and the resolution of the recalibration mesh at the functional the decision actually reads. All three are measurable before any recalibration is fitted. Under proportional costs the cost-aware optimum lies in its inaction region in 97.6% of one-minute periods in S&P 500 E-mini order-book data, where recalibration cannot change the current trade; in a value-at-risk sizing problem, moving the recalibration knots to the levels the limit evaluates halves realised breach rates, with the size of the gain proportional to how badly the base density's tail is mis-covered. CONTENTS code/ — 50 analysis modules (Python 3.12; numpy, pandas, scipy, arch, matplotlib), including the shared library implementing the friction operator, both decision rules, HAC and moving-block bootstrap inference, and the paired first- and second-order stochastic dominance tests. data_derived/ — every result table and log referenced in the paper. figures/ — the manuscript figures as vector PDF. manuscript/ — the complete LaTeX sources for the paper and the Online Appendix. MANIFEST_sha256.csv — SHA-256 of every file in the archive. README.md — layout, run order, and the notes reproduced below. DATA AVAILABILITY The raw CME Globex MDP-3.0 messages underlying the high-frequency panels are licensed from Databento and are not redistributable, so the panel builders cannot be run without licensed access. Every table and figure in the paper can nonetheless be regenerated from the derived outputs included here. Instruments are the front-month S&P 500 E-mini (ES), the 10-year Treasury note (ZN) and EUR/USD (6E) futures over 1 September to 30 December 2025, with a five-year daily equity-index panel for the applications in which the decision is interior. REPRODUCIBILITY Randomness is confined to the block bootstrap, the parametric bootstrap null distributions, the randomised probability integral transform draws, and a single synthetic loss-surface exercise; all use fixed, reported seeds. No reported empirical quantity is simulated. r2_digest.py reassembles every number reported in the paper from the stored outputs. The package also carries the correction that motivated this revision. The previous implementation selected positions with a cost-blind mean-variance target followed by clipping, which is not the rule the manuscript specified. r2lib.decide_clip reproduces that rule exactly and r2lib.decide_cost_aware implements the specified one, with the cost inside the objective and a closed-form solution; r2_decisions.py runs both on identical inputs, which is the comparison reported in the paper.