Dataset Package for Robust Portfolio Optimization under Computational Complexity: A P-vs-NP-Inspired Markowitz-CAPM Framework with Cardinality Constraints and Black-Scholes Derivative-Pricing Overlay

This dataset package supports the study “Robust Portfolio Optimization under Computational Complexity: A P-vs-NP-Inspired Markowitz-CAPM Framework with Cardinality Constraints and a Black-Scholes Derivative-Pricing Overlay.” The package provides a structured replication dataset for a computational-finance analysis of cardinality-constrained portfolio optimization using a fixed Damodaran January 2026 U.S. industry universe. The empirical universe consists of 94 U.S. industry portfolios after excluding aggregate market rows. The dataset records and organizes the numerical evidence used in the manuscript, including firm-count information, levered beta, annual equity volatility, CAPM-implied expected returns, CAPM Sharpe proxies, covariance and correlation diagnostics, eigenvalue-concentration results, combinatorial support-space calculations, sparse portfolio optimization outputs, K-grid frontier evidence, reduced exact-enumeration benchmarks, greedy/Monte Carlo/genetic-algorithm comparisons, GA plus continuous reoptimization results, runtime and selection-stability diagnostics, covariance-model robustness checks, expected-return sensitivity tests, transaction-cost scenarios, beta-drift diagnostics, risk-contribution metrics, Black-Scholes option-overlay calculations, stress-scenario results, falsification gates, integrated ranking tables and final claim-boundary matrices. The baseline calibration uses a risk-free rate of 3.97% and an equity risk premium of 4.23%, generating CAPM-implied expected returns according to μᵢ = Rf + βᵢ × ERP. The main sparse-selection problem is defined for n = 94 and K = 10, implying C(94,10) = 9,041,256,841,903 possible portfolio supports before continuous weight optimization. Because full enumeration of the main universe is computationally infeasible, the package distinguishes between documented full-universe heuristic search coverage and exact reduced benchmarking. The reduced benchmark uses n = 20 and K = 6, where all 38,760 supports are enumerated. The package includes a master Excel workbook, long-format table exports, table index, figure index, extracted numerical registry, equation register, data dictionary, README files, metadata files, checksum file, and documentation document. Empty tables and non-data templates are excluded. The dataset is intended to support transparency, auditability, replication, and Zenodo-based archival of the manuscript’s empirical-computational results. It should be interpreted as a replication package for a model-implied computational finance study, not as a realized historical trading-performance dataset, and not as a proof of the formal P versus NP problem.