regimebench: Boundary Pinning in Cluster-Count Selection -- A Controlled Benchmark

Abstract

A controlled benchmark for cluster-count selection on time series, using persistentheavy-tailed volatility processes as a deliberately hard case. It benchmarks Silhouette,Davies-Bouldin, Calinski-Harabasz, AIC, BIC, BIC(N_eff), ICL, the Gap Statistic andPrediction Strength against known ground truth, four null controls, two positive controls(one on the latent state variable, one running the identical criterion code onwell-separated i.i.d. Gaussian clusters), and a century of US daily returns. Every table and figure in the accompanying article regenerates from a single entry point(run_all.py), a seventeen-step pipeline whose final step is tools/verify_manuscript.py: itparses each reported number out of the LaTeX source and asserts it against the artifactthat produced it, failing on drift. A --self-test flag confirms the verifier detectsinjected corruption. Data: synthetic series are regenerated bit-for-bit from a seed recorded inexperiments/config/. Raw third-party payloads (Ken French Data Library, FRED, Binance) arenot redistributed here; the connectors fetch them from source on first run anddata/provenance.json records each payload's source URL, retrieval timestamp, row count,date range and SHA-256, so a reader verifies byte-for-byte.