Reproducibility Package: Basel-Aware Regime-Weighted Stacking (BARWS) for Multi-Asset Value-at-Risk Forecasting

This package contains the R source code and result artifacts for the Basel-Aware Regime-Weighted Stacking (BARWS) methodology from the author's Doctor of Computer Science dissertation at Bina Nusantara University. BARWS extends ensemble Value-at-Risk forecasting via (1) augmenting pinball loss with a Basel III Traffic Light surcharge, smooth Kupiec coverage penalty, and Christoffersen clustering penalty for joint accuracy + regulatory compliance optimization; and (2) regime-conditional weights (Normal/Volatile/Crisis) via EWMA-then-k-means clustering with optional soft mixture-of-experts. The stacking ensemble combines seven base learners — HS, VCV, MCS, LDA-FFT, LDA-Panjer, XGBoost (conformal calibration), and LSTM quantile regressor — over a five-asset portfolio (USDIDR, EURUSD, USDJPY, XAUUSD, BTCUSD) across three regulatory periods: A (2014-09-17 → 2019-12-31, pre-COVID), B (2020-01-01 → 2021-12-31, COVID shock), and C (2022-01-01 → 2026-03-10, post-COVID + crypto winter). The bundle includes: (a) twelve pedagogical R scripts (LAMPIRAN A–L) matching the dissertation appendix structure with full mathematical exposition; (b) three runnable master scripts — dcs-bky-std-v28.R (Phase-1 batch, 12 configurations = 3 periods × 4 loss modes), dcs-bky-std-v28b.R (canonical for the accompanying BARWS paper: reproduces +1.98 bps RCE and ΔL_pin = −2.95 [95% CI −6.30, −0.18] for Period A ENS_ALL), and dcs-bky-std-v28d.R (symmetric-Kupiec variant; empirically identical winners to v28b across all five feasible cells, included as validation companion); (c) full Phase-2 lambda-sweep Pareto-frontier output (246 child runs, 9×9 grid × 3 periods); (d) automation script (populate-results.R) mapping raw output to dissertation tables/figures; (e) representative single-run output from the Period A winning configuration (λ_basel=0.25, λ_kup=2, ENS_ALL; +1.98 bps RCE, Basel YELLOW 3.65, Kupiec p=0.4466); and (f) complete manifest of all Chapter IV artifacts (16 entries: 8 tables + 8 figures). WHAT'S NEW IN VERSION 2 (2026-05-20): Added master scripts v28b (paper canonical) and v28d (validation companion). Populated results/tables/ and results/figures/ with all eight final tables (4.1–4.8) and figures (4.1–4.8). Replaced data/sample_output/ with the Period A winning configuration. Extended manifest from 5 to 16 entries. Raw price data is fetched live from Yahoo Finance and Binance API on each run and is not redistributed (subject to provider ToS). All R code, documentation, and derived outputs released under CC BY 4.0. Reproduction time: ~5 h batch, ~14.5 h Phase-2 sweep (246 runs).