Reproduction archive: GAMLSS density forecasts and tail-risk backtests for agricultural futures

Code, derived data and results behind a study of GAMLSS (generalised additive models for location, scale and shape) density forecasts for the tail risk of eight agricultural futures (soybeans, corn, wheat, soybean meal, soybean oil, coffee, sugar and cotton), 2005 to 2025. The study asks whether the density chosen by an information criterion on the estimation sample is the one that passes a Value-at-Risk backtest out of sample, and compares GAMLSS models with GARCH-type benchmarks in a walk-forward evaluation from 2019. It was first submitted under the title "In-Sample Model Selection and Out-of-Sample Tail Risk: GAMLSS Density Forecasts for Agricultural Commodity Returns" and has been rewritten as a model-risk paper. The concept DOI always resolves to the latest version. Version 3.0.0 adds three analyses: GARCH and GJR-GARCH with skewed Student-t and Johnson SU innovations; the breach rate by day since the last refit, with a bootstrap over refit windows; and a full re-run of every model on returns that are not winsorised. It also adds daily-updated versions of the GARCH models, because the original GARCH forecasts were made 20 days ahead from a single fit while the GAMLSS models used the observed lagged return every day. The audit record is audit/NEW_ANALYSES_2026-09-29.md. Version 3.0.0 also corrects the Model Confidence Set table (results/table05_mcs.csv). The table of versions 1 and 2, in which every model stays in the 90% set in every cell, is not reproduced by R/06_backtests.R; the regenerated table excludes HAR-RV in all 24 market-level cells. It adds R/17_penalty_crossover.R, which writes the penalty crossover table that no script produced before, and corrects the README, whose run order and data description were inaccurate in earlier versions. The raw daily prices are NOT included. They were retrieved from Yahoo Finance with the R package quantmod, whose terms of service permit retrieval but not redistribution. R/01_load_data.R downloads them again and R/14_build_unwinsorised.R rebuilds the unwinsorised returns. The derived log-returns, winsorised and unwinsorised, are included. The R scripts, the walk-forward checkpoints and every table as CSV are included; README.md in the archive gives the run order. The three scripts new in this version were run with R 4.4.1, rugarch 1.5.3, gamlss 5.4.22 and MSGARCH 2.51 (not the renv versions in renv.lock); a validation run shows that they reproduce the earlier soybean forecasts exactly (MSGARCH excluded from that check). Code is released under the MIT licence and derived results under CC-BY-4.0.