Non-centred stochastic volatility produced stronger density-forecast scores than the centred version in a forecasting software study, with the advantage rising from about 6% at a one-quarter horizon to 15% at a one-year horizon. On point forecasts, however, the difference between the two volatility parameterizations was below 1%. In this exercise, the reported gap appeared in density scoring rather than point accuracy. The result is a model comparison, so it does not establish that the parameterization caused the higher scores.
The paper presents three R packages for forecasting Bayesian vector autoregressions, or VARs, with stochastic volatility and Student-t errors. It compares non-centred and centred versions of stochastic volatility, alongside constant-volatility models, normal and Student-t errors, and structural and reduced-form specifications. The central question is whether that package approach can improve density and point prediction.
The forecasting exercise
The empirical test used an expanding-window recursive forecasting exercise on a 10-variable quarterly system. The data ran from Q1 1959 through Q3 2025, and the initial forecast origin was Q4 2020. Across the exercise, 19 model specifications were compared, spanning non-centred, centred and constant-volatility models, together with normal and Student-t error assumptions. Structural and reduced-form versions were also included.
Forecasts were evaluated at one quarter and one year ahead. For density forecasts, the main measure was the average exponentiated predictive log-score, or PLS. For point forecasts, it was mean absolute scaled error, or MASE. The first score assesses the reported predictive distribution, while the second captures error in the reported point forecast, so the measures focus on different parts of forecasting performance.
Where the scores diverged
Non-centred stochastic volatility increased the average exponentiated PLS by approximately 6% at the one-quarter horizon and 15% at the four-quarter horizon, relative to centred stochastic volatility. The supplied analysis gives no confidence intervals, standard errors or formal significance tests for the comparison, so it does not show how much uncertainty surrounds those percentages.
Point accuracy told a different story. Average MASE was very similar across the alternative stochastic-volatility parameterizations, with relative differences below 1%. The reported density-score advantage was not matched by a substantial gain in the accuracy of the single-value forecasts in this exercise.
The error-distribution and structural choices produced another directional pattern. At both forecast horizons, Student-t models outperformed normal-error models, while structural models outperformed reduced-form models on both PLS and MASE. The analysis does not provide aggregate effect sizes for those comparisons, and no uncertainty estimate or significance test is reported.
The software's practical case
The software design combines C++ computational efficiency with the convenience of working with data in R. The packages emphasize normalised models with a well-specified covariance for the predictive density. Their unified forecasting workflow produces predictive-density draws together with their means and covariances.
The speed claims come from illustrative implementation benchmarks. In one test, RcppArmadillo obtained 1,000 standard-normal draws 33% faster on average than the Rcpp implementation. A C++ row-by-row Gibbs sampler for an 8-by-8 structural matrix was more than 60 times faster than equivalent optimized R code. The reported stochastic-volatility implementation delivered more than tenfold computational gains relative to less efficient implementations.
The supplied analysis does not report benchmark variability, hardware details or execution conditions. The figures should therefore be read as results from the reported implementations and tests, rather than as a universal ranking of C++ and R.
How far the finding reaches
The forecasting evidence rests on one 10-variable quarterly system and one expanding-window exercise. It does not establish that non-centred stochastic volatility will improve density forecasts on other data sets or at other horizons, or that the reported score differences would survive formal predictive-accuracy tests. Nor does the below-1% MASE gap support a claim of broad point-forecast superiority.
The document identifies itself as arXiv version 1 dated August 2026, and its external peer-review status is not reported in the supplied material. A note accompanying the reported results says the data and results can be reproduced by following the code. That provides a route for checking the reported comparisons, while broader validation remains open.
Paper data and sources
Original title: A Design Concept of Forecasting Software for Normalized Vector Autoregressions with Fat Tails and Stochastic Volatility
Authors: Fei Shang, Xiaolei Wang, Tomasz Woźniak
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text