Researchers publish diffusion copulas paper

StatsPapers on May 20 pointed readers to two new machine-learning papers that tackle a common problem from different directions: how to represent uncertainty without washing out the events researchers care about most. One paper, posted to arXiv on May 19 by David Huk, Dongshan Wang and Miha Bresar, proposes “Probabilistic Multivariate Time Series Forecasting with Diffusion Copulas.” A second paper, posted the same day by Olga Zaghen, Maksim Zhdanov, Dario Coscia, David R. Wessels and Erik J. Bekkers, describes “Uncertainty-aware Machine Learning Interatomic Potentials via Learned Functional Perturbations.” Both papers focus on probabilistic prediction rather than point estimates alone. The forecasting paper is framed around financial time series and tail risk in cryptocurrency markets, while the interatomic-potentials paper is aimed at atomistic simulation, where a model can be accurate on average but fail on out-of-distribution inputs. ### Why are these two papers being mentioned together? The May 20 StatsPapers post grouped them as examples of recent machine-learning work on uncertainty, forecasting and probabilistic modeling, according to the social briefing provided for this story. (arxiv.org) The pairing is loose in subject matter but tight in method: both papers argue that standard approaches can produce confident-looking outputs while missing the structure of uncertainty that matters in practice. David Huk and his co-authors say end-to-end diffusion forecasters can develop a “normality bias,” preserving joint coherence at the expense of marginal calibration and underestimating tail risk. Olga Zaghen and co-authors make a parallel complaint in atomistic modeling, writing that machine-learning interatomic potentials can suffer “silent failures” on out-of-distribution configurations unless uncertainty is modeled explicitly. ### What does the diffusion-copulas paper actually propose? (arxiv.org) The May 19 arXiv paper by Huk, Wang and Bresar says it “explicitly decouples” two tasks: learning each variable’s marginal distribution and learning the dependence structure across variables. The authors use deep mixture density networks for heavy-tailed asset dynamics and then a “Classification-Diffusion Copula” for the joint dependence structure. The application in the paper is cryptocurrency markets. (arxiv.org) The authors say their framework outperforms state-of-the-art baselines in forecasting “systemic extremes of both marginal and joint events,” and they single out crash scenarios as the main stress test. ### What is the paper’s central claim about crashes and tail risk? The most pointed claim in the abstract concerns simultaneous market selloffs. Huk, Wang and Bresar write that baseline models can classify simultaneous crashes as statistically impossible “Black Swans” with high surprise, while their framework identifies them as “Expected Crashes” with low surprise. (arxiv.org) That distinction matters because the paper is not only trying to improve average forecast quality. (arxiv.org) The authors say the goal is to preserve correlation structure during contagion events, which is the setting in which risk models are often tested hardest. ### What is new in the interatomic-potentials paper? Olga Zaghen and her co-authors propose turning a deterministic machine-learning interatomic potential into a probabilistic one through “learned functional perturbations,” then fine-tuning it end-to-end with the Continuous Ranked Probability Score, or CRPS. (arxiv.org) The paper says that approach is simpler than non-ensemble uncertainty-quantification methods that rely on variational inference or parametric distributional assumptions. The authors validate the method on two settings. On an N-body charged-particle benchmark, they report that P-EGNN improves CRPS over the Bayesian MLIP method BLIP by 19% to 32% across training sizes. On silica, they say P-Orb improves the Spearman correlation between predicted uncertainty and actual error to 0.84 from 0.75 for BLIP-Orb. ### Where do these papers sit in the publication pipeline? The diffusion-copulas paper was submitted to arXiv on May 19 and is also listed on OpenReview as an ICLR 2026 Workshop on Advances in Financial AI paper dated Feb. 28, 2026. (arxiv.org) The interatomic-potentials paper was submitted to arXiv on May 19. Orbital Materials’ public model repository shows Orb-v3 materials remain available on GitHub, including documentation on the model family referenced in the interatomic-potentials paper. (arxiv.org) The next step for readers is straightforward: the preprints are on arXiv, and the associated model and code materials are linked through public repositories and author accounts. (github.com) (arxiv.org)