ML predictive distributions thread
A technical thread proposed treating predictive distributions as points on a statistical manifold and regressing them on macro drivers like rates and VIX to generate trading signals when deviations occur. — the idea connects distributional forecasting with macro‑driven signal extraction for quant trading. (x.com)
Quant Beckman, who publishes the Substack "Trading the Breaking" and bills the project as a quant research lab, pushed the X thread proposing the manifold approach; the author also maintains a GitHub profile with code and notebooks. (quantbeckman.com) Single‑index Fréchet regression (Annals of Statistics) provides a formal framework for regressing metric‑space‑valued responses—explicitly including univariate probability distributions—on Euclidean predictors, enabling dimension reduction for distributional responses. (projecteuclid.org) Projected statistical methods built on the 2‑Wasserstein metric have been implemented for PCA and regression of distributions on the real line, giving a practical pipeline for mapping distributions to tangent/coordinate spaces before linear modeling. (jmlr.org) Recent distributional‑time‑series work embeds forecasts into Wasserstein geometry via a Koopman/DPDD extension with spectral convergence guarantees (Springer/ArXiv 2025), which directly addresses how probability measures evolve over time for forecasting. (link.springer.com) Estimation of instantaneous distributional dynamics has been demonstrated using Wasserstein temporal gradients and local Fréchet regression, providing consistent estimators for one‑dimensional distribution evolution that would be used when comparing a model’s predictive distribution to macro‑driven expectations. (projecteuclid.org) Finance applications include a 2025 BayesNP‑VAR study that generates joint predictive distributions for market return variables and reports robust out‑of‑sample forecasts at longer horizons, showing precedent for distributional forecasting in macro‑driven return prediction. (sciencedirect.com) Methodological extensions for mapping distributions to distributions using the Wasserstein metric—explicitly including Gaussian‑to‑Gaussian regression in multivariate settings—have been published recently, supplying candidate estimators for the “regress predictive distribution on rates/VIX” step. (sciencedirect.com) Empirically relevant macro drivers are supported by recent ML work: a 2024 Liverpool thesis documents improved daily VIX forecasts with adaptive ML methods, and a 2026 preprint (U.S. data through 2025) finds contrarian predictability following extreme VIX spikes—both motivating VIX as a regressor in distributional models. (livrepository.liverpool.ac.uk) Open‑source tooling commonly used to implement these ideas includes Python Optimal Transport (POT), a GitHub project with roughly 2.8k stars providing fast 1‑D OT and barycenter solvers, and geomstats (geomstats v2.8.0 on PyPI) for Riemannian/manifold computations; Quant Beckman’s own posts reference metric projections in portfolio constraints, aligning the thread’s prescriptions with existing practice. (github.com)
