Causal time-series methods advance

Two short papers arrived on arXiv in late March that sharpened how economists and quants separate prediction from intervention in time series. (arxiv.org 1) (arxiv.org 2) The first, by Jacob Carlson and Neil Shephard, asks a simple question with messy consequences: when does a time-series forecast actually reflect a causal effect you would get if you intervened? (arxiv.org) The paper builds a “potential system,” a nonparametric analogue of the potential-outcomes idea used in cross-sectional causal work, but designed for sequences: it treats an “assignment” at time t and tracks how changing that assignment would move an outcome at t + h. (arxiv.org) That framing makes concrete the difference between saying “past X predicts future Y” and saying “forcing X to a new value would change Y.” A machine-learning model trained to minimize one-step-ahead error can exploit correlations that would vanish under intervention; the potential system gives conditions under which familiar tools—time-series regressions, local projections, impulse response calculations, even SVARs—carry causal content rather than mere association. (arxiv.org) For a trader, the distinction is practical. If an algorithm finds that macro surprise S predicts a short-window equity return, that does not imply trading on S will change returns; the predictive edge could come from a confounder or a feedback loop that breaks when you act. Carlson and Shephard’s formalism tells you what extra assumptions or experimental designs you need before treating a forecast as an actionable causal statement. (arxiv.org) The second paper, by Roberto Fuentes-Martínez and Irene Crimaldi, moves in the opposite direction: it gives a sharper test for a particular kind of causal influence in time series—Granger causality—when we care about the distribution tails. (arxiv.org) They replace the usual mean-focused Granger test with one based on expectiles, a risk-aware target that weights tail outcomes and, unlike simple quantiles, is both coherent and amenable to forecast evaluation. (arxiv.org) To model the joint dynamics they use M‑vine copulas—flexible constructions that let you build a high-dimensional joint distribution from many pairwise building blocks without assuming linear Gaussian structure. That lets the test detect non-linear and non-Gaussian dependencies across multiple series simultaneously, not just pairwise links. The authors show the test is consistent in theory, performs well in simulations, and finds joint causal links among international stock indices in their empirical example. (arxiv.org) For background on vine-copula approaches to Granger causality you can also look at prior work that used vine copulas to capture non-linear dependence in time series. (arxiv.org) Together these notes give two complementary tools. Carlson and Shephard give a conceptual scaffold that tells you when a time-series estimator is plausibly causal; Fuentes‑Martínez and Crimaldi give a practical test that looks for tail-sensitive predictive content under complex dependence. The combination matters for portfolio projects: you can use the potential-system ideas to design an experiment or identification strategy, then apply an expectile-based M‑vine test to see whether a proposed driver really carries tail risk information beyond the rest of the system. (arxiv.org 1) (arxiv.org 2) If you want a hands-on capstone: implement a local-projection estimator under the potential-system assumptions and compare its impulse responses to naive forecasts; then fit an M‑vine copula to regional equity returns and run the expectile Granger test to identify joint tail-drivers. Both arXiv PDFs are from March 2026 and are available for download. (arxiv.org) (arxiv.org)