Causal graphs for time series without the hand-waving

What it does

Tigramite estimates causal graphs from high-dimensional time series datasets. It pairs discovery algorithms (PCMCI, PCMCIplus, LPCMCI, and others) with conditional independence tests tailored to your data type—linear or nonlinear, continuous, discrete, or mixed. Once you have a graph, the CausalEffects class estimates direct, total, and mediated effects, and there’s a prediction module that uses sklearn models with causal feature selection.

The interesting bit

Most causal inference tools bury their assumptions in footnotes. Tigramite puts them front and center: each method and each test statistic comes with a clear table of prerequisites—causal stationarity, faithfulness, hidden variable tolerance, and so on. The README is essentially a decision tree for responsible modeling.

Key highlights

• Five discovery methods with different robustness to hidden confounders, contemporaneous links, and regime changes
• Nine conditional independence tests ranging from simple partial correlation to permutation-based mutual information and Gaussian process regression
• Handles missing values, masks, and heteroskedasticity
• PairwiseMultCI wrapper turns any univariate test multivariate
• Academic pedigree: methods published in Science Advances, NeurIPS, UAI, Nature Reviews Earth & Environment

Caveats

• Optional dependencies stack up quickly: torch, gpytorch, dcor, ortools depending on which methods you use
• Numba/numpy version compatibility is currently a known friction point
• The user agreement requires citing specific papers per method used, which is fair but administratively fussy

Verdict

Researchers in climate science, neuroscience, or econometrics who need defensible causal claims from observational time series should look here. If your data isn’t temporal or you just want quick correlations, this is overkill.