OnlineCML
Causal inference for the real world — one observation at a time.
OnlineCML is the first Python library for causal inference in a purely online, one-observation-at-a-time setting. It is the streaming counterpart to EconML, CausalML, and DoWhy — built on River's infrastructure and conventions.
Why OnlineCML?
Most causal inference libraries assume you have a complete dataset before you start. But the real world doesn't wait for your batch job: treatment decisions are made continuously, data arrives as a stream, and effects shift over time.
OnlineCML lets you:
- Estimate treatment effects without storing data (O(1) memory)
- Make treatment assignment decisions in real time
- Detect concept drift in causal relationships
- Stay compatible with River pipelines and any River model
Quick Install
5-Line Quickstart
from onlinecml.datasets import LinearCausalStream
from onlinecml.reweighting import OnlineIPW
estimator = OnlineIPW()
for x, treatment, outcome, _ in LinearCausalStream(n=1000, true_ate=2.0, seed=42):
estimator.learn_one(x, treatment, outcome)
print(f"ATE: {estimator.predict_ate():.3f}")
print(f"95% CI: {estimator.predict_ci()}")
Methods (v1.0)
| Method | Class | CATE | Category |
|---|---|---|---|
| Inverse Probability Weighting | OnlineIPW |
No | Reweighting |
| Augmented IPW (Doubly Robust) | OnlineAIPW |
Yes | Reweighting |
| Overlap Weights | OnlineOverlapWeights |
No | Reweighting |
| Covariate Balancing PS | OnlineCBPS |
No | Reweighting |
| S-Learner | OnlineSLearner |
Yes | Meta-Learner |
| T-Learner | OnlineTLearner |
Yes | Meta-Learner |
| X-Learner | OnlineXLearner |
Yes | Meta-Learner |
| R-Learner (Double ML) | OnlineRLearner |
Yes | Meta-Learner |
| Causal Hoeffding Tree | CausalHoeffdingTree |
Yes | Forest |
| Online Causal Forest | OnlineCausalForest |
Yes | Forest |
| Nearest-Neighbour Matching | OnlineMatching |
Yes | Matching |
| Caliper Matching | OnlineCaliperMatching |
Yes | Matching |
| Kernel Matching | OnlineKernelMatching |
Yes | Matching |
| Epsilon-Greedy | EpsilonGreedy |
— | Policy |
| Thompson Sampling | ThompsonSampling |
— | Policy |
| Gaussian Thompson Sampling | GaussianThompsonSampling |
— | Policy |
| UCB | UCB |
— | Policy |