Causal inference for the real world — one observation at a time.
Every major causal inference library (EconML, CausalML, DoWhy) requires a complete dataset before you begin. But many real-world applications don't have that luxury:
- A/B tests where you want decisions now, not after 30 days
- Clinical trials where treatment effects must be monitored continuously
- Marketing systems where customer data arrives as a stream
- Any setting where the treatment effect might shift over time
OnlineCML processes one observation at a time. No batches. No waiting.
pip install onlinecmlfromonlinecml.datasetsimportLinearCausalStreamfromonlinecml.reweightingimportOnlineIPWestimator=OnlineIPW()
forx, treatment, outcome, _inLinearCausalStream(n=1000, true_ate=2.0, seed=42):
estimator.learn_one(x, treatment, outcome)
print(f"ATE: {estimator.predict_ate():.3f}") # → ~2.0print(f"95%CI: {estimator.predict_ci()}")| Method | Class | ATE | Individual CATE | Doubly Robust |
|---|---|---|---|---|
| Inverse Probability Weighting | OnlineIPW | ✓ | — | — |
| Augmented IPW | OnlineAIPW | ✓ | ✓ | ✓ |
| Overlap Weights | OnlineOverlapWeights | ✓ | — | — |
| S-Learner | OnlineSLearner | ✓ | ✓ | — |
| T-Learner | OnlineTLearner | ✓ | ✓ | — |
| X-Learner | OnlineXLearner | ✓ | ✓ | — |
| R-Learner | OnlineRLearner | ✓ | ✓ | — |
| Online Matching | OnlineMatching | ✓ | ✓ | — |
| Caliper Matching | OnlineCaliperMatching | ✓ | ✓ | — |
| Causal Hoeffding Tree | CausalHoeffdingTree | ✓ | ✓ | ✓ |
| Online Causal Forest | OnlineCausalForest | ✓ | ✓ | ✓ |
Novel contributions:CausalHoeffdingTree and OnlineCausalForest implement a
custom causal split criterion that maximises between-child CATE variance rather than
outcome MSE, with linear leaf models, doubly robust correction, multi-threshold split
search, and per-tree ADWIN drift detection.
Policies:EpsilonGreedy, ThompsonSampling, UCB
Diagnostics:OnlineSMD, ATETracker (with convergence plot and forgetting factor),
OverlapChecker, ConceptDriftMonitor
Datasets:LinearCausalStream, HeterogeneousCausalStream, DriftingCausalStream,
UnbalancedCausalStream, ContinuousTreatmentStream
Evaluation:progressive_causal_score, PEHE, ATEError, UpliftAUC, QiniCoefficient
| DoWhy | EconML | CausalML | OnlineCML | |
|---|---|---|---|---|
| Online / streaming | ✗ | ✗ | ✗ | ✓ |
| One-obs-at-a-time | ✗ | ✗ | ✗ | ✓ |
| Concept drift | ✗ | ✗ | ✗ | ✓ |
| Exploration policy | ✗ | ✗ | ✗ | ✓ |
| River compatible | ✗ | ✗ | ✗ | ✓ |
| Online causal forest | ✗ | ✗ | ✗ | ✓ |
| IPW / DR / Overlap | ✓ | ✓ | ✓ | ✓ |
| Meta-learners | ✓ | ✓ | ✓ | ✓ |
| CATE estimation | ✓ | ✓ | ✓ | ✓ |
Full documentation and example notebooks at athammad.github.io/onlinecml.
See CONTRIBUTING.md. All PRs require unit tests and must maintain >90% coverage.
@software{onlinecml2025,
title = {OnlineCML: Online Causal Machine Learning in Python},
author = {Hammad, Ahmed},
year = {2025},
url = {https://github.com/athammad/onlinecml}
}MIT