causal-inference

Minimal M5PHET causal provider

The CPU-only ATE provider lives in the isolated provider/ subproject; see installation, explicit demo preparation and contract. It is inference-only over studies that were fitted on purpose beforehand, it requires caller-declared identifying assumptions, and it never fits during chat inference.

It reports no accuracy, and refuses to: a causal estimate has no held-out truth, because the counterfactual outcome of a row is never observed. It reports the estimate, its interval, the assumptions it rests on and its diagnostics.

The application instructions below are the historical tool's, not the provider's installation path.

Status: experimental research repository — unverified. This repository hosts exploratory causal-inference research for regime-based trading strategies. Nothing here is a released tool: the scripts are one-off research artifacts, the inherited application code is unmaintained, and no claim in this README has been verified by running the code for this document beyond inspecting the committed tree.

What this repository actually is

Two distinct layers share this repository:

1. An inherited fork of rl-optimizer. The repository began as a copy of that project and most of the tree still comes from it: the app/ package (CLI, config handling, plugin loader, data handler), the plugin modules under app/plugins/, the legacy tests/ suite and the helper .bat/.sh scripts.

2. Causal research scripts added on top for regime-analysis experiments (see below).

Unrepaired inherited package identity — disclosure

The package metadata still identifies the inherited rl-optimizer package: setup.py declares name='rl-optimizer', a rl_optimizer=app.main:main console script, rl_optimizer.* entry-point groups and the upstream rl-optimizer project URL. This has not been repaired: installing this repository would install a package named rl-optimizer, not causal-inference. Several declared entry points also reference module paths that do not exist in the committed tree (for example app.plugins.optimizer_plugin_openrl — the file actually lives at app/plugins/transformation/optimizer_plugin_openrl.py, and the NEAT optimizer plugins are absent entirely). Code ownership and packaging identity will be separated in a future repair; until then treat the packaging metadata as stale.

Causal research content committed on master

• causal_regime_analysis.py — the main research script. It applies three techniques to 12 regime features derived from EURUSD hourly OHLC data (resampled to 4h):

  1. NOTEARS causal discovery (DAG structure among features),
  2. Invariant Causal Prediction (stability of feature/forward-return relationships across regimes),
  3. DoWhy refutation tests (placebo and random-confounder checks).

  It expects its input CSV from a sibling feature-eng checkout (path overridable via the OHLC_FILE environment variable). The input data is not committed here.

• results/causal_analysis_results.json — the recorded output of one historical run of that script (ICP scores, p-values and per-regime coefficients per feature). It documents a specific past run and is not regenerated automatically; it has not been independently reproduced for this README.

Additional causal research scripts (a v2 regime analysis, cluster-based regime analysis and cross-asset comparison/audit work) exist only as uncommitted work-in-progress in the owner's local working tree. They are not part of the repository history and are deliberately not described here; they will be documented if and when they are committed.

Inherited application code (unverified)

The app/ package provides a plugin-based CLI inherited from rl-optimizer: app/main.py, configuration merging (app/config_handler.py, app/config_merger.py), a plugin loader (app/plugin_loader.py) and plugin modules under app/plugins/ (inference, preprocessing and transformation groups, including an OpenRL PPO agent and prediction/custom environment plugins). The legacy tests/ suite also predates the causal work. None of this has been exercised or verified in the causal-inference context — no claim is made that the CLI runs, that the plugins load, or that the tests pass.

Dependencies

requirements.txt mixes inherited RL dependencies (tensorflow-gpu, openrl, neat-python, stable-baselines3, gym) with general scientific packages. The causal scripts additionally import libraries (e.g. for NOTEARS/DoWhy) that are not all pinned there. No tested installation procedure is claimed.

Relationship to sibling repositories

• rl-optimizer — the upstream project this repository was forked from; its packaging identity is still present here (see disclosure above).
• feature-eng — source of the regime features and the OHLC test data consumed by causal_regime_analysis.py.

License

LICENSE.txt (inherited from the upstream project).

---

This README describes only what is committed on master as of 2026-08-10. Experimental, unverified research — not a supported tool, not financial advice.