argenta-cml 0.1.0

Warehouse-native causal ML experimentation platform

Warehouse-native causal ML for experimentation.

Most A/B testing tools answer one question: Did the experiment work?

Argenta answers: For whom did it work, why, and what should you do next?

---

What Argenta Is

Argenta is an analytics layer that connects to your existing data warehouse, runs a SQL pipeline inside it to construct an experiment dataset, and then applies causal ML methods to produce results your current tool doesn't: heterogeneous treatment effects (HTE), per-user CATE scores, automatic subgroup discovery, and targeted rollout recommendations.

It works with experiments you've already run — using whatever tool assigned variants (Statsig, LaunchDarkly, Optimizely, homegrown feature flags). Argenta only analyzes.

What Argenta Is NOT

• Not an A/B testing platform. Argenta does not assign variants or run feature flags.
• Not a real-time system. Analysis runs post-hoc, after experiments have collected data.
• Not a data pipeline tool. Argenta does not move or replicate your data. SQL runs inside your warehouse; results are written back to a schema you control.

---

Quick Start

1. Install

# Pick your warehouse:
pip install "argenta-cml[snowflake]"
pip install "argenta-cml[bigquery]"
pip install "argenta-cml[redshift]"
pip install "argenta-cml[all]"   # all warehouses

2. Create a config file

# argenta.yaml
warehouse:
  warehouse_type: snowflake
  output_schema: argenta
  credentials:
    account: my_account
    user: argenta_svc
    password: "${SNOWFLAKE_PASSWORD}"
    database: ANALYTICS
    schema: PUBLIC
    warehouse: COMPUTE_WH

exposures:
  table: ANALYTICS.PUBLIC.STATSIG_EXPOSURES
  user_id_col: user_id
  experiment_id_col: experiment_name
  variant_col: group
  timestamp_col: exposure_time

outcomes:
  table: ANALYTICS.PUBLIC.EVENTS
  user_id_col: user_id
  event_name_col: event_type
  value_col: revenue
  timestamp_col: event_time
  target_events:
    - purchase
    - add_to_cart

user_features:
  table: ANALYTICS.PUBLIC.USER_DIM
  user_id_col: user_id
  feature_cols:
    - country
    - device_type
    - account_age_days
  covariate_col: pre_experiment_revenue

experiment:
  experiment_id: checkout_redesign_2024
  control_variant: control
  treatment_variant: treatment
  alpha: 0.05
  winsorize_percentile: 0.99
  use_cuped: true

3. Run

from argenta import ArgentoConfig, PipelineRunner
from argenta.config.loader import load_config

config = load_config("argenta.yaml")
runner = PipelineRunner(config)
result = runner.run("checkout_redesign_2024")

print(result)

Results are also written back to your warehouse at argenta.experiment_results.

---

Architecture

Your warehouse (Snowflake / BigQuery / Redshift)
         │
         │  read-only + write to argenta schema
         ▼
Argenta SQL pipeline (runs INSIDE your warehouse)
  ├── 1. Exposure deduplication   — first exposure per user wins
  ├── 2. Outcome join             — only events after first exposure
  └── 3. User feature join        — covariates for CUPED + future CATE
         │
         ▼
Argenta stats layer (Python)
  ├── ATE + Welch CI + p-value
  ├── Winsorization
  ├── SRM detection
  └── CUPED variance reduction
         │
         ▼
Results written back to your warehouse
  ├── argenta.experiment_results   — per-metric ATE, CI, p-value
  ├── argenta.user_cate_scores     — per-user CATE (Phase 2)
  └── argenta.segment_effects      — HTE by segment (Phase 2)

---

Statistical Methods

Method	Purpose
Welch's t-test	ATE + p-value
Confidence intervals	95% CI on ATE
Winsorization	Outlier handling for revenue metrics
SRM detection	Sample ratio mismatch check
CUPED	Variance reduction via pre-experiment covariate
Causal Forest (CausalForestDML)	Non-parametric CATE estimation
X-Learner	CATE with unbalanced treatment/control
Uplift modeling	Score users not in experiment
Sequential testing (mSPRT)	Always-valid p-values
Multiple testing correction	Bonferroni / BH-FDR across metrics

---

Input Tables Required

Argenta needs three tables in your warehouse (all column names are configurable):

Table	Required columns
Exposures	user_id, experiment_id, variant, timestamp
Outcomes / events	user_id, event_name, value, timestamp
User features	user_id, + any feature columns

The user features table is required for CUPED variance reduction. Without it, set use_cuped: false.

---

Output Tables

Argenta writes results back to {output_schema} in your warehouse (default: argenta):

Table	Contents
argenta.experiment_results	ATE, CI, p-value, SRM flag per metric
argenta.user_cate_scores	Per-user CATE score
argenta.segment_effects	HTE by segment

---

Supported Warehouses

Warehouse	Extra	Status
Snowflake	argenta[snowflake]	Supported
BigQuery	argenta[bigquery]	Supported
Redshift	argenta[redshift]	Supported
Databricks	argenta[databricks]	Planned

---

Documentation

• Architecture
• Onboarding a new warehouse
• SQL pipeline explained
• Statistics reference
• Contributing

---

Contributing

Contributions are welcome. See CONTRIBUTING.md for setup instructions, coding conventions, and the PR process.

---

License

Apache 2.0. See LICENSE.

---

Side note

The name "Argenta" is inspired by the name of a professor who introduced me to causal inference during my bachelor's degree.