limen.cohort

Canonical docs

In the research projects

The Python project composes selected, already evaluated fitted objects through Cohort.from_python. The Compiled project walk_forward scenario binds reconstructed qualified folds. Member identity, output units and available context must agree before aggregation.

What this package owns

This package owns the cohort-level inference surface:

Cohort.from_python(sensors) for live mean/vote/custom aggregation, and artifact-backed Cohort construction for selected experiment permutations
single-file cohort (sfc) strategies for choosing those permutations from experiment artifacts

It does not own:

experiment search execution (owned by limen.experiment)
raw market data infrastructure
downstream decisioning/execution logic (outside Limen)

Key entry points

Entry point	Use case	Notes
`Cohort`	Aggregate retained Python fits or compatible reconstructed native members at inference time	Supports probability-weighted and majority-vote fallback modes
`bayesian_cohort`	Bind observed binary outcomes to live Python members	Fixed Beta–Bernoulli reliability-weighted voting; see the canonical Cohort documentation
`BUILTIN_SELECTORS`	Map selector names to built-in selector modules	Exported at the package root
`limen.cohort.sfc.all.select`	Default "use every round" selector	Preserves existing omitted-`permutation_ids` behavior
`limen.cohort.sfc.top_n.select`	Results-column ranker	Requires `results.csv`
`limen.cohort.sfc.backtest_pareto.select`	Trading-metric Pareto selection	Uses backtest return/risk columns
`limen.cohort.sfc.diverse_metrics.select`	Metric-diverse member selection	Uses PCA/KMeans medoids over metric columns

Package map

Module or family	Responsibility
`cohort.py`	Cohort entry points + compiled aggregation
`_python_cohort.py`	Live composition and callback boundary
`sfc/`	Single-file cohort strategies
`all.py`	select(context) -> ids
`top_n.py`	select(context, *, column, n) -> ids
`backtest_pareto.py`	select(context, *, target_count) -> ids
`diverse_metrics.py`	select(context, *, target_count) -> ids

text · README.md

1cohort/2├── README.md3├── __init__.py4├── _fold_sources.py5├── _python_cohort.py6├── _round_sources.py7├── bayesian.py8├── cohort.py9└── sfc/

Merged source PR · 277022e9Package layout · Limen 6.0.1 · current package validation · package source

Cohort quick behavior

For live fitted Sensors, start with Cohort.from_python(sensors). Matching explicit Sensor output IDs permit default mean; optional preparation, output mapping and aggregation functions provide custom control. Compiled members require an explicit mapper/common meaning/cutoff. See the Python contract for alignment, blocking, availability and live provenance.

The artifact-backed constructor follows these rules:

Construct from exactly one experiment source (experiment_id or experiment_log_path)
Validate explicit or selector-produced permutation_ids
Bind each member to its selected source and architecture; admit mixed declared binary outputs
Infer homogeneous aggregation from architecture capability; bind mixed aggregation from native binary declarations:
- probability_weighted when probability output is supported
- majority_vote fallback otherwise
predict(raw_klines) returns one BarPrediction for the last bar
predict_all(raw_klines) returns list[BarPrediction]
__call__(raw_klines) aliases predict_all(raw_klines)

See Cohort for full contract and examples.

Selector quick behavior

Selectors are one-file strategies with one public select(context) function
Selectors receive a context dict and return permutation IDs only
Cohort owns all validation after selector execution
Built-ins cover all-round, single-column rank, backtest Pareto, and metric diversity selection

Adjacent modules

limen.experiment owns manifests and parameter search; limen.inference owns Trainer and Sensor reconstruction used by Cohort flows
limen.log provides analysis surfaces commonly used before selector-driven promotion
limen.sfd and reference architectures define member model output behavior