limen.cohort

Inference-time cohort surfaces for aggregating decoder outputs from completed Limen experiments.

Canonical docs

• Cohort
• Trainer

What this package owns

This package owns the cohort-level inference surface:

1. Cohort for direct multi-member aggregation from one experiment and selected permutations
2. 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 selected decoders from one completed experiment at inference time	Supports probability-weighted and majority-vote fallback modes
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

cohort/
├── cohort.py                # Cohort constructor + aggregation logic
└── 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

Cohort quick behavior

• Construct from exactly one experiment source (experiment_id or experiment_log_path)
• Validate explicit or selector-produced permutation_ids
• Enforce single-architecture selection
• Infer aggregation mode from architecture capability:
  • 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

Read next

• Cohort
• Trainer