Canonical docs
In the research projects
The Python project fits its ordinary caller-owned Ridge pipeline. The Compiled project search scenario runs the declared CPU model comparisons, retains their metrics and prediction payloads; the selection scenario separately proves CLI reconstruction. The architecture reference describes the prepared-data and output contracts.
What this package owns
Owns the reference model classes (and their function forms) that implement fit/predict for each decoder family, without the packaged data pipeline.
Does not own the experiment shape or manifest (owned by the matching limen.sfd.foundational_sfd module) or experiment execution.
Key entry points
| Entry point | Use case | Notes |
|---|---|---|
DirectionSizingModel / direction_sizing |
Direction and signed exposure model | Native allocation/execution contract |
meta_allocation.py |
Fold-qualified allocation labels and policy fitting | Used by allocation manifests; distinct from entry classification |
ReferenceModel |
Base class for reference models | Shared fit/predict contract |
LogRegBinary / logreg_binary |
Logistic-regression binary model | Class and function forms |
DLinearRegressor / dlinear_regressor |
Linear time-series regressor | Class and function forms |
LightGBMBinary / lightgbm_binary |
LightGBM binary classifier | Class and function forms |
RidgeRegressor / ridge_regressor |
sklearn ridge regressor | Class and function forms |
XGBoostRegressor / xgboost_regressor |
XGBoost regressor | Class and function forms |
RandomBinary / random_binary |
Baseline random classifier | Class and function forms |
RuleBasedStrategy / rule_based |
Predicate-driven, non-learned strategy | Uses limen.sfd.rule_based |
TabPFNBinary / tabpfn_binary |
TabPFN binary classifier | Optional; requires tabpfn |
Adjacent modules
limen.sfd.foundational_sfdpairs each of these models with an experiment shape and manifest.limen.inference.Sensorwraps aReferenceModelfor live bar-by-bar prediction.limen.metricsis commonly used inside these model functions to score outputs.
Quick orientation
| Module or family | Responsibility |
|---|---|
base.py |
ReferenceModel base class |
tabpfn_binary.py |
Optional |
reference_architecture/├── README.md├── __init__.py├── _backtest_evaluation.py├── _chronological_link.py├── _pseudo_pool.py├── _sequence_binary.py├── base.py├── cnn1d_binary.py├── direction_sizing.py├── dlinear_regressor.py├── entropy_var_regressor.py├── feature_link_binary.py├── gaussian_hmm_binary.py├── lightgbm_binary.py├── logreg_binary.py├── lstm_binary.py├── meta_allocation.py├── pseudo_label_binary.py├── random_binary.py├── rbf_svm_binary.py├── residual_regressor.py├── ridge_regressor.py├── rule_based.py├── tabpfn_binary.py├── weight_link_binary.py└── xgboost_regressor.pyMerged source PR · 25e47866Package layout · Limen 6.0.1 · current package validation · package source
Things to know
- Each module exposes both a class (e.g.
LogRegBinary) and a function form (e.g.logreg_binary); pick the class for custom subclassing and the function for manifest wiring. - These models are the template for custom SFD work: subclass
ReferenceModelor copy a module to start. - A reference architecture owns the training logic; the matching foundational SFD owns the manifest and data pipeline.