Find an API by task

Start with the task below, then follow the linked signature and its input contract. The module index groups concepts; the Python reference lists importable namespaces, and rustdoc lists Rust paths. A module guide’s title is not necessarily its Python import name. If a call rejects your inputs or returns an unsuitable result, use Diagnose input errors.

Choose the operation

Task	Python entry point	What to check before using it
Preserve selected source values through cleaning	provenance.select_revisions, rejoin_selected	Select before cleaning and rejoin by stable IDs. Exact typed identity does not prove provider authenticity or history completeness.
Assemble complete declared feature rows	feature_assembly.assemble_features	Specify feature order and required source fields. Retain missing-input exclusions and remap original row indices after filtering.
Audit training-time availability	availability.select_available_training	Supply explicit source and label clocks and the splitter’s original training indices; this filter does not replace purging or fold-local preprocessing.
Build meta-labels from a primary model’s sides	labeling.get_events, labeling.get_bins	Preserve primary-side provenance. These functions consume supplied sides; they do not fit either model or enforce out-of-sample prediction boundaries.
Load and inspect local OHLCV data	data.load_ohlcv, data.data_quality_report	First resolve point-in-time revision availability; loading and cleaning do not enforce it. The Python interface takes frames/files; the native _core.data exposes separate column and frame calls. Rust uses data_processing.
Explain why rows left a validation fold	cross_validation.split_with_diagnostics	Supply ordered label spans in consistent units. Diagnostics describe purge/embargo exclusions, not whether feature construction leaked.
Bootstrap labels inside a training fold	sampling.seq_bootstrap; worked handoff	Draws index training-local label columns. Map through training indices to original rows, preserve multiplicity across fitting arrays, and keep test rows excluded. Repeated draws are not independent new observations.
Score predictions already produced out of sample	feature_importance.mda_from_probabilities, sfi_from_probabilities	These consume predictions you supply; they do not fit your model. Preserve original rows and feature columns; see saved-prediction handoffs.
Fit the built-in diagnostic models on features	feature_diagnostics.mda_importance, sfi_importance	These are different interfaces from the prediction consumers above. Read the diagnostic model and preprocessing assumptions.
Track the trial count used for selection-adjusted reporting	evaluation.TrialRegistry	Counts unique explicit configuration keys; repeating a key replaces its entry. Include all attempted choices in those keys; see the trial-identity contract.
Allocate using supplied returns or covariance	portfolio.allocate_from_inputs	Follow observations to named weights for timestamp/asset alignment, then check covariance validity, bounds and units in portfolio optimization.
Assemble predictions into CPCV paths	backtesting_engine.assemble_cpcv_paths	Restore split and sample order before assembly. Paths reuse observations and are dependent; their spread is descriptive.
Convert an API result into a frame for analysis	adapters	Choose the adapter for the result type. Conversion does not establish timestamp alignment, calibration, or account-level P&L.

For an integrated example, use the Python workflow or Rust workflow. Both are API demonstrations; training and held-out evaluation remain separate work. For a fitted baseline and held-out predictions, follow the end-to-end research journey. For parameter selection before one final chronological assessment, use the model-selection walkthrough. It connects purged inner splits, training-only preprocessing and explicit refitting; its final classification score is separate from the search score and is not trading P&L.

Python wrappers and native calls

Use openquant.<namespace> as the starting point. Seven namespaces have a Python wrapper and a separately documented native surface: backtesting_engine, bars, cross_validation, data, feature_importance, hyperparameter_tuning, and pipeline. Their low-level references are named openquant._core.<namespace>. A shared function name does not imply identical accepted inputs or return types.

For example, the validation wrapper returns NumPy index arrays while its native counterpart returns lists. This small example also checks that fold row positions agree; the integers here are abstract, consistently ordered time units.

import numpy as np
from openquant import cross_validation
from openquant._core import cross_validation as native_cv

t0 = [0, 2, 4, 6, 8, 10]
t1 = [1, 3, 5, 7, 9, 11]
wrapped = cross_validation.purged_kfold_splits(t0, t1, 3, 0.0)
native = native_cv.purged_kfold_splits(t0, t1, 3, 0.0)
assert isinstance(wrapped[0][0], np.ndarray)
assert isinstance(native[0][0], list)
assert [(train.tolist(), test.tolist()) for train, test in wrapped] == native
print(wrapped[0][0].tolist(), wrapped[0][1].tolist())

[2, 3, 4, 5] [0, 1]

Other native namespaces such as risk, hrp, and fracdiff are re-exported directly at the package root. Pure-Python namespaces such as evaluation, feature_diagnostics, and viz have no same-name _core module.

Translate a name between languages

These common aliases help when moving from a Python example to rustdoc. They identify related modules, not interchangeable function signatures.

Python namespace	Rust module
bars	data_structures
data	data_processing for OHLCV operations; Python caching/fetching is a separate layer
risk	risk_metrics
portfolio	portfolio_optimization
backtest_stats	backtest_statistics
synthetic_bt	synthetic_backtesting
microstructural	microstructural_features
ensemble	ensemble_methods
fast_ewma, volatility	util::fast_ewma, util::volatility

Tasks that currently require Rust

These public Rust APIs have no Python binding. Similar names elsewhere do not supply the same operation.

Task	Rust entry point and worked guide
Build a rebalanced futures basket or roll series	etf_trick::EtfTrick, get_futures_roll_series; guide
Decompose a fitted predictor’s feature effects	fingerprint::RegressionModelFingerprint; guide. Python feature importance measures are different diagnostics.
Dispatch a user-supplied computation over partitions	hpc_parallel::run_parallel; guide. Python streaming_hpc supplies specific streaming calculations, not this generic executor.
Supply a custom integer objective or solver adapter	combinatorial_optimization::IntegerObjective, SolverAdapter; guide. Python dynamic_allocation exposes a particular allocation problem, not these traits.