Research models and strategies for Bitcoin trading

Limen is a research engine for Bitcoin trading. The engine prepares observations, evaluates model and strategy parameters, records research evidence, and carries selected models into inference through Sensors and Cohorts.

Limen supports activity and adaptive bars, feature and target design, classification, continuous and quantile forecasts, rule strategies, and linked learning. Search, controlled perturbations, walk-forward evaluation, costed execution, and protected final assessment connect those choices to recorded outcomes. The research book develops that sequence through two complete executable projects.

Two interfaces, different purposes

Limen offers Python Research and Compiled Research. Python Research defines preparation, fitting, backtesting, and selection through ordinary Python functions and caller-chosen libraries. Compiled Research declares those choices in YAML and executes them through the command-line interface (CLI), within the Manifest contract. Both use the same experiment engine. Indicators, features, targets, scalers, backtesting tools, and inference adapters support composition in Python and declared use in Compiled Research.

Choose your path

Install

Both paths use the same installation:

python -m pip install "vaquum-limen[data]"

Requires Python >=3.10,<3.14 (3.10–3.13) on macOS or Linux. Optional extras: boosting for LightGBM/XGBoost, tabpfn for TabPFN, neural for LSTM/1D CNN, hmm for GaussianHMM, indicators for TA-Lib comparison tooling, stats for statistical helpers, or vaquum-limen[all] for the full stack.

Python Research

Python functions define the research methodology: data acquisition, features, targets, fitting, backtesting, and selection. Ordinary estimators, custom learning stages, and Limen components can participate in the same experiment.

The path is observations → experiment → search → selection → fitted models → Sensor. The complete Python project defines every research component explicitly, including its imports, entry point, and recorded input checks. The excerpts below belong to that complete source.

Define your experiment

Start with a momentum feature and the next bar's open-to-close return:

def features(bars, window):
    return bars.with_columns(
        (pl.col("close") / pl.col("close").shift(window) - 1).alias("momentum")
    )

def prepare_split(bars, round_params):
    return features(bars, round_params["window"]).with_columns(
        (pl.col("close").shift(-1) / pl.col("open").shift(-1) - 1)
        .alias("next_return")
    ).slice(12).drop_nulls(["momentum", "next_return"])

Split before computing features and forward labels. Each candidate loses the same 12-bar warm-up and its partition's final forward-label row:

def prep(bars, round_params):
    # Split before features and labels; no target crosses a boundary.
    return {
        "train": prepare_split(bars[:960], round_params),
        "validation": prepare_split(bars[960:1440], round_params),
    }

Fit a scaler and Ridge model on training observations only, then score the validation forecasts with an explicit one-bar policy:

def fit(train, round_params):
    return make_pipeline(StandardScaler(), Ridge(alpha=round_params["alpha"])).fit(
        train.select("momentum"), train["next_return"].to_numpy()
    )

def backtest(validation, forecasts):
    positions = np.sign(forecasts)
    net = positions * validation["next_return"].to_numpy() - abs(positions) * 0.0002
    return float(np.mean(net))

The score uses signed forecasts, fixed notional, and a round-trip cost of two basis points. The caller defines this calculation. Limen's event backtester instead applies a declared execution, product, and funding policy.

Run the experiment

The project searches four candidates, retains their evaluated fitted pipelines and chooses the best validation result per feature window. Alpha breaks ties:

def run_research(bars, experiment_name):
    research = bars[:1440]
    uel, retained = run_trials(research, experiment_name)
    selected = choose_parameters(uel)
    sensors = selected_sensors(bars, selected, retained)
    cohort = Cohort.from_python(sensors)
    return cohort, uel, selected, retained

Use the project and recorded input bundle linked with the example evidence. From the bundle root, run:

python python-research/research.py --scenario selection --data-dir inputs --output-dir results/python

The command checks the recorded source digest and writes research.csv and summary.json. The evaluated models and Cohort remain in caller-owned memory during the run. Import run_research to continue using the returned objects in the same process; the two output files do not restore arbitrary fitted state.

Create a Sensor

The function above wraps each selected evaluated pipeline with Sensor.from_model. The Sensor reuses the fitted scaler and model, selects the feature names recorded during fitting, and blocks null or nonfinite features and decisions before the selection cutoff. sensor.predict(history) returns the latest numeric forecast or an explicit unavailable reason. Availability remains a separate declared clock.

Python functions supply preparation and its causal assumptions. Sensor.from_python accepts arbitrary fitted state and inference logic, with custom prediction and output mapping. Captured fitted state preserves explicitly retained estimators in a .uel object. The project restores those estimators and verifies their forecasts. See Python Sensor construction.

Create a sensor cohort

Python selection supplies the cohort members. The project combines two selected Sensors with Cohort.from_python, averaging compatible next-bar return forecasts. cohort.predict(history) returns the aggregated prediction. Members declare the same output_id to identify compatible meanings; Limen checks timestamp alignment and preserves unavailable results. Binary voting and custom aggregation support other declared outputs. See Python cohorts.

This recorded case fits on the first 960 observations, compares candidates on the next 480, and reserves the final 480 for later inference and assessment. The complete project extends to adaptive bars, fitted feature transformations, targets, search and feedback, recovery, walk-forward evaluation, calibration, event execution, and holdout decisions. A linked-learning case fits Ridge forecasts and an XGBoost residual correction, retains both stages, and selects the correction depth on validation.

Continue with custom SFDs, UEL, or search and feedback.

Compiled Research

Compiled Research executes a declared pipeline. The Manifest specifies data, chronological splits, preparation, model, and parameter space. The CLI validates and compiles the complete YAML definition before execution.

The complete Compiled project starts with a recorded-data logistic experiment. Its selection declaration fixes the source, dates, train-fitted target/scaler, execution policy and validation objective:

# region declaration
schema_version: "1.0"
metadata:
  name: book_compiled_selection
  mode: development
sfd:
  manifest:
    type: ml
# region recorded_source
    data_source:
      method: limen.data.HistoricalData.get_any_file
      params:
        file_path_or_url: inputs/spot_15m_20250101_20250531.parquet
# endregion recorded_source
    split_dates:
      train_start: "2025-01-01"
      train_end: "2025-01-15"
      val_start: "2025-01-15"
      val_end: "2025-01-18"
      test_start: "2025-01-18"
      test_end: "2025-01-22"
# region momentum_indicator
    indicators:
      - func: limen.indicators.roc
        params:
          period: 1
          group: momentum
# endregion momentum_indicator
# region fitted_target
    target:
      name: quantile_flag
      class: limen.targets.QuantileBinaryTarget
      fit_params:
        source_column: roc_1
        quantile: 0.5
      transform_params:
        shift: -1
# endregion fitted_target
# region validation_objective
    objective:
      metric: backtest_total_return
      direction: maximize
# endregion validation_objective
    backtest:
      source_interval_seconds: 900
      fee_bps: 1.0
      slip_bps: 0.0
      product:
        kind: cash_spot
        instrument: BTCUSDT
        base_currency: BTC
        quote_currency: USDT
        quantity_step: 1.0e-9
        min_notional: 0.0
# region fitted_scaler
    scaler:
      from_params: scaler_type
# endregion fitted_scaler
    reference_architecture: limen.sfd.reference_architecture.logreg_binary
  params:
    scaler_type: [logreg]
    C: [1.0, 0.5]
    random_state: [42]
uel:
  output_path: compiled-cli
  n_permutations: 2
  search_strategy:
    type: grid
  output_format: csv

# region assessment
assessment:
  metric: val_backtest_total_return
  direction: maximize
  top_k: 1
  predict: true
# endregion assessment
# endregion declaration

From the example bundle root, run that exact declaration through the installed CLI:

limen run --no-progress-bar compiled-research/selection.yaml

The command validates and compiles the declaration, executes two candidates and writes results/dev/compiled-cli. The CLI lifecycle also exercises template initialization, explicit validation/profiling, dry compilation, immutable committed definitions, recovery and project backup.

limen assess compiled-research/selection.yaml --results results/dev/compiled-cli

After execution, limen assess applies the YAML's validation selection, reconstructs the selected Sensors, and binds their Cohort. Reconstruction refits under the recorded pipeline. When results.csv is present, Trainer compares eligible overlapping numeric metrics within declared tolerances. The comprehensive YAML continues through feature ablation, PCA, calibration, threshold selection, full recording, and inference. Separate declarations cover model families, rule strategies, adaptive bars, linked learning, protected research, and final assessment; the supported-workflow table records their admission boundaries.

Row evidence and recovery depend on the recording policy. Choose it before a run when later replay or continuation needs those payloads.

Continue with the CLI workflow, Manifest contract, or end-to-end workflow.

Shared safety

Both interfaces use the experiment engine and reusable research components. Python functions compose or replace those components. Compiled Research connects their safeguards through the declared pipeline.

Common foundations

The Mutable Search Queue (MSQ) and Universal Experiment Loop (UEL) provide parameter identities, explicit budgets, round results, and convergence-warning records. Search includes grid, seeded shuffled traversal, random, and Sobol strategies, with feedback-driven queue updates. Reducers prune the remaining queue from observed results. Compatible checkpoint and full-recording artifacts support continuation, while factorization reuses eligible preparation across candidates. Search seeds control candidate ordering; model, data, and environment identities remain separate. See search and recovery.

Sensors wrap fitted state and distinguish usable predictions from unavailable decisions. Cohorts check member timestamp alignment before aggregation. Python adapters preserve declared continuous forecasts or signed exposure through compatible aggregation. Native Cohorts aggregate binary decisions under their member-output contract; model and fold compatibility governs admission. See selected behavior for selection, cutoffs, and output meanings.

In Python Research, safeguards follow the selected Limen components. Caller-defined replacements supply their own checks and evidence.

What Compiled Research adds

The CLI validates schema, component references, parameter links, split declarations, and execution settings before running. Its native architectures cover binary classification, continuous and quantile forecasts, neural sequences, latent states, linked learning, and explicit rule strategies. Preparation supports grouped indicators, fitted targets and scalers, causal feature transformations, PCA, sampling, and controlled ablation. The model-choice chapter connects model outputs to execution and reconstruction contracts. The declared pipeline coordinates the following controls:

See the preparation contract, backtest mechanics and inference contract for each control's scope.

Additional validation, when enabled

• Walk-forward validation refits each fold with declared purge and embargo gaps. Recorded fold Sensors carry their fitting and selection cutoffs into compatible cohorts. Purge and embargo lengths must cover the declared label horizon.
• A validation-return objective binds supported threshold tuning and reducers to validation execution. DSR and PBO reports assess retained research evidence; failed or unavailable criteria are reported rather than stopping the run.
• Opt-in project holdout reservations exclude reserved observations from admitted research, recovery, and replay. Frozen selection binds the chosen members and assessment scenarios before final observations are consumed. Protected research and final assessment have distinct model and workflow requirements.

Compilation does not establish arbitrary feature causality, data correctness or future performance. Validation used for tuning is selection evidence; statistical reports do not create an untouched holdout.

Research scope

Limen covers research from recorded observations to reusable Sensors and Cohorts: preparation, model and rule comparison, robustness analysis, execution policy, selection, and assessment. Its built-in components focus on Bitcoin markets. The surrounding Vaquum tools have distinct roles:

Sensors and Cohorts also expose inference interfaces for integration with other trading systems.

Risk boundary

Limen is research software. Benchmark and backtest outputs are not investment advice, trading advice, regulatory approval, or a promise of future performance. Past performance is not predictive, and trading digital assets can result in total loss of capital.

Supply chain and provenance

Limen makes its package origin, build process and release evidence inspectable:

• Reviewed source. Protected merges require approval and passing checks for tests, coverage, static analysis, dependency vulnerabilities and packaging.
• Controlled builds. Workflow actions are pinned to full commit hashes; CI requirements are installed with hash verification. Packaging checks build the wheel and source distribution twice in the same CI environment, with a fixed timestamp, and compare their SHA-256 digests.
• Dependency maintenance. Weekly dependency update proposals accompany vulnerability audits of declared core dependencies. Exceptions require a recorded reason and expiry date.
• Verifiable releases. PyPI publishing authenticates through OpenID Connect using Trusted Publishing. GitHub releases carry the same wheel and source distribution, build-provenance attestations, SHA-256 digests and a CycloneDX software bill of materials (SBOM) describing the release environment.

Limen holds OpenSSF Best Practices Silver, based on a public project self-assessment. Silver's requirements include repeatable builds, cryptographically signed releases, dependency inventory and monitoring, automated testing and documented security assurance.

See how to verify a release, inspect the security assurance case, or report a vulnerability.

Learn more

See the documentation hub, Roadmap, and Security Assurance Case.

Contributing

Contribution starts through CONTRIBUTING.md, docs changes, or open issues.

Before contributing, start with the Developer docs.

Support

Use SUPPORT.md for support routes and scope boundaries.

Vulnerabilities

Report vulnerabilities privately through GitHub Security Advisories. Do not report vulnerabilities through public issues.

Citations

Published work should cite:

Vaquum Limen [Computer software]. (2026). Retrieved from GitHub.

Machine-readable citation metadata lives in CITATION.cff.

License

MIT License.