ml-pipeline（毫升管道）

ML Pipeline

Unified skill for the complete ML pipeline within a quant trading research system. Consolidates eight prior skills into a single authoritative reference covering the full lifecycle: data validation, feature creation, selection, transformation, anti-leakage checks, pipeline automation, deep learning optimization, and deployment.

1. When to Use

Activate this skill when the task involves any of the following:

• Creating, selecting, or transforming features for an ML-driven strategy.
• Auditing an existing feature pipeline for data leakage or overfitting risk.
• Automating an end-to-end ML pipeline (data prep through model export).
• Evaluating feature importance, scaling, encoding, or interaction effects.
• Integrating features with a feature store (Feast, Tecton, custom Parquet store).
• Explaining core ML concepts (bias-variance, cross-validation, regularisation)

in the context of feature engineering decisions.

2. Inputs to Gather

Before starting work, collect or confirm:

3. Feature Creation Patterns

3.1 Numerical Features

• Interaction terms: price * volume, high / low, close - open.
• Rolling statistics: mean, std, skew, kurtosis over configurable windows.
• Polynomial / log transforms: log(volume + 1), spread^2.
• Binning / discretisation: equal-width, quantile-based, or domain-driven bins.

3.2 Categorical Features

• One-hot encoding: for low-cardinality categoricals (sector, exchange).
• Target encoding: mean-target per category with smoothing (careful of leakage -- use only in-fold means).
• Ordinal encoding: when categories have a natural order (credit rating).

3.3 Time-Series Specific

• Lag features: return_{t-1}, return_{t-5}, etc.
• Calendar features: day-of-week, month, quarter, options-expiry flag.
• Rolling z-score: (x - rolling_mean) / rolling_std for stationarity.
• Fractional differentiation: preserve memory while achieving stationarity (Lopez de Prado).

3.4 Feature Selection Techniques

• Filter methods: mutual information, variance threshold, correlation pruning.
• Wrapper methods: recursive feature elimination (RFE), forward/backward selection.
• Embedded methods: L1 regularisation, tree-based importance, SHAP values.
• Permutation importance: model-agnostic; run on out-of-fold predictions.

4. Anti-Leakage Checks

Data leakage is the single most common cause of inflated backtest results. Apply these checks at every pipeline stage:

4.1 Label Leakage

• Labels must be computed from future returns relative to the feature

timestamp. Verify that the label window does not overlap the feature window.

• Use purging and embargo when labels span multiple bars.

4.2 Feature Leakage

• No feature may use information from time t+1 or later at prediction time t.
• Rolling statistics must use a closed left window: df['feat'].rolling(20).mean().shift(1).
• Target-encoded categoricals must be computed on the training fold only.

4.3 Cross-Validation Leakage

• Use purged k-fold or walk-forward CV for time-series. Never use random

k-fold on ordered data.

• Insert an embargo gap between train and test folds to prevent bleed-through

from autocorrelation.

4.4 Survivorship & Selection Bias

• Ensure the universe of instruments at time t reflects what was actually

tradable at that time (delisted stocks, halted symbols removed later).

• Backfill from point-in-time databases where available.

4.5 Validation Checklist

Run before every backtest:

[ ] Labels computed strictly from future returns (no overlap with features)
[ ] All rolling features shifted by at least 1 bar
[ ] Target encoding uses in-fold means only
[ ] Walk-forward or purged CV used (no random shuffle on time-series)
[ ] Embargo gap >= max(label_horizon, autocorrelation_lag)
[ ] Universe is point-in-time (no survivorship bias)
[ ] No global scaling fitted on full dataset (fit on train, transform test)

5. Pipeline Automation (AutoML)

5.1 Prerequisites

• Python environment with one or more AutoML libraries:

Auto-sklearn, TPOT, H2O AutoML, PyCaret, Optuna, or custom Optuna pipelines.

• Training data in CSV / Parquet / database.
• Problem type identified: classification, regression, or time-series forecasting.

5.2 Pipeline Steps

5.3 Pipeline Configuration Template

pipeline_config = {
    "task_type": "classification",        # or "regression", "time_series"
    "time_budget_seconds": 3600,
    "algorithms": ["rf", "xgboost", "catboost", "lightgbm"],
    "preprocessing": ["scaling", "encoding", "imputation"],
    "tuning_strategy": "bayesian",        # or "random", "hyperband"
    "cv_folds": 5,
    "cv_type": "purged_kfold",            # or "walk_forward"
    "embargo_bars": 10,
    "early_stopping_rounds": 50,
    "metric": "sharpe_ratio",             # domain-specific metric
}

5.4 Output Artifacts

• automl_config.py -- pipeline configuration.
• best_model.pkl / .joblib / .onnx -- serialised model.
• feature_pipeline.pkl -- fitted preprocessing + feature transforms.
• evaluation_report.json -- metrics, confusion matrix / residuals, feature rankings.
• deployment/ -- prediction API code, input validation, requirements.txt.

6. Core ML Fundamentals (Feature-Engineering Context)

6.1 Bias-Variance Trade-off

• More features increase model capacity (lower bias) but risk overfitting (higher variance).
• Use regularisation (L1/L2), feature selection, or dimensionality reduction to manage.

6.2 Evaluation Strategy

• Walk-forward validation: the gold standard for time-series strategies.

Roll a fixed-width training window forward; test on the next out-of-sample period.

• Monte Carlo permutation tests: shuffle labels and re-evaluate to estimate

the probability that observed performance is due to chance.

• Combinatorial purged CV (CPCV): generate many train/test combinations with

purging for more robust performance estimates.

6.3 Feature Scaling

• Fit scalers (StandardScaler, MinMaxScaler, RobustScaler) on the training set only.
• Apply the same fitted scaler to validation and test sets.
• RobustScaler is often preferred for financial data due to heavy tails.

6.4 Handling Missing Data

• Forward-fill then backward-fill for price data (be aware of leakage on backfill).
• Indicator column for missingness can itself be informative.
• Tree-based models can handle NaN natively; linear models cannot.

7. Workflow

For any feature engineering task, follow this sequence:

1. Restate the task in measurable terms (metric, constraints, deadline).
2. Enumerate required artifacts: datasets, feature definitions, configs, scripts, reports.
3. Propose a default approach and 1-2 alternatives with trade-offs.
4. Implement feature pipeline with anti-leakage checks built in.
5. Validate with walk-forward CV, Monte Carlo, and the leakage checklist above.
6. Deliver repo-ready code, documentation, and a run command.

8. Deep Learning Optimization

8.1 Optimizer Selection

8.2 Learning Rate Scheduling

• OneCycleLR: Best for short training, fast convergence
• CosineAnnealing: Smooth decay, good generalization
• ReduceOnPlateau: Adaptive when validation loss plateaus
• Warmup + Decay: Standard for transformers

8.3 Regularization Techniques

• Dropout: 0.1-0.5 for fully connected layers
• L2 (Weight Decay): 1e-4 to 1e-2
• Batch Normalization: Stabilizes training
• Early Stopping: Monitor validation loss, patience 5-10 epochs

8.4 PyTorch Lightning Integration

import pytorch_lightning as pl

class TradingModel(pl.LightningModule):
    def configure_optimizers(self):
        optimizer = torch.optim.AdamW(self.parameters(), lr=1e-4)
        scheduler = torch.optim.lr_scheduler.OneCycleLR(
            optimizer, max_lr=1e-3, total_steps=self.trainer.estimated_stepping_batches
        )
        return [optimizer], [scheduler]

8.5 Financial Reinforcement Learning

• State: Market features, portfolio state, position
• Action: Buy/Sell/Hold, position sizing
• Reward: Risk-adjusted returns (Sharpe, Sortino)
• Frameworks: Stable-Baselines3, RLlib, FinRL

9. Error Handling

10. Bundled Scripts

All scripts live in scripts/ within this skill directory.

11. Resources

Frameworks

• scikit-learn -- preprocessing, feature selection, pipelines.
• Auto-sklearn / TPOT / H2O AutoML / PyCaret -- automated pipeline search.
• Optuna -- flexible hyperparameter optimisation.
• SHAP -- model-agnostic feature importance.
• Feast / Tecton -- feature store management.
• PyTorch Lightning -- https://lightning.ai/docs/pytorch/stable/
• Stable-Baselines3 -- https://stable-baselines3.readthedocs.io/
• FinRL -- https://github.com/AI4Finance-Foundation/FinRL

Key References

• Lopez de Prado, *Advances in Financial Machine Learning* (2018) -- purged CV, fractional differentiation, meta-labelling.
• Hastie, Tibshirani & Friedman, *The Elements of Statistical Learning* -- bias-variance, regularisation, model selection.
• scikit-learn user guide: feature extraction, preprocessing, model selection.

Best Practices

• Always start with a simple baseline before running AutoML.
• Balance automation with domain knowledge -- blind search rarely beats informed priors.
• Monitor resource consumption; set hard timeouts.
• Validate on true out-of-sample holdout data, not just cross-validation.
• Document every pipeline decision for reproducibility.