run_ml_job

Overview

run_ml_job trains a classical machine learning model on feature-engineered data and produces per-ticker next-day log-return predictions for the ML-enhanced portfolio path. Supported libraries:

scikit-learn — linear, ridge, lasso, elastic net, quantile, random forest, …
XGBoost, LightGBM, CatBoost — gradient-boosting regressors

This stage is optional. The default QuantLab path is data → PO → trading. Use ML when you want predicted returns as μ (mu_mode: custom in run_po_job).

For XGBoost, LightGBM, and CatBoost, QuantLab uses the regressor APIs (e.g. XGBRegressor, LGBMRegressor, CatBoostRegressor) to predict continuous returns — not classifiers or clustering.

Hyperparameters — full library power

Pass any constructor argument the chosen model accepts via model_params. The worker forwards those kwargs into the library’s estimator __init__, so you can use the full API surface of sklearn / XGBoost / LightGBM / CatBoost (learning rate, depth, subsample, reg_alpha, l1_ratio, random_state, and so on). Complete config values are shown below. Use one alongside the required feature_url and data_extractor_url from your own preceding jobs; these objects are not standalone tool requests.

{
  "ML params": {
    "model": "xgboost",
    "model_params": {
      "n_estimators": 300,
      "max_depth": 5,
      "learning_rate": 0.05,
      "subsample": 0.8,
      "colsample_bytree": 0.8,
      "random_state": 42
    }
  }
}

{
  "ML params": {
    "model": "lightgbm",
    "model_params": {
      "n_estimators": 400,
      "num_leaves": 63,
      "learning_rate": 0.03,
      "min_child_samples": 20
    }
  }
}

{
  "ML params": {
    "model": "catboost",
    "model_params": {
      "iterations": 500,
      "depth": 6,
      "learning_rate": 0.05,
      "random_seed": 42
    }
  }
}

In chat you can just say “XGBoost with max_depth 5 and 300 trees” — MCP skills fill in model_params.

Parameters

string

required

URL of feature_engine_*.json from run_feature_worker.

string

required

URL of data_extractor_*.json from run_data_extraction (train/test date alignment).

object

required

ML configuration.

Show config fields

object

required

Show ML params fields

string

required

Model key (case-insensitive):

Value	Backend
`"random_forest"` / `"rf"`	scikit-learn RandomForestRegressor
`"xgboost"` / `"xgb"`	XGBoost regressor
`"lightgbm"` / `"lgbm"`	LightGBM regressor
`"catboost"`	CatBoost regressor
`"linear"` / `"linear_regression"`	sklearn LinearRegression
`"ridge"`	sklearn Ridge
`"lasso"`	sklearn Lasso
`"elastic_net"`	sklearn ElasticNet
`"quantile"`	sklearn QuantileRegressor

The left-hand names are the canonical values; a short alias is remapped to one of them and the substitution is listed in coercions. Check coercions and warnings in the result rather than assuming the model you named is the one that ran.

object

Dict of hyperparameters for the model constructor — any kwargs supported by that library’s regressor API.

integer

Samples per CV test fold (walk-forward). Default: 15. Legacy alias: test_size.

integer

Number of CV splits. Default: 15.

object

Optional MAPIE conformal intervals (enabled, alpha, bound).

Returns

{
  "status": "Succeeded",
  "output_url": "https://stqsnpprod.blob.core.windows.net/data/ml_engine_7d4e2a91c0bf.json",
  "output_name": "ml_engine_7d4e2a91c0bf.json",
  "execution_name": "ml-worker-job-abc123xyz",
  "effective_params": {
    "model": "elastic_net",
    "model_params": { "alpha": 0.001 },
    "n_splits": 15,
    "cv_test_size": 15,
    "intervals_enabled": false
  },
  "coercions": [],
  "warnings": []
}

effective_params is what the run actually used. A remapped alias, a rounded cv_test_size or an ignored block leaves the run Succeeded and shows up in coercions / warnings — read them before passing the forecasts on. If you want neural networks instead of classical models, use run_dl_job — its nn_engine_*.json is interchangeable with this output downstream.

Example

{
  "feature_url": "https://stqsnpprod.blob.core.windows.net/data/feature_engine_a1c3e5f70912.json",
  "data_extractor_url": "https://stqsnpprod.blob.core.windows.net/data/data_extractor_2b510101b9b7.json",
  "config": {
    "ML params": {
      "model": "xgboost",
      "model_params": {
        "n_estimators": 200,
        "max_depth": 4,
        "learning_rate": 0.05,
        "random_state": 42
      },
      "cv_test_size": 15,
      "n_splits": 15,
      "intervals": {
        "enabled": false,
        "alpha": 0.99,
        "bound": "center"
      }
    }
  }
}

Next Step

Pass output_url to run_po_job as input_url. On an ml_engine_ input the optimizer defaults to mu_mode: "custom", so the forecasts act as expected returns.
Before that, it is worth looking at the forecasts themselves: run_plot_job with block_name: "ml_visualization" and tool_name: "plot_compare_pred_vs_real_for_ticker" puts predicted and realized returns on one chart.

Overview

How To

Tools description

Overview

Hyperparameters — full library power

Parameters

Returns

Example

Next Step

​Overview

​Hyperparameters — full library power

​Parameters

​Returns

​Example

​Next Step

Overview

Hyperparameters — full library power

Parameters

Returns

Example

Next Step