NexQuant/rdagent/scenarios/qlib/developer/factor_runner.py

from pathlib import Path

import pandas as pd
from pandarallel import pandarallel

from rdagent.core.conf import RD_AGENT_SETTINGS
from rdagent.core.utils import cache_with_pickle

pandarallel.initialize(verbose=1)

from rdagent.app.qlib_rd_loop.conf import FactorBasePropSetting
from rdagent.components.runner import CachedRunner
from rdagent.core.exception import FactorEmptyError
from rdagent.log import rdagent_logger as logger
from rdagent.scenarios.qlib.developer.utils import process_factor_data
from rdagent.scenarios.qlib.experiment.factor_experiment import QlibFactorExperiment
from rdagent.scenarios.qlib.experiment.model_experiment import QlibModelExperiment

DIRNAME = Path(__file__).absolute().resolve().parent
DIRNAME_local = Path.cwd()

# TODO: supporting multiprocessing and keep previous results


class QlibFactorRunner(CachedRunner[QlibFactorExperiment]):
    """
    Docker run
    Everything in a folder
    - config.yaml
    - price-volume data dumper
    - `data.py` + Adaptor to Factor implementation
    - results in `mlflow`
    """

    def calculate_information_coefficient(
        self, concat_feature: pd.DataFrame, SOTA_feature_column_size: int, new_feature_columns_size: int
    ) -> pd.DataFrame:
        res = pd.Series(index=range(SOTA_feature_column_size * new_feature_columns_size))
        for col1 in range(SOTA_feature_column_size):
            for col2 in range(SOTA_feature_column_size, SOTA_feature_column_size + new_feature_columns_size):
                res.loc[col1 * new_feature_columns_size + col2 - SOTA_feature_column_size] = concat_feature.iloc[
                    :, col1
                ].corr(concat_feature.iloc[:, col2])
        return res

    def deduplicate_new_factors(self, SOTA_feature: pd.DataFrame, new_feature: pd.DataFrame) -> pd.DataFrame:
        # calculate the IC between each column of SOTA_feature and new_feature
        # if the IC is larger than a threshold, remove the new_feature column
        # return the new_feature

        concat_feature = pd.concat([SOTA_feature, new_feature], axis=1)
        IC_max = (
            concat_feature.groupby("datetime")
            .parallel_apply(
                lambda x: self.calculate_information_coefficient(x, SOTA_feature.shape[1], new_feature.shape[1])
            )
            .mean()
        )
        IC_max.index = pd.MultiIndex.from_product([range(SOTA_feature.shape[1]), range(new_feature.shape[1])])
        IC_max = IC_max.unstack().max(axis=0)
        return new_feature.iloc[:, IC_max[IC_max < 0.99].index]

    @cache_with_pickle(CachedRunner.get_cache_key, CachedRunner.assign_cached_result)
    def develop(self, exp: QlibFactorExperiment) -> QlibFactorExperiment:
        """
        Generate the experiment by processing and combining factor data,
        then passing the combined data to Docker for backtest results.
        """
        if exp.based_experiments and exp.based_experiments[-1].result is None:
            logger.info(f"Baseline experiment execution ...")
            exp.based_experiments[-1] = self.develop(exp.based_experiments[-1])

        fbps = FactorBasePropSetting()
        env_to_use = {
            "PYTHONPATH": "./",
            "train_start": fbps.train_start,
            "train_end": fbps.train_end,
            "valid_start": fbps.valid_start,
            "valid_end": fbps.valid_end,
            "test_start": fbps.test_start,
            "feature_names": str(list(exp.base_features.keys())),
            "feature_expressions": str(list(exp.base_features.values())),
        }
        if fbps.test_end is not None:
            env_to_use.update({"test_end": fbps.test_end})

        if exp.based_experiments:
            SOTA_factor = None
            # Filter and retain only QlibFactorExperiment instances
            sota_factor_experiments_list = [
                base_exp for base_exp in exp.based_experiments if isinstance(base_exp, QlibFactorExperiment)
            ]
            if len(sota_factor_experiments_list) > 1:
                logger.info(f"SOTA factor processing ...")
                SOTA_factor = process_factor_data(sota_factor_experiments_list)

            # Process the new factors data
            logger.info(f"New factor processing ...")
            new_factors = process_factor_data(exp)

            if new_factors.empty:
                raise FactorEmptyError("Factors failed to run on the full sample, this round of experiment failed.")

            # Combine the SOTA factor and new factors if SOTA factor exists
            if SOTA_factor is not None and not SOTA_factor.empty:
                new_factors = self.deduplicate_new_factors(SOTA_factor, new_factors)
                if new_factors.empty:
                    raise FactorEmptyError(
                        "The factors generated in this round are highly similar to the previous factors. Please change the direction for creating new factors."
                    )
                combined_factors = pd.concat([SOTA_factor, new_factors], axis=1).dropna()
            else:
                combined_factors = new_factors

            # Sort and nest the combined factors under 'feature'
            combined_factors = combined_factors.sort_index()
            combined_factors = combined_factors.loc[:, ~combined_factors.columns.duplicated(keep="last")]
            new_columns = pd.MultiIndex.from_product([["feature"], combined_factors.columns])
            combined_factors.columns = new_columns
            logger.info(f"Factor data processing completed.")

            num_features = len(exp.base_features) + len(combined_factors.columns)

            # Due to the rdagent and qlib docker image in the numpy version of the difference,
            # the `combined_factors_df.pkl` file could not be loaded correctly in qlib dokcer,
            # so we changed the file type of `combined_factors_df` from pkl to parquet.
            target_path = exp.experiment_workspace.workspace_path / "combined_factors_df.parquet"

            # Save the combined factors to the workspace
            combined_factors.to_parquet(target_path, engine="pyarrow")

            # If model exp exists in the previous experiment
            exist_sota_model_exp = False
            for base_exp in reversed(exp.based_experiments):
                if isinstance(base_exp, QlibModelExperiment):
                    sota_model_exp = base_exp
                    exist_sota_model_exp = True
                    break
            logger.info(f"Experiment execution ...")
            if exist_sota_model_exp:
                exp.experiment_workspace.inject_files(
                    **{"model.py": sota_model_exp.sub_workspace_list[0].file_dict["model.py"]}
                )
                sota_training_hyperparameters = sota_model_exp.sub_tasks[0].training_hyperparameters
                if sota_training_hyperparameters:
                    env_to_use.update(
                        {
                            "n_epochs": str(sota_training_hyperparameters.get("n_epochs", "100")),
                            "lr": str(sota_training_hyperparameters.get("lr", "2e-4")),
                            "early_stop": str(sota_training_hyperparameters.get("early_stop", 10)),
                            "batch_size": str(sota_training_hyperparameters.get("batch_size", 256)),
                            "weight_decay": str(sota_training_hyperparameters.get("weight_decay", 0.0001)),
                        }
                    )
                sota_model_type = sota_model_exp.sub_tasks[0].model_type
                if sota_model_type == "TimeSeries":
                    env_to_use.update(
                        {"dataset_cls": "TSDatasetH", "num_features": num_features, "step_len": 20, "num_timesteps": 20}
                    )
                elif sota_model_type == "Tabular":
                    env_to_use.update({"dataset_cls": "DatasetH", "num_features": num_features})

                # model + combined factors
                result, stdout = exp.experiment_workspace.execute(
                    qlib_config_name="conf_combined_factors_sota_model.yaml", run_env=env_to_use
                )
            else:
                # LGBM + combined factors
                result, stdout = exp.experiment_workspace.execute(
                    qlib_config_name="conf_combined_factors.yaml",
                    run_env=env_to_use,
                )
        else:
            logger.info(f"Experiment execution ...")
            if exp.base_feature_codes:
                factors = process_factor_data(exp)
                factors = factors.sort_index()
                factors = factors.loc[:, ~factors.columns.duplicated(keep="last")]
                new_columns = pd.MultiIndex.from_product([["feature"], factors.columns])
                factors.columns = new_columns
                target_path = exp.experiment_workspace.workspace_path / "combined_factors_df.parquet"
                # Save the combined factors to the workspace
                factors.to_parquet(target_path, engine="pyarrow")
                logger.info(f"Factor data processing completed.")
                result, stdout = exp.experiment_workspace.execute(
                    qlib_config_name="conf_combined_factors.yaml",
                    run_env=env_to_use,
                )
            else:
                result, stdout = exp.experiment_workspace.execute(
                    qlib_config_name="conf_baseline.yaml",
                    run_env=env_to_use,
                )

        if result is None:
            logger.error(f"Failed to run this experiment, because {stdout}")
            raise FactorEmptyError(f"Failed to run this experiment, because {stdout}")

        exp.result = result
        exp.stdout = stdout

        # Protection Manager: Check if factor passes risk criteria
        try:
            self._run_protection_check(exp, result)
        except Exception as e:
            logger.warning(f"Protection check failed for factor {exp.hypothesis.hypothesis}: {e}")
            # Don't block the workflow, just log the warning

        return exp

    def _run_protection_check(self, exp, result: dict) -> None:
        """
        Run protection checks on backtest results.

        Parameters
        ----------
        exp : QlibFactorExperiment
            The experiment with backtest results
        result : dict
            Backtest metrics dictionary
        """
        from rdagent.components.backtesting.protections import ProtectionManager

        protection_manager = ProtectionManager()
        protection_manager.create_default_protections()

        # Extract returns and equity curve from backtest results
        returns = result.get("returns", [])
        timestamps = result.get("timestamps", [])
        current_equity = result.get("final_equity", 100000)
        peak_equity = result.get("peak_equity", current_equity)

        # Get factor name from hypothesis
        factor_name = "unknown"
        if hasattr(exp, "hypothesis") and exp.hypothesis is not None:
            factor_name = getattr(exp.hypothesis, "hypothesis", "unknown")

        protection_result = protection_manager.check_all(
            returns=returns,
            timestamps=timestamps,
            current_equity=current_equity,
            peak_equity=peak_equity,
            factor_name=factor_name,
        )

        if protection_result.should_block:
            logger.warning(
                f"Factor {factor_name} rejected by protection manager: {protection_result.reason}"
            )
            # Mark factor as rejected by protection
            exp.rejected_by_protection = True
            exp.protection_reason = protection_result.reason