code/methods/naive.py

"""Naive / statistical methods for the MacroLens unified API.

Five concrete classes, one per (model × subset of tasks):

    Persistence          T1            (N, horizon) -- last-close tile
    HistoricalAnalogue   T4            (N,)         -- per-event-type train mean
    LogSizeOLS           T2, T5        (N,)         -- log1p OLS on numerics
    SectorMedian         T3, T6        long-form    -- per-(sector, field) train median
    MetroMedian          T7            (N, 3)       -- per-(state, property_type) median

Every class:

* Inherits :class:`_JoblibSaveMixin` (state.joblib + manifest.json).
* Registers via :func:`register` so `ml.methods.<Name>` finds it.
* Receives ``task`` as a constructor arg; a ``ValueError`` is raised if
  the requested task is not in :attr:`tasks`.

Hard rules (enforced by ``tests/test_layer_isolation.py``):

* Zero IO (no ``pd.read_parquet``, no file reads, no config imports).
* Zero ``meta`` consumption — methods take only ``X`` (and at fit time,
  ``y``).
* Zero canonical-indices imports / subsampling.
* Zero eval imports.
"""

from __future__ import annotations

import importlib.metadata
import logging
from typing import Any, ClassVar

import numpy as np
import pandas as pd

from ._config import (
    HistoricalAnalogueConfig,
    LogSizeOLSConfig,
    MetroMedianConfig,
    PersistenceConfig,
    SectorMedianConfig,
)
from ._registry import register
from .base import Method, _JoblibSaveMixin

logger = logging.getLogger(__name__)
_PERSISTENCE_DEFAULT_CLOSE_IDX_LOGGED: bool = False


# ── Internal lib_versions helper ──────────────────────────────────────────


def _naive_lib_versions() -> dict[str, str]:
    """Versions of the three libs the naive family actually uses.

    Overrides the broader default in :class:`Method` (which probes torch /
    transformers / vllm too) -- those are not imported here and recording
    them on a naive-method RunRecord would be misleading.
    """
    out: dict[str, str] = {}
    for pkg in ("numpy", "pandas", "scikit-learn"):
        try:
            out[pkg] = importlib.metadata.version(pkg)
        except importlib.metadata.PackageNotFoundError:
            pass
    return out


# ── T1 — Persistence ──────────────────────────────────────────────────────


@register(
    name="persistence",
    family="naive",
    tasks=frozenset({"T1"}),
    config_class=PersistenceConfig,
)
class Persistence(_JoblibSaveMixin, Method):
    """Forecast the future close trajectory by tiling the last lookback close.

    Predict shape: ``(N, horizon)``. The class emits a constant trajectory
    per row equal to ``X[:, -1, close_feature_idx]`` tiled ``horizon``
    times. This is the trivial-floor baseline; any real T1 method should
    beat it.

    Notes
    -----
    * ``X`` is a numpy ndarray ``(N, lookback, F)`` with no column names,
      so the runner is responsible for passing the close-feature index
      via ``PersistenceConfig.close_feature_idx`` (default ``0``).
    * ``horizon`` is read from ``y.shape[1]`` at fit time and stored as
      ``self._horizon``; if ``config.horizon`` is set it overrides.
    """

    name: ClassVar[str] = "persistence"
    family: ClassVar[str] = "naive"
    tasks: ClassVar[frozenset[str]] = frozenset({"T1"})
    schema_version: ClassVar[int] = 1

    def __init__(
        self,
        *,
        task: str,
        config: PersistenceConfig | None = None,
        **kwargs: Any,
    ) -> None:
        if task not in self.tasks:
            raise ValueError(
                f"Persistence does not support task={task!r}; "
                f"supported: {sorted(self.tasks)}"
            )
        self.task = task
        self.config = config or PersistenceConfig(**kwargs)
        if self.task == "T1" and self.config.close_feature_idx == 0:
            # Information-only: the canonical loader puts close at idx 0 so the
            # default is fine in practice. Emit once per process via logger so
            # batch instantiation doesn't pollute stderr.
            global _PERSISTENCE_DEFAULT_CLOSE_IDX_LOGGED
            if not _PERSISTENCE_DEFAULT_CLOSE_IDX_LOGGED:
                logger.info(
                    "Persistence(task='T1'): close_feature_idx defaults to 0; "
                    "the runner should resolve the actual close column index "
                    "from meta.attrs['feature_names'] and pass it via config."
                )
                _PERSISTENCE_DEFAULT_CLOSE_IDX_LOGGED = True
        self._horizon: int | None = None
        self._close_feature_idx: int = int(self.config.close_feature_idx)

    @classmethod
    def default_config(cls) -> PersistenceConfig:
        return PersistenceConfig()

    def fit(self, X: np.ndarray, y: np.ndarray, *, seed: int = 42) -> "Persistence":
        if not isinstance(X, np.ndarray) or X.ndim != 3:
            raise ValueError(
                f"Persistence.fit: expected X shape (N, lookback, F); got "
                f"type={type(X).__name__} shape={getattr(X, 'shape', '?')}"
            )
        if not isinstance(y, np.ndarray) or y.ndim != 2:
            raise ValueError(
                f"Persistence.fit: expected y shape (N, horizon); got "
                f"type={type(y).__name__} shape={getattr(y, 'shape', '?')}"
            )
        # Honour explicit override; otherwise read horizon from y.
        self._horizon = (
            int(self.config.horizon)
            if self.config.horizon is not None
            else int(y.shape[1])
        )
        if self._close_feature_idx >= X.shape[2]:
            raise ValueError(
                f"Persistence.fit: close_feature_idx={self._close_feature_idx} "
                f"out of bounds for F={X.shape[2]}"
            )
        return self

    def predict(self, X: np.ndarray) -> np.ndarray:
        if self._horizon is None:
            raise RuntimeError("Persistence: call .fit(X, y) before .predict().")
        if not isinstance(X, np.ndarray) or X.ndim != 3:
            raise ValueError(
                f"Persistence.predict: expected (N, lookback, F); got "
                f"type={type(X).__name__} shape={getattr(X, 'shape', '?')}"
            )
        if self._close_feature_idx >= X.shape[2]:
            raise ValueError(
                f"Persistence.predict: close_feature_idx={self._close_feature_idx} "
                f"out of bounds for F={X.shape[2]}"
            )
        last_close = X[:, -1, self._close_feature_idx][:, np.newaxis]
        return np.tile(last_close, (1, self._horizon)).astype(np.float32)

    def lib_versions(self) -> dict[str, str]:
        return _naive_lib_versions()


# ── T4 — HistoricalAnalogue ───────────────────────────────────────────────


@register(
    name="historical_analogue",
    family="naive",
    tasks=frozenset({"T4"}),
    config_class=HistoricalAnalogueConfig,
)
class HistoricalAnalogue(_JoblibSaveMixin, Method):
    """Predict scenario return as the per-``event_type`` train mean.

    ``X`` is the T4 DataFrame ``[lookback, event_type, event_description]``;
    only ``event_type`` is consumed. Unseen event types fall back to the
    global train mean.

    Predict shape: ``(N,)`` float32.
    """

    name: ClassVar[str] = "historical_analogue"
    family: ClassVar[str] = "naive"
    tasks: ClassVar[frozenset[str]] = frozenset({"T4"})
    schema_version: ClassVar[int] = 1

    def __init__(
        self,
        *,
        task: str,
        config: HistoricalAnalogueConfig | None = None,
        **kwargs: Any,
    ) -> None:
        if task not in self.tasks:
            raise ValueError(
                f"HistoricalAnalogue does not support task={task!r}; "
                f"supported: {sorted(self.tasks)}"
            )
        self.task = task
        self.config = config or HistoricalAnalogueConfig(**kwargs)
        self._type_mean: dict[str, float] = {}
        self._global_mean: float = 0.0
        self._fitted: bool = False

    @classmethod
    def default_config(cls) -> HistoricalAnalogueConfig:
        return HistoricalAnalogueConfig()

    @staticmethod
    def _event_types(X: pd.DataFrame) -> np.ndarray:
        if not isinstance(X, pd.DataFrame):
            raise TypeError(
                f"HistoricalAnalogue: expected DataFrame X; got "
                f"{type(X).__name__}"
            )
        if "event_type" not in X.columns:
            raise ValueError(
                "HistoricalAnalogue: X is missing 'event_type' column."
            )
        return np.asarray(X["event_type"].values, dtype=object).astype(str)

    def fit(
        self, X: pd.DataFrame, y: np.ndarray, *, seed: int = 42
    ) -> "HistoricalAnalogue":
        et = self._event_types(X)
        y_arr = np.asarray(y, dtype=np.float64).ravel()
        if et.shape[0] != y_arr.shape[0]:
            raise ValueError(
                f"HistoricalAnalogue.fit: event_type/y length mismatch "
                f"({et.shape[0]} vs {y_arr.shape[0]})"
            )
        df = pd.DataFrame({"event_type": et, "ret": y_arr}).dropna(subset=["ret"])
        if df.empty:
            raise RuntimeError("HistoricalAnalogue.fit: no usable training rows.")
        self._type_mean = df.groupby("event_type")["ret"].mean().to_dict()
        self._global_mean = float(df["ret"].mean())
        self._fitted = True
        return self

    def predict(self, X: pd.DataFrame) -> np.ndarray:
        if not self._fitted:
            raise RuntimeError(
                "HistoricalAnalogue: call .fit(X, y) before .predict()."
            )
        et = self._event_types(X)
        return np.array(
            [self._type_mean.get(e, self._global_mean) for e in et],
            dtype=np.float32,
        )

    def lib_versions(self) -> dict[str, str]:
        return _naive_lib_versions()


# ── T2 / T5 — LogSizeOLS ──────────────────────────────────────────────────


# LogSizeOLS removed from the registry per panel-design decision (replaced by
# RandomForest in classical.py). Class kept to preserve save/load
# compatibility for any old checkpoint, but no longer registered → not
# discovered by ``methods.ALL_METHODS`` and not runnable through the
# unified runner.
class LogSizeOLS(_JoblibSaveMixin, Method):
    """OLS regression of ``log1p(market_cap)`` on ``log1p(numeric features)``.

    Classical-ML baseline (fitted parametric model): for T2 / T5 inputs,
    fits a ``sklearn.linear_model.LinearRegression`` on
    ``log1p(numeric_columns)`` plus optional sector dummies, with target
    ``log1p(actual_market_cap)``. Predicts in log space and returns the
    dollar-space prediction via ``np.expm1``.

    Predict shape: ``(N,)`` float32.
    """

    name: ClassVar[str] = "log_size_ols"
    family: ClassVar[str] = "classical"
    tasks: ClassVar[frozenset[str]] = frozenset({"T2", "T5"})
    schema_version: ClassVar[int] = 1

    def __init__(
        self,
        *,
        task: str,
        config: LogSizeOLSConfig | None = None,
        **kwargs: Any,
    ) -> None:
        if task not in self.tasks:
            raise ValueError(
                f"LogSizeOLS does not support task={task!r}; "
                f"supported: {sorted(self.tasks)}"
            )
        self.task = task
        self.config = config or LogSizeOLSConfig(**kwargs)
        self._model: Any = None
        self._numeric_cols: list[str] = []
        self._sector_dummy_cols: list[str] = []
        self._train_columns: list[str] = []  # post-concat order

    @classmethod
    def default_config(cls) -> LogSizeOLSConfig:
        return LogSizeOLSConfig()

    @staticmethod
    def _select_numeric(df: pd.DataFrame) -> list[str]:
        """Return the columns this method treats as size-proxy numerics.

        Anything numeric that is neither a key (``ticker``/``date``) nor
        the target itself qualifies. The exact set depends on the inputs
        file the loader projected onto -- T5 strips price-derived
        features upstream, so the same selector works for T2 and T5.
        """
        skip = {"ticker", "date", "actual_market_cap", "derived_market_cap"}
        return [
            c
            for c in df.columns
            if c not in skip and pd.api.types.is_numeric_dtype(df[c])
        ]

    def _build_features(
        self, X: pd.DataFrame, *, train_columns: list[str] | None
    ) -> pd.DataFrame:
        """log1p of numerics + (optional) sector dummies, aligned to train.

        ``train_columns`` is None at fit time; the post-concat order is
        captured for predict-time alignment.
        """
        numeric_cols = (
            self._numeric_cols if train_columns is not None else self._select_numeric(X)
        )

        num = X.reindex(columns=numeric_cols).apply(
            lambda s: pd.to_numeric(s, errors="coerce").fillna(0.0)
        )
        # log1p; clip negative values to 0 to keep the log defined.
        num = num.clip(lower=0.0)
        num = np.log1p(num)

        if self.config.sector_dummies and "sector" in X.columns:
            dum = pd.get_dummies(X["sector"], prefix="sec", dtype=np.float32)
        else:
            dum = pd.DataFrame(index=X.index)

        feat = pd.concat(
            [num.reset_index(drop=True), dum.reset_index(drop=True)], axis=1
        ).fillna(0.0)

        if train_columns is None:
            # fit-time -- record the resolved numeric cols + dummy cols
            self._numeric_cols = list(numeric_cols)
            self._sector_dummy_cols = list(dum.columns)
            self._train_columns = list(feat.columns)
            return feat

        # predict-time: align to train_columns (add missing as 0; drop extras)
        for c in train_columns:
            if c not in feat.columns:
                feat[c] = 0.0
        return feat[train_columns].fillna(0.0)

    def fit(self, X: pd.DataFrame, y: np.ndarray, *, seed: int = 42) -> "LogSizeOLS":
        from sklearn.linear_model import LinearRegression

        if not isinstance(X, pd.DataFrame):
            raise TypeError(
                f"LogSizeOLS.fit: expected DataFrame X; got {type(X).__name__}"
            )
        y_arr = np.asarray(y, dtype=np.float64).ravel()
        if y_arr.shape[0] != len(X):
            raise ValueError(
                f"LogSizeOLS.fit: y length {y_arr.shape[0]} != "
                f"X length {len(X)}"
            )

        # Drop rows with non-positive / NaN target (log1p needs >= 0).
        mask = np.isfinite(y_arr) & (y_arr > 0)
        if not mask.any():
            raise RuntimeError(
                "LogSizeOLS.fit: zero rows after dropping non-positive / NaN targets."
            )
        X_fit = X.loc[mask].reset_index(drop=True)
        y_fit = y_arr[mask]

        feat = self._build_features(X_fit, train_columns=None)
        target = np.log1p(y_fit.astype(np.float64))

        model = LinearRegression()
        model.fit(feat.values, target)
        self._model = model
        return self

    def predict(self, X: pd.DataFrame) -> np.ndarray:
        if self._model is None:
            raise RuntimeError("LogSizeOLS: call .fit(X, y) before .predict().")
        if not isinstance(X, pd.DataFrame):
            raise TypeError(
                f"LogSizeOLS.predict: expected DataFrame; got {type(X).__name__}"
            )
        feat = self._build_features(X, train_columns=self._train_columns)
        pred_log = self._model.predict(feat.values)
        # Clip in log space to avoid expm1 overflow.
        pred_log = np.clip(pred_log, 0.0, 50.0)
        return np.expm1(pred_log).astype(np.float32)

    def lib_versions(self) -> dict[str, str]:
        return _naive_lib_versions()


# ── T3 / T6 — SectorMedian ────────────────────────────────────────────────


@register(
    name="sector_median",
    family="naive",
    tasks=frozenset({"T3", "T6"}),
    config_class=SectorMedianConfig,
)
class SectorMedian(_JoblibSaveMixin, Method):
    """Predict each XBRL field as the per-(sector, field) train median.

    At fit time the method:
      1. Reads the set of fields to predict from ``y["field"].unique()``
         (locked in :attr:`fitted_fields`).
      2. Inner-joins ``y`` against ``X[ticker, fiscal_year, sector]`` to
         attach the train sector per row.
      3. Groups by ``(sector, field)`` -> median(value), with a per-field
         global median fallback for unseen sectors.

    Predict emits a long-form DataFrame with one row per
    ``(ticker, fiscal_year)`` × every fitted field, with the looked-up
    sector median (fallback to per-field global median).

    Predict columns: ``[ticker, fiscal_year, field, pred]``.
    """

    name: ClassVar[str] = "sector_median"
    family: ClassVar[str] = "naive"
    tasks: ClassVar[frozenset[str]] = frozenset({"T3", "T6"})
    schema_version: ClassVar[int] = 1

    def __init__(
        self,
        *,
        task: str,
        config: SectorMedianConfig | None = None,
        **kwargs: Any,
    ) -> None:
        if task not in self.tasks:
            raise ValueError(
                f"SectorMedian does not support task={task!r}; "
                f"supported: {sorted(self.tasks)}"
            )
        self.task = task
        self.config = config or SectorMedianConfig(**kwargs)
        self.fitted_fields: list[str] = []
        self._field_sector_median: dict[tuple[str, str], float] = {}
        self._field_global_median: dict[str, float] = {}

    @classmethod
    def default_config(cls) -> SectorMedianConfig:
        return SectorMedianConfig()

    @staticmethod
    def _check_xy(X: pd.DataFrame, y: pd.DataFrame) -> None:
        if not isinstance(X, pd.DataFrame):
            raise TypeError(
                f"SectorMedian: expected DataFrame X; got {type(X).__name__}"
            )
        if not isinstance(y, pd.DataFrame):
            raise TypeError(
                f"SectorMedian: expected DataFrame y; got {type(y).__name__}"
            )
        for col in ("ticker", "fiscal_year"):
            if col not in X.columns:
                raise ValueError(f"SectorMedian: X missing '{col}'")
        for col in ("ticker", "fiscal_year", "field", "value"):
            if col not in y.columns:
                raise ValueError(f"SectorMedian: y missing '{col}'")

    def fit(
        self, X: pd.DataFrame, y: pd.DataFrame, *, seed: int = 42
    ) -> "SectorMedian":
        self._check_xy(X, y)

        # Lock the set of fields to predict from y at fit time. The dataloader
        # projects T3/T6 y onto a curated dense panel, so what's in y IS the
        # full panel -- no need to extend with hardcoded defaults.
        fitted_fields = sorted(y["field"].astype(str).unique())
        if not fitted_fields:
            raise RuntimeError("SectorMedian.fit: y has zero distinct 'field' values.")
        self.fitted_fields = fitted_fields

        # Build the (ticker, fiscal_year) -> sector lookup from X.
        sec_col = "sector" if "sector" in X.columns else None
        if sec_col is None:
            # Without sector, we still fit but every row falls back to "Unknown".
            x_keys = X[["ticker", "fiscal_year"]].copy()
            x_keys["sector"] = "Unknown"
        else:
            x_keys = X[["ticker", "fiscal_year", "sector"]].copy()
        x_keys["ticker"] = x_keys["ticker"].astype(str)
        x_keys["fiscal_year"] = pd.to_numeric(
            x_keys["fiscal_year"], errors="coerce"
        ).astype("Int64")
        x_keys["sector"] = x_keys["sector"].astype(str).fillna("Unknown")
        x_keys = x_keys.drop_duplicates(subset=["ticker", "fiscal_year"])

        # Attach sector to y via inner join.
        gt = y[["ticker", "fiscal_year", "field", "value"]].copy()
        gt["ticker"] = gt["ticker"].astype(str)
        gt["fiscal_year"] = pd.to_numeric(gt["fiscal_year"], errors="coerce").astype(
            "Int64"
        )
        gt["field"] = gt["field"].astype(str)
        gt["value_num"] = pd.to_numeric(gt["value"], errors="coerce")

        joined = gt.merge(x_keys, on=["ticker", "fiscal_year"], how="left")
        joined["sector"] = joined["sector"].fillna("Unknown")

        valid = joined.dropna(subset=["value_num"])
        if valid.empty:
            raise RuntimeError(
                "SectorMedian.fit: no rows with finite numeric values after coercion."
            )

        self._field_sector_median = (
            valid.groupby(["field", "sector"])["value_num"].median().to_dict()
        )
        self._field_global_median = (
            valid.groupby("field")["value_num"].median().to_dict()
        )
        return self

    def predict(self, X: pd.DataFrame) -> pd.DataFrame:
        if not self.fitted_fields:
            raise RuntimeError("SectorMedian: call .fit(X, y) before .predict().")
        if not isinstance(X, pd.DataFrame):
            raise TypeError(
                f"SectorMedian.predict: expected DataFrame; got {type(X).__name__}"
            )
        for col in ("ticker", "fiscal_year"):
            if col not in X.columns:
                raise ValueError(f"SectorMedian.predict: X missing '{col}'")

        keys = X[["ticker", "fiscal_year"]].copy()
        if "sector" in X.columns:
            keys["sector"] = X["sector"].astype(str).fillna("Unknown")
        else:
            keys["sector"] = "Unknown"
        keys["ticker"] = keys["ticker"].astype(str)
        keys["fiscal_year"] = pd.to_numeric(
            keys["fiscal_year"], errors="coerce"
        ).astype("Int64")

        rows: list[dict[str, Any]] = []
        for ticker, fy, sector in zip(
            keys["ticker"].values,
            keys["fiscal_year"].values,
            keys["sector"].values,
        ):
            for fld in self.fitted_fields:
                med = self._field_sector_median.get(
                    (fld, sector),
                    self._field_global_median.get(fld, 0.0),
                )
                rows.append(
                    {
                        "ticker": ticker,
                        "fiscal_year": fy,
                        "field": fld,
                        "pred": float(med) if pd.notna(med) else 0.0,
                    }
                )
        return pd.DataFrame(rows, columns=["ticker", "fiscal_year", "field", "pred"])

    def lib_versions(self) -> dict[str, str]:
        return _naive_lib_versions()


# Backward-compat alias for the legacy class name still referenced by
# methods/__init__.py (its registry-driven rewrite is the end-of-Phase-2
# deliverable; keeping the alias avoids breaking the package import in
# the meantime). Both names point at the same class.
LogSizeRegression = LogSizeOLS


# ── T7 — MetroMedian ──────────────────────────────────────────────────────


@register(
    name="metro_median",
    family="naive",
    tasks=frozenset({"T7"}),
    config_class=MetroMedianConfig,
)
class MetroMedian(_JoblibSaveMixin, Method):
    """Predict rent / price as per-(state, property_type) train medians.

    Default metro key is ``state_property_type``; ``state`` and
    ``city_state`` are also supported via :attr:`MetroMedianConfig.metro_key`.

    Fit consumes ``y[address, rent, price]`` for the train labels and
    ``X[state, property_type]`` for the metro key. Unseen metros fall back
    to the global train median (per output).

    Predict columns: ``[address, pred_rent, pred_price]``.
    """

    name: ClassVar[str] = "metro_median"
    family: ClassVar[str] = "naive"
    tasks: ClassVar[frozenset[str]] = frozenset({"T7"})
    schema_version: ClassVar[int] = 1

    def __init__(
        self,
        *,
        task: str,
        config: MetroMedianConfig | None = None,
        **kwargs: Any,
    ) -> None:
        if task not in self.tasks:
            raise ValueError(
                f"MetroMedian does not support task={task!r}; "
                f"supported: {sorted(self.tasks)}"
            )
        self.task = task
        self.config = config or MetroMedianConfig(**kwargs)
        self._metro_rent_median: dict[str, float] = {}
        self._metro_price_median: dict[str, float] = {}
        self._global_rent_median: float = 0.0
        self._global_price_median: float = 0.0
        self._has_rent: bool = False
        self._has_price: bool = False
        self._fitted: bool = False

    @classmethod
    def default_config(cls) -> MetroMedianConfig:
        return MetroMedianConfig()

    def _metro_key(self, df: pd.DataFrame) -> np.ndarray:
        """Build the per-row metro key from X according to config.metro_key."""
        kind = self.config.metro_key
        if kind == "state":
            cols = ["state"]
        elif kind == "city_state":
            cols = ["city", "state"]
        elif kind == "state_property_type":
            cols = ["state", "property_type"]
        else:  # pragma: no cover -- pydantic Literal forbids other values
            raise ValueError(f"Unknown metro_key={kind!r}")

        for c in cols:
            if c not in df.columns:
                raise ValueError(
                    f"MetroMedian: column '{c}' missing from X "
                    f"(metro_key={kind!r} requires {cols})"
                )

        parts = [df[c].astype(str).fillna("").str.strip() for c in cols]
        out = parts[0]
        for p in parts[1:]:
            out = out.str.cat(p, sep="|")
        return np.where(out.values == "", "Unknown", out.values)

    def fit(
        self, X: pd.DataFrame, y: pd.DataFrame, *, seed: int = 42
    ) -> "MetroMedian":
        if not isinstance(X, pd.DataFrame):
            raise TypeError(
                f"MetroMedian.fit: expected DataFrame X; got {type(X).__name__}"
            )
        if not isinstance(y, pd.DataFrame):
            raise TypeError(
                f"MetroMedian.fit: expected DataFrame y; got {type(y).__name__}"
            )

        # Align X and y on address (y is the canonical labels frame).
        if "address" not in X.columns or "address" not in y.columns:
            raise ValueError(
                "MetroMedian.fit: both X and y must contain 'address'."
            )
        df = X.merge(
            y[["address", *[c for c in ("rent", "price") if c in y.columns]]],
            on="address",
            how="inner",
        )
        if df.empty:
            raise RuntimeError(
                "MetroMedian.fit: zero rows after joining X with y on address."
            )

        metro = self._metro_key(df)
        df = df.assign(_metro=metro)

        if "rent" in df.columns:
            rv = pd.to_numeric(df["rent"], errors="coerce")
            r = pd.DataFrame({"metro": df["_metro"].values, "rent": rv.values}).dropna()
            if not r.empty:
                self._metro_rent_median = r.groupby("metro")["rent"].median().to_dict()
                self._global_rent_median = float(r["rent"].median())
                self._has_rent = True
        if "price" in df.columns:
            pv = pd.to_numeric(df["price"], errors="coerce")
            p = pd.DataFrame({"metro": df["_metro"].values, "price": pv.values}).dropna()
            if not p.empty:
                self._metro_price_median = p.groupby("metro")["price"].median().to_dict()
                self._global_price_median = float(p["price"].median())
                self._has_price = True
        self._fitted = True
        return self

    def predict(self, X: pd.DataFrame) -> pd.DataFrame:
        if not self._fitted:
            raise RuntimeError("MetroMedian: call .fit(X, y) before .predict().")
        if not isinstance(X, pd.DataFrame):
            raise TypeError(
                f"MetroMedian.predict: expected DataFrame; got {type(X).__name__}"
            )
        if "address" not in X.columns:
            raise ValueError("MetroMedian.predict: X missing 'address'.")

        metro = self._metro_key(X)
        out = pd.DataFrame({"address": X["address"].astype(str).values})
        if self._has_rent:
            out["pred_rent"] = [
                self._metro_rent_median.get(m, self._global_rent_median) for m in metro
            ]
        else:
            out["pred_rent"] = np.nan
        if self._has_price:
            out["pred_price"] = [
                self._metro_price_median.get(m, self._global_price_median) for m in metro
            ]
        else:
            out["pred_price"] = np.nan
        return out

    def lib_versions(self) -> dict[str, str]:
        return _naive_lib_versions()