code/macrolens/_fast.py

"""Fast array-based data access for model training.

Provides pre-materialized numpy arrays and PyTorch-compatible datasets
that bypass the slow per-sample DataFrame slicing of WhatIfTSFDataset.
"""

from __future__ import annotations

import logging
from typing import Any

import numpy as np
import pandas as pd

from ._compat import config

logger = logging.getLogger(__name__)


# ── Feature catalogue ────────────────────────────────────────────────────

_FEATURE_GROUPS: dict[str, list[str]] = {}  # lazily built


def _build_feature_groups(granularity: str = "daily") -> dict[str, list[str]]:
    """Build feature group catalogue from the panel columns."""
    bench_dir = config.get_benchmark_dir(granularity)
    panel_path = bench_dir / "panel_train.parquet"
    if not panel_path.exists():
        return {}

    # Read just the column names (no data)
    cols = pd.read_parquet(panel_path, columns=None).columns.tolist()  # type: ignore[arg-type]

    groups: dict[str, list[str]] = {
        "price": [],
        "fundamentals": [],
        "macro": [],
        "derived": [],
        "other": [],
    }

    for c in cols:
        if c in ("ticker", "date", "label", "split", "sector", "industry",
                  "nearest_filing_type", "nearest_filing_date", "nearest_filing_path"):
            continue
        if c in ("open", "high", "low", "close", "volume", "adj_close"):
            groups["price"].append(c)
        elif c.startswith("stmt_"):
            groups["fundamentals"].append(c)
        elif c.startswith(("fred_", "eia_")):
            groups["macro"].append(c)
        elif c.startswith("derived_"):
            groups["derived"].append(c)
        else:
            groups["other"].append(c)

    return {k: sorted(v) for k, v in groups.items() if v}


def features(granularity: str = "daily", verbose: bool = True) -> dict[str, list[str]]:
    """List available features grouped by category.

    Parameters
    ----------
    granularity : str
        ``"daily"`` (default).
    verbose : bool
        If True, print a summary table.

    Returns
    -------
    dict[str, list[str]]
        Mapping of group name to feature column names.

    Example
    -------
    >>> groups = macrolens.features()
    Feature groups (daily):
      price          :   6 features  [open, high, low, close, volume, adj_close]
      fundamentals   :  22 features  [stmt_revenue, stmt_net_income, ...]
      macro          :  80 features  [fred_DFF, fred_DGS10, ...]
      derived        :  15 features  [derived_market_cap, derived_pe, ...]
    Total: 123 features
    """
    groups = _build_feature_groups(granularity)

    if verbose:
        total = sum(len(v) for v in groups.values())
        print(f"\nFeature groups ({granularity}):")
        for name, cols in groups.items():
            preview = cols[:3]
            more = f", ... +{len(cols)-3}" if len(cols) > 3 else ""
            print(f"  {name:<16}: {len(cols):>3} features  [{', '.join(preview)}{more}]")
        print(f"  Total: {total} features\n")

    return groups


# ── Fast numpy array materializer ─────────────────────────────────────────

def to_arrays(
    split: str = "train",
    horizon: int = 21,
    lookback: int | None = None,
    granularity: str = "daily",
    target_col: str = "close",
    max_instances: int | None = None,
    feature_cols: list[str] | None = None,
    seed: int = 42,
) -> tuple[np.ndarray, np.ndarray, list[str]]:
    """Materialize benchmark data as numpy arrays for fast training.

    Pre-builds all sliding windows into contiguous arrays. Much faster
    than iterating ``WhatIfTSFDataset`` for model training.

    Parameters
    ----------
    split : str
        ``"train"`` or ``"test"``.
    horizon : int
        Forecast horizon (default: 21).
    lookback : int, optional
        Lookback window (default: 63 for daily).
    granularity : str
        ``"daily"`` (default).
    target_col : str
        Target column (default: ``"close"``).
    max_instances : int, optional
        Subsample to this many instances (random). Default: all.
    feature_cols : list[str], optional
        Subset of feature columns. Default: all numeric columns.
    seed : int
        Random seed for subsampling.

    Returns
    -------
    X : np.ndarray
        Shape ``(N, lookback, n_features)`` float32.
    y : np.ndarray
        Shape ``(N, horizon)`` float32.
    feature_names : list[str]
        Column names corresponding to X's last axis.

    Example
    -------
    >>> X_train, y_train, features = macrolens.to_arrays("train", horizon=21)
    >>> X_train.shape
    (50000, 63, 103)
    >>> y_train.shape
    (50000, 21)

    >>> # For sklearn:
    >>> X_flat = X_train.reshape(len(X_train), -1)
    >>> model.fit(X_flat, y_train[:, -1])
    """
    if lookback is None:
        lookback = config.get_lookback_windows(granularity)[0]

    bench_dir = config.get_benchmark_dir(granularity)
    panel_path = bench_dir / f"panel_{split}.parquet"
    if not panel_path.exists():
        raise FileNotFoundError(f"Panel not found at {panel_path}.")

    logger.info("Loading panel %s...", panel_path)
    panel = pd.read_parquet(panel_path)
    panel["date"] = pd.to_datetime(panel["date"])
    panel = panel.sort_values(["ticker", "date"]).reset_index(drop=True)

    # Determine feature columns
    exclude = {"ticker", "date", "label", "split", "sector", "industry",
               "nearest_filing_type", "nearest_filing_date", "nearest_filing_path"}
    if feature_cols is None:
        feature_cols = [
            c for c in panel.columns
            if c not in exclude and panel[c].dtype.kind in "fiub"
        ]
    feat_names = sorted(feature_cols)

    if target_col not in panel.columns:
        raise ValueError(f"target_col={target_col!r} not in panel columns.")

    # Get target column index for extraction
    target_idx = panel.columns.get_loc(target_col)

    required_len = lookback + horizon
    feat_idx = [panel.columns.get_loc(c) for c in feat_names]
    n_features = len(feat_names)

    # ── Pass 1: count available windows per ticker ──
    ticker_info: list[tuple[str, int, int]] = []  # (ticker, group_start, n_windows)
    total_windows = 0
    for ticker, grp in panel.groupby("ticker", sort=False):
        n = len(grp)
        if n < required_len:
            continue
        n_windows = n - required_len + 1
        ticker_info.append((ticker, grp.index[0], n_windows))
        total_windows += n_windows

    logger.info("Total available windows: %d across %d tickers",
                total_windows, len(ticker_info))

    # ── Decide how many windows to take per ticker ──
    rng = np.random.RandomState(seed)
    budget = max_instances if max_instances is not None else total_windows

    if budget >= total_windows:
        # Take all windows
        per_ticker_take = {t: nw for t, _, nw in ticker_info}
    else:
        # Proportional sampling per ticker (at least 1 if selected)
        per_ticker_take: dict[str, int] = {}
        remaining = budget
        for t, _, nw in ticker_info:
            alloc = max(1, int(round(nw / total_windows * budget)))
            alloc = min(alloc, nw, remaining)
            if alloc > 0:
                per_ticker_take[t] = alloc
                remaining -= alloc
            if remaining <= 0:
                break
        # Distribute leftover budget
        if remaining > 0:
            for t, _, nw in ticker_info:
                if t not in per_ticker_take:
                    continue
                extra = min(remaining, nw - per_ticker_take[t])
                if extra > 0:
                    per_ticker_take[t] += extra
                    remaining -= extra
                if remaining <= 0:
                    break

    actual_n = sum(per_ticker_take.values())

    # ── Pass 2: pre-allocate and fill ──
    X = np.empty((actual_n, lookback, n_features), dtype=np.float32)
    y = np.empty((actual_n, horizon), dtype=np.float32)
    offset = 0

    panel_vals = panel.values

    for ticker, grp_start, n_windows in ticker_info:
        take = per_ticker_take.get(ticker, 0)
        if take <= 0:
            continue

        # Slice this ticker's data from the panel (index is 0-based after reset_index)
        grp_len = n_windows + required_len - 1
        vals = panel_vals[grp_start : grp_start + grp_len]

        feats = vals[:, feat_idx].astype(np.float32)
        targets = vals[:, target_idx].astype(np.float32)

        # Select which windows to extract
        if take >= n_windows:
            chosen = np.arange(n_windows)
        else:
            chosen = rng.choice(n_windows, take, replace=False)
            chosen.sort()

        # Extract windows for chosen starts
        for i, start in enumerate(chosen):
            X[offset + i] = feats[start : start + lookback]
            y[offset + i] = targets[start + lookback : start + required_len]

        offset += len(chosen)

    # Trim in case of rounding
    X = X[:offset]
    y = y[:offset]

    # Handle NaN
    np.nan_to_num(X, copy=False, nan=0.0)
    np.nan_to_num(y, copy=False, nan=0.0)

    logger.info("Built %d instances: X=%s, y=%s", len(X), X.shape, y.shape)

    return X, y, feat_names


# ── PyTorch Dataset wrapper ───────────────────────────────────────────────

class TSFTorchDataset:
    """PyTorch-compatible dataset backed by pre-materialized numpy arrays.

    Works directly with ``torch.utils.data.DataLoader`` — no custom
    collate function needed.

    Parameters
    ----------
    X : np.ndarray
        Shape ``(N, lookback, n_features)`` float32.
    y : np.ndarray
        Shape ``(N, horizon)`` float32.

    Example
    -------
    >>> X, y, _ = macrolens.to_arrays("train", horizon=21, max_instances=50000)
    >>> ds = macrolens.TSFTorchDataset(X, y)
    >>> dl = DataLoader(ds, batch_size=256, shuffle=True, num_workers=4)
    >>> for x_batch, y_batch in dl:
    ...     pred = model(x_batch)  # (256, 63, 103) -> (256, 21)
    """

    def __init__(self, X: np.ndarray, y: np.ndarray) -> None:
        self.X = X
        self.y = y

    def __len__(self) -> int:
        return len(self.X)

    def __getitem__(self, idx: int) -> tuple:
        import torch

        return (
            torch.from_numpy(self.X[idx]),
            torch.from_numpy(self.y[idx]),
        )


def load_torch(
    split: str = "train",
    horizon: int = 21,
    lookback: int | None = None,
    granularity: str = "daily",
    target_col: str = "close",
    max_instances: int | None = None,
    **kwargs: Any,
) -> TSFTorchDataset:
    """Load a PyTorch-compatible TSF dataset for fast training.

    This materializes the data as numpy arrays, then wraps them in a
    ``TSFTorchDataset`` that returns ``(x_tensor, y_tensor)`` tuples
    compatible with ``torch.utils.data.DataLoader``.

    Parameters
    ----------
    split : str
        ``"train"`` or ``"test"``.
    horizon : int
        Forecast horizon (default: 21).
    lookback : int, optional
        Lookback window (default: 63).
    max_instances : int, optional
        Subsample to this many instances. Recommended for prototyping.
    **kwargs
        Passed to :func:`to_arrays`.

    Returns
    -------
    TSFTorchDataset
        PyTorch-compatible dataset.

    Example
    -------
    >>> from torch.utils.data import DataLoader
    >>> ds = macrolens.load_torch("train", horizon=21, max_instances=50000)
    >>> dl = DataLoader(ds, batch_size=256, shuffle=True, num_workers=4)
    >>> for x, y in dl:
    ...     print(x.shape, y.shape)  # (256, 63, 103) (256, 21)
    ...     break
    """
    X, y, _ = to_arrays(
        split=split,
        horizon=horizon,
        lookback=lookback,
        granularity=granularity,
        target_col=target_col,
        max_instances=max_instances,
        **kwargs,
    )
    return TSFTorchDataset(X, y)