sanguo_vnpy_v2/sanguo_factor/tears_data.py

"""Tears 分层序列序列化(方案A tears 端点数据源).

create_full_tear_sheet 的 matplotlib 图,本质是"已算好的序列被画成图"——
本模块把同一批 alphalens performance 序列(日度IC / 分组日度收益 / 因子秩自相关)
直接序列化成前端 ECharts 可渲染的 dict,替代 iframe 内嵌浅色 HTML 报告
(用户 2026-08-29 拍板方案A:原生重构)。
"""
import pandas as pd

# 与 analyzer.py 同款幂等补丁:alphalens-reloaded 的 demean_forward_returns 用
# groupby.transform(lambda) 在 pandas2 崩,独立 import 本模块(单测路径)时
# analyzer 可能未加载 → 补丁不在 → mean_return_by_quantile 崩。重复 patch 无害。
try:
    import alphalens.utils as _al_utils

    def _demean_forward_returns_pandas2(factor_data, grouper=None):
        factor_data = factor_data.copy()
        if not grouper:
            grouper = factor_data.index.get_level_values("date")
        cols = _al_utils.get_forward_returns_columns(factor_data.columns)
        means = factor_data.groupby(grouper)[cols].transform("mean")
        factor_data[cols] = factor_data[cols] - means
        return factor_data

    _al_utils.demean_forward_returns = _demean_forward_returns_pandas2
except ImportError:
    pass

# 同款幂等补丁:alphalens quantize_factor 的 pd.qcut(x, q) 不传 duplicates
# (默认 'raise')——秩类因子(cs_rank/ts_rank 族,如 alpha16)截面值高度并列,
# 分位边界必重合 → ValueError: Bin edges must be unique → 被 no_raise 吞成
# 全 NaN → binning 丢 100% → merged 空 → IC 全空 + tears 崩(NAS 2026-08-30
# 实锤:alpha16 303/303 日期全抛,中位日 25 值仅 6 唯一)。改默认 duplicates='drop'
# 后并列值并入少数分位(quantile_keys 变短,前端已按实际 keys 动态渲染)。
_orig_qcut = pd.qcut


def _qcut_dup_drop(*args, **kwargs):
    kwargs.setdefault("duplicates", "drop")
    return _orig_qcut(*args, **kwargs)


if not getattr(pd.qcut, "_sanguo_dup_drop", False):
    _qcut_dup_drop._sanguo_dup_drop = True
    pd.qcut = _qcut_dup_drop

_TRADING_DAYS = 252


def _nav(daily: pd.Series) -> pd.Series:
    """日度收益(可含重叠窗口)→ 累计净值,首值 1;NaN 日视为空仓(0 收益)."""
    return (1.0 + daily.fillna(0.0)).cumprod()


def _max_drawdown(nav: pd.Series) -> float:
    """净值序列最大回撤(≤0)."""
    v = (nav / nav.cummax() - 1.0).min()
    return float(v) if pd.notna(v) else 0.0


def _annualized(daily: pd.Series, period: int) -> float:
    """日均 × 252 / period:线性去重叠年化,1/5/10D 三周期可比(1D≈eval 链路口径)."""
    s = daily.dropna()
    if s.empty:
        return 0.0
    return float(s.mean() * _TRADING_DAYS / period)


def _monthly_ic(ic: pd.Series) -> list[dict]:
    """日度 IC 按月聚合(月度柱/热力图数据),口径同 sanguo_factor.metrics.monthly_ic."""
    s = ic.dropna()
    if s.empty:
        return []
    g = s.groupby(s.index.to_period("M")).mean()
    return [{"month": t.strftime("%Y-%m"), "ic": round(float(v), 6)} for t, v in g.items()]


def build_tears_data(merged_data, periods: tuple = (1, 5, 10)) -> dict:
    """alphalens factor_data → tears 序列 dict.

    每序列自带日期轴(ic_dates/nav_dates);IC/分组收益均调 alphalens 原函数,
    与 tearsheet 同源。quantile_nav 给全 10 组,前端按需画 Q1/Q5/Q10。
    """
    from alphalens.performance import (
        factor_information_coefficient,
        factor_rank_autocorrelation,
        mean_return_by_quantile,
    )

    ic_df = factor_information_coefficient(merged_data)
    qr_by_date, _ = mean_return_by_quantile(merged_data, by_date=True, demeaned=True)

    # 因子秩自相关(1D,指标条一项):失败不致命 → None
    try:
        ac = factor_rank_autocorrelation(merged_data, period=1).dropna()
        factor_autocorr = round(float(ac.mean()), 4) if len(ac) else None
    except Exception:
        factor_autocorr = None

    out: dict = {"factor_autocorr": factor_autocorr, "periods": {}}
    for p in periods:
        # 列名兼容:alphalens-reloaded 生成 "1D",老版纯数字 "1"(同 analyzer 口径)
        col = next((c for c in ic_df.columns if c in (f"{p}D", str(p))), None)
        if col is None:
            continue

        ic = ic_df[col].dropna()
        qd = qr_by_date[col].unstack("factor_quantile")  # date × quantile
        qkeys = sorted(int(q) for q in qd.columns)

        quantile_nav: dict = {}
        quantile_annual: dict = {}
        for q in qkeys:
            daily_q = qd[q]
            quantile_nav[str(q)] = [round(float(v), 6) for v in _nav(daily_q)]
            quantile_annual[str(q)] = round(_annualized(daily_q, p), 4)

        # 多空 = 最高分位 − 最低分位(每日,demeaned 超额口径同 tearsheet)。
        # 离散因子 duplicates='drop' 后极端形态可能仅 1 组:多空退化为零序列。
        if len(qkeys) >= 2:
            ls_daily = qd[qkeys[-1]] - qd[qkeys[0]]
        elif qkeys:
            ls_daily = qd[qkeys[0]] * 0.0
        else:
            ls_daily = pd.Series(dtype=float)
        ls_nav = _nav(ls_daily)

        n = len(ic)
        m = float(ic.mean()) if n else 0.0
        sd = float(ic.std()) if n > 1 else 0.0
        out["periods"][f"{p}D"] = {
            "count": int(n),
            "ic_mean": round(m, 6) if n else None,
            "ic_std": round(sd, 6) if n > 1 else None,
            "icir": round(m / sd, 4) if sd > 0 else None,
            "t_stat": round(m / (sd / n ** 0.5), 4) if sd > 0 and n > 1 else None,
            "win_rate": round(float((ic > 0).mean()), 4) if n else None,
            "ic_dates": [d.strftime("%Y-%m-%d") for d in ic.index],
            "ic_values": [round(float(v), 6) for v in ic],
            "monthly_ic": _monthly_ic(ic),
            "quantile_keys": [str(q) for q in qkeys],
            "nav_dates": [d.strftime("%Y-%m-%d") for d in qd.index],
            "quantile_nav": quantile_nav,
            "quantile_annual": quantile_annual,
            "ls_nav": [round(float(v), 6) for v in ls_nav],
            "ls_annual": round(_annualized(ls_daily, p), 4),
            "ls_max_dd": round(_max_drawdown(ls_nav), 4),
        }
    return out