第 25 篇：参数优化与 Walk-Forward 防过拟合

窗口1: 训练 2015-2016 → 测试 2017
窗口2: 训练 2015-2017 → 测试 2018   (或滚动式: 训练 2016-2017 → 测试 2018)
窗口3: 训练 2015-2018 → 测试 2019
...依次前推，把每段测试期的成绩拼接成最终的评估序列

import numpy as np
import pandas as pd

# ---------- 1. 准备数据：akshare 优先，本地造数据 fallback ----------
def get_data():
    try:  # akshare 接口可能随版本变化，请以官方文档为准
        import akshare as ak
        df = ak.stock_zh_a_hist(symbol="600519", period="daily",
                                start_date="20150101", end_date="20241231",
                                adjust="qfq")
        df = df.rename(columns={"日期": "date", "收盘": "close"})
        df["date"] = pd.to_datetime(df["date"])
        return df.set_index("date")[["close"]]
    except Exception:
        rng = np.random.default_rng(42)  # 固定种子，保证可复现
        dates = pd.bdate_range("2015-01-01", "2024-12-31")
        ret = rng.normal(0.0004, 0.015, len(dates))  # 示意用随机收益
        price = 100 * np.exp(np.cumsum(ret))
        return pd.DataFrame({"close": price}, index=dates)

df = get_data()
print(f"数据条数: {len(df)}, 区间: {df.index[0].date()} ~ {df.index[-1].date()}")

# ---------- 2. 双均线向量化回测函数（供网格调用） ----------
def backtest_ma(df, fast, slow, fee=0.001):
    """返回策略日收益序列；信号 shift(1) 避免未来函数"""
    ma_fast = df["close"].rolling(fast).mean()
    ma_slow = df["close"].rolling(slow).mean()
    sig = (ma_fast > ma_slow).astype(int)      # 当日收盘后确定的持仓
    pos = sig.shift(1).fillna(0)               # 次日才执行 → 无未来函数
    ret = df["close"].pct_change().fillna(0)
    turnover = pos.diff().abs().fillna(0)      # 换手 → 双边 0.1% 成本
    strat = pos * ret - turnover * fee
    return strat

# ---------- 3. 网格搜索（训练期） ----------
is_data = df[:"2022-12-31"]  # 样本内：2015~2022
results = []
for fast in range(3, 16, 2):
    for slow in range(20, 121, 10):
        if fast >= slow:
            continue
        s = backtest_ma(is_data, fast, slow)
        n = len(s)
        ann = (1 + s.sum()) ** (252 / n) - 1  # 简化年化（对数近似教学用）
        results.append((fast, slow, ann))

grid = pd.DataFrame(results, columns=["fast", "slow", "ann"])
pivot = grid.pivot(index="fast", columns="slow", values="ann")
best = grid.loc[grid["ann"].idxmax()]
print("样本内最优参数:", int(best["fast"]), "/", int(best["slow"]),
      f"年化(示意): {best['ann']:.1%}")

# ---------- 4. 样本外验证 ----------
oos_data = df["2023-01-01":]  # 样本外：2023~2024
oos = backtest_ma(oos_data, int(best["fast"]), int(best["slow"]))
oos_n = len(oos)
oos_ann = (1 + oos.sum()) ** (252 / oos_n) - 1
print(f"样本外年化(示意): {oos_ann:.1%}")
if oos_ann < 0.02:
    print(">>> 警告：样本外明显弱于样本内，疑似过拟合")

# ---------- 5. Walk-Forward 滚动优化 ----------
def walk_forward(df, train_years=4, test_years=1, fee=0.001):
    years = sorted(set(df.index.year))
    oos_all = []
    for i in range(0, len(years) - train_years - test_years + 1):
        train_y = years[i:i + train_years]
        test_y = years[train_y[-1] + 1:train_y[-1] + test_years]
        tr = df[df.index.year.isin(train_y)]
        te = df[df.index.year.isin(test_y)]
        if len(te) == 0:
            continue
        best_f, best_s, best_v = 5, 60, -np.inf
        for f in range(3, 16, 2):          # 每轮重新选参，只用训练段
            for s in range(20, 121, 10):
                if f >= s:
                    continue
                v = backtest_ma(tr, f, s, fee).sum()
                if v > best_v:
                    best_f, best_s, best_v = f, s, v
        seg = backtest_ma(te, best_f, best_s, fee)  # 用选出的参数跑测试段
        oos_all.append(seg)
        print(f"训练{train_y[0]}-{train_y[-1]} → 测试{test_y[0]}: "
              f"参数{best_f}/{best_s}")
    return pd.concat(oos_all)

wf = walk_forward(df)
n = len(wf)
wf_ann = (1 + wf.sum()) ** (252 / n) - 1
print(f"Walk-Forward 拼接年化(示意): {wf_ann:.1%}")

# ---------- 6. 参数热力图 ----------
import matplotlib.pyplot as plt
plt.rcParams["font.sans-serif"] = ["SimHei"]
plt.rcParams["axes.unicode_minus"] = False

fig, ax = plt.subplots(figsize=(9, 5))
im = ax.imshow(pivot.values, aspect="auto", cmap="RdYlGn")
ax.set_xticks(range(len(pivot.columns)))
ax.set_xticklabels(pivot.columns)
ax.set_yticks(range(len(pivot.index)))
ax.set_yticklabels(pivot.index)
ax.set_xlabel("慢线周期")
ax.set_ylabel("快线周期")
ax.set_title("双均线参数网格年化收益热力图（样本内，示意）")
fig.colorbar(im, label="年化收益")
plt.tight_layout()
plt.savefig("wf_heatmap.png", dpi=120)
plt.show()

示例输出（示意，数值因数据与随机种子而异）:
数据条数: 2437, 区间: 2015-01-05 ~ 2024-12-31
样本内最优参数: 5 / 110 年化(示意): 19.8%
样本外年化(示意): 3.1%
>>> 警告：样本外明显弱于样本内，疑似过拟合
训练2015-2018 → 测试2019: 参数9/100
训练2016-2019 → 测试2020: 参数5/110
训练2017-2020 → 测试2021: 参数13/90
训练2018-2021 → 测试2022: 参数5/110
训练2019-2022 → 测试2023: 参数7/60
Walk-Forward 拼接年化(示意): 5.4%