Backtesting the Gold Macro Scorecard: Does the Signal Deliver?

为什么要对黄金宏观成绩单进行反测试?

在同行文章中 采用宏观数据预测黄金价格我们建立了一个综合宏观评分卡,将指向信号分配给六个美国宏观指标 TIPS 10Y实际收益率,平衡通胀率,美联储政策利率,Fed总资产,M2货币供应量和贸易加权美元,并将它们汇总成净黄金偏差.

这篇文章通过对 每日 黄金价格从FXMacroData 商品终点我们将在每次宏观数据发布时计算分数表, 基于净信号, 持有金币的简单长/平仓,

步骤1:获取每日黄金价格和宏观系列

后台测试的基础是FXMacroData的每日黄金价格. 商品/黄金 终点 LBMA PM 固定价格以美元为特洛伊司.与月度或每周的汇总数据不同,每日价格使我们能够准确衡量每个宏观信号转换的影响.

import requests
import pandas as pd
from datetime import date

BASE = "https://fxmacrodata.com/api/v1"
KEY  = "YOUR_API_KEY"

def get_series(path: str, start: str = "2020-01-01") -> pd.DataFrame:
    """Fetch a time series and return as a DataFrame with date index."""
    r = requests.get(f"{BASE}{path}", params={"api_key": KEY, "start_date": start})
    r.raise_for_status()
    data = r.json().get("data", [])
    df = pd.DataFrame(data)
    if not df.empty:
        df["date"] = pd.to_datetime(df["date"])
        df = df.set_index("date").sort_index()
    return df

# Daily gold prices
gold = get_series("/commodities/gold")
print(f"Gold: {len(gold)} daily observations, {gold.index[0].date()} to {gold.index[-1].date()}")
# Gold: ~1350 daily observations, 2020-01-02 to 2026-04-15

接下来,抽取六个给分卡提供料的宏观指标系列.这些都以不同的频率发布,有些每周 (TIPS收益率,平衡),有些每月 (CPI,M2),有些在FOMC日期 (政策利率) ,但每个观察都保持为"当前"值,直到下一个发布.

# Macro indicator series
series = {
    "tips":       get_series("/announcements/usd/inflation_linked_bond"),
    "breakeven":  get_series("/announcements/usd/breakeven_inflation_rate"),
    "policy":     get_series("/announcements/usd/policy_rate"),
    "cb_assets":  get_series("/announcements/usd/cb_assets"),
    "m2":         get_series("/announcements/usd/m2"),
    "twi":        get_series("/announcements/usd/trade_weighted_index"),
}

for name, df in series.items():
    print(f"  {name:12s}: {len(df):4d} obs  ({df.index[0].date()} – {df.index[-1].date()})")

步骤2:排列序列并进行前列填写

将所有宏观系列合并到日期黄金指数中.每个宏观值都被前进填写,从发布日期推向到下一个发布日,因此后验从不使用未来信息.

# Align all series to the daily gold date index
aligned = gold[["val"]].rename(columns={"val": "gold"}).copy()

for name, df in series.items():
    # Reindex to gold dates and forward-fill
    macro = df[["val"]].rename(columns={"val": name})
    macro = macro.reindex(aligned.index, method="ffill")
    aligned = aligned.join(macro)

# Drop rows where any macro series hasn't started yet
aligned = aligned.dropna()
print(f"Aligned dataset: {len(aligned)} trading days")
print(aligned.tail())

步骤3:计算每日分数卡信号

每个交易日,我们从原始文章计算相同的分数卡,但不是比较最近的两个观察,我们将当前的前期填充值与前30天的值进行比较. 这比日复一日噪音更强有力的方向衡量.

LOOKBACK = 30  # calendar days for direction detection

def score_column(col: pd.Series, mode: str) -> pd.Series:
    """Score a macro series: +1 bullish gold, 0 neutral, -1 bearish."""
    prev = col.shift(LOOKBACK)
    change = col - prev

    if mode == "falling":
        return pd.Series(
            [1.0 if c < -0.05 else (-1.0 if c > 0.05 else 0.0) for c in change],
            index=col.index
        )
    elif mode == "rising":
        return pd.Series(
            [1.0 if c > 0.05 else (-1.0 if c < -0.05 else 0.0) for c in change],
            index=col.index
        )
    elif mode == "negative":
        return pd.Series(
            [1.0 if v < 0 else (-1.0 if v > 1.0 else 0.0) for v in col],
            index=col.index
        )
    return pd.Series(0.0, index=col.index)


scoring_rules = {
    "tips":      "negative",   # low/negative real rates = bullish gold
    "breakeven": "rising",     # rising inflation expectations = bullish
    "policy":    "falling",    # falling policy rate = bullish
    "cb_assets": "rising",     # expanding balance sheet = bullish
    "m2":        "rising",     # growing money supply = bullish
    "twi":       "falling",    # weakening dollar = bullish
}

for name, mode in scoring_rules.items():
    aligned[f"sig_{name}"] = score_column(aligned[name], mode)

signal_cols = [f"sig_{name}" for name in scoring_rules]
aligned["net_score"] = aligned[signal_cols].sum(axis=1)

print(aligned[["gold", "net_score"]].tail(10))

第四步:确定交易规则

后台测试使用了一套简单,现实的规则:

• 长信号: 如果净分数卡 ≥ +2,就去黄金交易 (我们有机会看到黄金升值).
• 没有信号: 如果净分数卡 < +2,持有现金 (没有黄金位置).
• 没有空头销售: 宏观指数表确定了黄金的有利模式,但没有同样的信心来产生黄金空期信号.
• 没有杆: 位置是100%的黄金或100%的现金.
• 每天重新平衡: 信号在日结束时进行评估; 位置变化适用于下一天的交易回报.
• 交易成本: 我们每次往返交易 (入口+出口) 扣除5个基点,以计算黄金ETF或期货合约的差距和滑动.

# Trading rules
THRESHOLD = 2.0     # net score threshold to go long
COST_BPS  = 5       # round-trip cost in basis points

# Daily gold returns
aligned["gold_ret"] = aligned["gold"].pct_change()

# Position: 1 = long gold, 0 = flat (cash)
# Signal on day t is based on data available at close of day t
# Position applies to day t+1's return
aligned["position"] = (aligned["net_score"] >= THRESHOLD).astype(float)

# Detect trade events (position changes)
aligned["trade"] = aligned["position"].diff().abs()
aligned.loc[aligned.index[0], "trade"] = 0  # no trade on first day

# Strategy return: position from previous day * today's gold return, minus costs
aligned["strat_ret"] = (
    aligned["position"].shift(1) * aligned["gold_ret"]
    - aligned["trade"].shift(1) * (COST_BPS / 10_000)
)

# Cumulative returns
aligned["gold_cum"]  = (1 + aligned["gold_ret"]).cumprod()
aligned["strat_cum"] = (1 + aligned["strat_ret"].fillna(0)).cumprod()

print(f"Buy-and-hold return: {(aligned['gold_cum'].iloc[-1] - 1) * 100:.1f}%")
print(f"Strategy return:     {(aligned['strat_cum'].iloc[-1] - 1) * 100:.1f}%")

第五步:测量性能

粗积累性回报只是图片的一部分. 风险调整后的指标告诉我们,战略的表现是否来自技能 (计时宏观制度) 或者只是因为承担更多的风险.

import numpy as np

def performance_stats(returns: pd.Series, trades: pd.Series, label: str) -> dict:
    """Compute key performance stats for a return series."""
    total_ret = (1 + returns).prod() - 1
    ann_ret   = (1 + total_ret) ** (252 / len(returns)) - 1
    ann_vol   = returns.std() * np.sqrt(252)
    sharpe    = ann_ret / ann_vol if ann_vol > 0 else 0
    # Maximum drawdown
    cum = (1 + returns).cumprod()
    peak = cum.cummax()
    dd = (cum - peak) / peak
    max_dd = dd.min()
    # Win rate
    invested_days = returns[returns != 0]
    win_rate = (invested_days > 0).mean() if len(invested_days) > 0 else 0
    n_trades = int(trades.sum() / 2)  # round trips

    return {
        "label":        label,
        "total_return":  f"{total_ret * 100:.1f}%",
        "annual_return": f"{ann_ret * 100:.1f}%",
        "annual_vol":    f"{ann_vol * 100:.1f}%",
        "sharpe_ratio":  f"{sharpe:.2f}",
        "max_drawdown":  f"{max_dd * 100:.1f}%",
        "win_rate":      f"{win_rate * 100:.1f}%",
        "trades":        n_trades,
    }

strat_stats = performance_stats(
    aligned["strat_ret"].dropna(),
    aligned["trade"].fillna(0),
    "Macro Scorecard"
)
bnh_stats = performance_stats(
    aligned["gold_ret"].dropna(),
    pd.Series(0, index=aligned.index),
    "Buy & Hold"
)

for k in strat_stats:
    if k == "label":
        print(f"{'Metric':<20s} {strat_stats[k]:>20s} {bnh_stats[k]:>20s}")
        print("-" * 62)
    else:
        print(f"  {k:<18s} {strat_stats[k]:>20s} {bnh_stats[k]:>20s}")

第6步:分析吸收量和信号质量

宏观时间模型中最重要的价值主张不是捕捉每一个上行动作,而是避免最糟糕的下行动向.让我们来看看策略平稳 (黄金不足) 的时期,以及这些时期是否与有意义的下跌相对应.

# Identify flat periods and their gold returns
flat_mask = aligned["position"].shift(1) == 0
flat_gold_ret = aligned.loc[flat_mask, "gold_ret"]
long_gold_ret = aligned.loc[~flat_mask, "gold_ret"]

print(f"Days long gold:       {(~flat_mask).sum()}")
print(f"Days flat (cash):     {flat_mask.sum()}")
print(f"Avg daily ret (long): {long_gold_ret.mean()*100:.3f}%")
print(f"Avg daily ret (flat): {flat_gold_ret.mean()*100:.3f}%")
print(f"Avoided loss days:    {(flat_gold_ret < 0).sum()} "
      f"(total loss: {flat_gold_ret[flat_gold_ret < 0].sum()*100:.1f}%)")

2022年拉回率是最明确的例子.随着美联储从3月至10月份激烈提高利率,2022年,TIPS 10Y收益率从接近零升至1.6%,贸易加权美元大幅升值,M2增长变为负值.分数表正确读出所有三个信号都是下行,并转向现金,避免黄金从峰值到低谷下跌20%左右.

第7步:信号模式的分解

不是所有的分数卡水平都是相同的.按净分数水平分解平均期货黄金回报率,可以发现信号如何区分有利和不利的模式.

# Forward 20-day gold return by score level
aligned["fwd_20d"] = aligned["gold"].pct_change(20).shift(-20)

regime_stats = (
    aligned.groupby("net_score")["fwd_20d"]
    .agg(["mean", "std", "count"])
    .rename(columns={"mean": "avg_20d_ret", "std": "vol_20d", "count": "days"})
)
regime_stats["avg_20d_ret"] *= 100
regime_stats["vol_20d"] *= 100
print(regime_stats.round(2))

第八步:稳定性检查

一个单个后验配置可能过量.在这里,我们通过变化关键参数来检查结果是否脆弱.

极限敏感度

results = []
for thresh in range(-2, 6):
    pos = (aligned["net_score"] >= thresh).astype(float)
    ret = pos.shift(1) * aligned["gold_ret"]
    trades = pos.diff().abs().fillna(0)
    ret -= trades.shift(1) * (COST_BPS / 10_000)
    cum = (1 + ret.fillna(0)).prod()
    vol = ret.std() * np.sqrt(252)
    ann = cum ** (252 / len(ret)) - 1
    sharpe = ann / vol if vol > 0 else 0
    results.append({"threshold": thresh, "total_ret": f"{(cum-1)*100:.1f}%",
                    "sharpe": round(sharpe, 2), "pct_invested": f"{pos.mean()*100:.0f}%"})

pd.DataFrame(results).set_index("threshold")

追溯期敏感性

for lb in [15, 30, 60, 90]:
    # Recompute scores with different lookback
    sig_sum = pd.Series(0.0, index=aligned.index)
    for name, mode in scoring_rules.items():
        prev = aligned[name].shift(lb)
        chg  = aligned[name] - prev
        if mode == "falling":
            sig = pd.Series([1 if c < -0.05 else (-1 if c > 0.05 else 0) for c in chg], index=aligned.index)
        elif mode == "rising":
            sig = pd.Series([1 if c > 0.05 else (-1 if c < -0.05 else 0) for c in chg], index=aligned.index)
        elif mode == "negative":
            sig = pd.Series([1 if v < 0 else (-1 if v > 1 else 0) for v in aligned[name]], index=aligned.index)
        else:
            sig = pd.Series(0, index=aligned.index)
        sig_sum += sig
    pos = (sig_sum >= THRESHOLD).astype(float)
    ret = pos.shift(1) * aligned["gold_ret"]
    cum = (1 + ret.fillna(0)).prod()
    vol = ret.std() * np.sqrt(252)
    ann = cum ** (252/len(ret)) - 1
    print(f"  Lookback {lb:3d}d: return {(cum-1)*100:+.1f}%  Sharpe {ann/vol:.2f}")

第9步:完成后验脚本

这是一个完整的自主后台测试,从FXMacroData获取所有数据,运行分数卡策略,并打印出一个图表准备的输出.

"""
Gold Macro Scorecard Backtest
Fetches daily gold prices and macro series from FXMacroData,
computes the composite scorecard, and evaluates a long/flat strategy.
"""
import requests
import pandas as pd
import numpy as np
from datetime import date

BASE = "https://fxmacrodata.com/api/v1"
KEY  = "YOUR_API_KEY"
START = "2020-01-01"
THRESHOLD = 2
LOOKBACK  = 30
COST_BPS  = 5

def get(path: str) -> pd.DataFrame:
    r = requests.get(f"{BASE}{path}", params={"api_key": KEY, "start_date": START})
    r.raise_for_status()
    df = pd.DataFrame(r.json().get("data", []))
    df["date"] = pd.to_datetime(df["date"])
    return df.set_index("date").sort_index()

# ── Fetch data ──
gold = get("/commodities/gold")[["val"]].rename(columns={"val": "gold"})

macro = {
    "tips":      (get("/announcements/usd/inflation_linked_bond"),  "negative"),
    "breakeven": (get("/announcements/usd/breakeven_inflation_rate"), "rising"),
    "policy":    (get("/announcements/usd/policy_rate"),             "falling"),
    "cb_assets": (get("/announcements/usd/cb_assets"),               "rising"),
    "m2":        (get("/announcements/usd/m2"),                      "rising"),
    "twi":       (get("/announcements/usd/trade_weighted_index"),    "falling"),
}

# ── Align and forward-fill ──
df = gold.copy()
for name, (series, _) in macro.items():
    s = series[["val"]].rename(columns={"val": name})
    df = df.join(s.reindex(df.index, method="ffill"))
df = df.dropna()

# ── Score ──
def score(col, mode):
    prev = col.shift(LOOKBACK)
    chg  = col - prev
    if mode == "negative":
        return col.apply(lambda v: 1 if v < 0 else (-1 if v > 1 else 0)).astype(float)
    if mode == "falling":
        return chg.apply(lambda c: 1 if c < -0.05 else (-1 if c > 0.05 else 0)).astype(float)
    if mode == "rising":
        return chg.apply(lambda c: 1 if c > 0.05 else (-1 if c < -0.05 else 0)).astype(float)
    return pd.Series(0.0, index=col.index)

df["net_score"] = sum(score(df[n], m) for n, (_, m) in macro.items())

# ── Trade ──
df["ret"] = df["gold"].pct_change()
df["pos"] = (df["net_score"] >= THRESHOLD).astype(float)
df["trade"] = df["pos"].diff().abs().fillna(0)
df["strat_ret"] = df["pos"].shift(1) * df["ret"] - df["trade"].shift(1) * (COST_BPS/1e4)
df["gold_cum"]  = (1 + df["ret"].fillna(0)).cumprod()
df["strat_cum"] = (1 + df["strat_ret"].fillna(0)).cumprod()

# ── Report ──
for label, cum_col, ret_col in [("Strategy", "strat_cum", "strat_ret"),
                                  ("Buy&Hold", "gold_cum", "ret")]:
    total = df[cum_col].iloc[-1] - 1
    vol   = df[ret_col].std() * np.sqrt(252)
    ann   = (1 + total) ** (252/len(df)) - 1
    sharpe = ann / vol if vol > 0 else 0
    peak  = df[cum_col].cummax()
    mdd   = ((df[cum_col] - peak) / peak).min()
    print(f"{label:12s}  Return: {total*100:+.1f}%  Sharpe: {sharpe:.2f}  MaxDD: {mdd*100:.1f}%")

print(f"\nDays invested: {df['pos'].mean()*100:.0f}%  |  Round-trips: {int(df['trade'].sum()/2)}")

关键发现与实用教训

限制和警告

• 解释性后台测试. 本文中显示的样本结果使用代表性数据来说明方法.您应该使用自己的密钥对活跃API运行完整的脚本,以产生您喜欢的日期范围内的验证结果.
• 选择指标时存在生存偏差. 我们选择了这六个指标,因为它们对黄金有很强的理论先验,但选择本身是一种隐含曲线适应的形式.真正的样本外测试需要在查看黄金数据之前选择指标.
• 没有位置尺寸. 双边长/平的方法是故意简单的.更复杂的仓位大小 (例如,按净分数大小进行扩展) 可以提高风险调整的回报率,但增加可能过度适用的自由参数.
• 现金收益被忽略. 在平稳期内,该策略的收益为零.在实践中,短期利率在此期间为0.5.5% 包括空置现金的无风险收益率将进一步提高该策划的风险调整边缘.
• 没有期货的交易成本建模. 如果通过黄金期货而不是ETF实施,则适用滚动成本和保证金要求. 假设5个基础点的往返成本代表黄金ETF差距,但可能低估期货执行成本.
• 宏观数据发布延迟 后台测试使用实际发布日期,没有前性.但在实时交易中,数据发布和您的系统处理之间可能会有几个小时.每日重新平衡的节奏使得这对这个策略无关紧要.

在本后台测试中使用的所有数据 日均黄金价格和六个美国宏观指标均可从FXMacroData API中获取. 黄金商品终点 提供每天的LBMA PM固定价格, 美国宏观终点 开始免费试用在 没有任何信息可以提供.现在我们要做什么?