GC Mean Reversion Band

Results

Backtested performance (illustrative)

Testing detail

• Backtest window: Jul 8, 2024 to Jul 8, 2026 (1 hour bars, the full 730 day window Yahoo Finance provides at that granularity).
• Cost model: 5 bps round trip transaction cost (spread and slippage), applied on every position change. No commission modeled separately.
• Sizing convention: Full notional when in a position, flat otherwise. No leverage, no volatility targeting.
• Redundancy check: Not checked against Lemma's other signals in this run, see validation transparency below.
• Trade concentration: Top 5 trades contributed 21 percent of total PnL out of 96 trades.

How to read this research

• Metrics are computed from the same simulated backtested performance shown below, so the headline numbers and the chart always agree with each other.
• Net return is a plain percent over the fixed backtest window for this horizon, not an annualized figure.
• The Risk Officer note below is part of the research, not a footnote. Read it before acting on this idea.

Risk Officer

Validation transparency

• Data window: Jul 8, 2024 to Jul 8, 2026.
• Mechanical mean reversion rule simulated on the fixed bar size and window for this horizon, with parameters fixed before each backtest. 5 bps round trip transaction cost assumed. 18 parameter variants across the full setup template library for this horizon were backtested, and the strongest performer by profit factor and net return is presented.
• News reviewed: Gold consolidating in a multi month range as rate cut odds stay in flux.; No major central bank meetings scheduled in the final week of the backtest window.

Not checked

• Tested on recent, in sample data only (730 days).
• Limited market regime coverage.
• Multiple candidate setups and parameter variants were backtested; presenting the strongest performer carries some risk of overfitting to this specific window.
• No out of sample test.
• Past performance does not predict future results.

Script handoff

Approved dossiers export as a standalone Python script that downloads the same instrument at the same bar size and window, and mechanically replicates the rule above. Below is an excerpt, not the full script.

import numpy as np
import pandas as pd
import yfinance as yf
from datetime import datetime, timedelta, timezone

TICKER = "GC=F"
INTERVAL = "1h"  # bar size actually backtested
WINDOW_DAYS = 730  # fixed lookback window for this horizon
COST_BPS = 5  # round trip transaction cost assumption, spread and slippage
SESSION_FLATTEN = False  # force flat at the last bar of each day
MAX_HOLD_DAYS = 3  # force flat once a position spans more than this many days


def build_signal(df: pd.DataFrame, lookback: int = 40, entry_z: float = 2.0, exit_z: float = 0.5) -> pd.Series:
    """Fades dislocations back toward a rolling mean; holds until the z score reverts."""
    prices = df["Close"]
    mean = prices.rolling(lookback).mean()
    std = prices.rolling(lookback).std()
    z = (prices - mean) / std.replace(0, np.nan)
    position = pd.Series(0.0, index=prices.index)
    in_position = False
    for i in range(len(prices)):
        zi = z.iloc[i]
        if pd.isna(zi):
            continue
        dislocated = zi <= -entry_z
        reverted = zi >= -exit_z
        if not in_position and dislocated:
            in_position = True
        elif in_position and reverted:
            in_position = False
        position.iloc[i] = 1.0 if in_position else 0.0
    return position


# load_data(), apply_holding_rules(), backtest(), and summarize() ship with
# the full script on export. Preview only.