Fading the VWAP bands on the S&P 500

Abstract

We backtested the VWAP mean reversion rules on Micro E-mini S&P 500 (MES) 5-minute bars from 2016-01-11 to 2026-10-30: 4,784 trades, with commission and a tick of slippage on every market and stop fill. In sample (to 2023-08-03) it made −0.34R per trade with a profit factor of 0.62; out of sample (from 2023-08-04) −0.26R per trade and 0.68 over 1,431 trades. 2 of the 25 parameter settings we tried were profitable out of sample. Sized at $250 of risk a trade, simulated 50K evaluations passed 1% to 3% of the time across 36 tracked plans.

1Hypothesis

When price closes more than two standard deviations away from the day's VWAP, bet on a return to VWAP. The stop is one more standard deviation away.

If intraday stretches away from VWAP are mostly order-flow noise, a close beyond the 2-sigma band should revert to VWAP more often than it runs one more sigma, enough to pay for costs.

2Data

Instrument

Micro E-mini S&P 500 (MES), traded on the ES price series (ES.c.0: front month, rolled on the calendar). CME Globex via Databento.

Period

2016-01-11 to 2026-10-30: 2,746 sessions, 1,056,820 one-minute bars, built into 210,816 5-minute bars of the regular session (09:30 to 16:00 ET).

Costs

$0.75 commission a contract each way and 1 tick ($1.25) of slippage on every market and stop fill. Limit orders fill when price touches them; market orders fill at the next bar's open and stops at their price, each with one tick of slippage; when a bar touches both the stop and the target, the stop counts.

Position size

About $250 of risk a trade: micro contracts, rounded down, at least one. R is a trade's net P&L divided by its risk at entry.

Out of sample

The last 30% of weekdays: 2023-08-04 on (846 days), after 1,974 in-sample days. The rules were fixed before either side was run.

3Method

Our reading of the publicly described rules; the creator may trade it differently. The rules, exactly as the code runs them:

1. VWAP and its bands restart at 09:30 ET. VWAP is the volume-weighted average of each 5-minute bar's typical price (high + low + close) / 3; sigma is the volume-weighted standard deviation of the same prices around it.
2. Entry: a 5-minute bar that CLOSES below VWAP - band_sd (2) sigma buys at the next bar's open; one that closes above VWAP + 2 sigma sells. Signals only from 10:00 to 15:00 ET, so the bands have half an hour of data first.
3. Stop: stop_sd (1) sigma beyond the signal bar's close.
4. Target: VWAP as it stood at the signal.
5. Exit: the stop, the target, or flat at 15:55 ET. At most max_trades_per_day (2) trades a day, one at a time.
6. Size: about risk_usd dollars of risk a trade (micro contracts, rounded down, at least 1 and at most max_contracts).

1"""VWAP mean reversion with standard-deviation bands, on MES 5-minute bars, regular session.2 3Our reading of the commonly taught rules; traders who teach it may trade it differently.4 5Rules:6  1. VWAP and its bands restart at 09:30 ET. VWAP is the volume-weighted average of each 5-minute7     bar's typical price (high + low + close) / 3; sigma is the volume-weighted standard deviation of8     the same prices around it.9  2. Entry: a 5-minute bar that CLOSES below VWAP - `band_sd` (2) sigma buys at the next bar's open;10     one that closes above VWAP + 2 sigma sells. Signals only from 10:00 to 15:00 ET, so the bands11     have half an hour of data first.12  3. Stop: `stop_sd` (1) sigma beyond the signal bar's close.13  4. Target: VWAP as it stood at the signal.14  5. Exit: the stop, the target, or flat at 15:55 ET. At most `max_trades_per_day` (2) trades a day,15     one at a time.16  6. Size: about `risk_usd` dollars of risk a trade (micro contracts, rounded down, at least 1 and at17     most `max_contracts`).18"""19import math20 21import qy22 23 24class Bot(qy.Strategy):25    """Fade a close outside the 2-sigma VWAP band, target VWAP."""26 27    symbol = "MES"28    timeframe = "5m"29    session = "rth"30    params = {"band_sd": 2.0, "stop_sd": 1.0, "max_trades_per_day": 2, "risk_usd": 250, "max_contracts": 50}31    max_position = 5032    entry_window = ("10:00", "15:00")33    flatten_time = "15:55"34    overnight = False35 36    def setup(self):37        self.day = None38 39    def _size(self, risk_points):40        per_contract = risk_points * self.point_value41        n = int(self.p.risk_usd // per_contract) if per_contract > 0 else 142        return max(1, min(self.p.max_contracts, n))43 44    def on_bar(self, bar):45        if self.day != bar.date:46            self.day = bar.date47            self.sv = self.svp = self.svp2 = 0.048            self.open_px = bar.open49        tp = (bar.high + bar.low + bar.close) / 350        v = bar.volume or 0.051        self.sv += v52        self.svp += v * tp53        self.svp2 += v * tp * tp54        if self.sv <= 0:55            return56        vwap = self.svp / self.sv57        sigma = math.sqrt(max(0.0, self.svp2 / self.sv - vwap * vwap))58        if sigma < self.tick_size or not self.position.is_flat or self.orders:59            return60        signal_at = bar.dt.hour * 60 + bar.dt.minute + 5  # the bar's close, when the order goes in61        if self.trades_today >= self.p.max_trades_per_day or not 600 <= signal_at < 900:62            return63        close = bar.close64        stretch = (close - vwap) / sigma65        move_pct = abs(close - self.open_px) / self.open_px * 10066        risk = self.p.stop_sd * sigma67        if stretch <= -self.p.band_sd:68            self.buy(self._size(risk), stop_price=close - risk, target_price=vwap,69                     tag=f"VWAP fade long | f={move_pct:.4f}")70        elif stretch >= self.p.band_sd:71            self.sell(self._size(risk), stop_price=close + risk, target_price=vwap,72                      tag=f"VWAP fade short | f={move_pct:.4f}")

4Results

5Robustness

We reran the strategy on a grid of 25 settings of its two main parameters, each on the full history with the same split. 1 were profitable in sample and 2 out of sample.

6Prop-firm implications

The backtest's sessions run through the rules of 36 tracked 50K evaluations: 2,000 runs of up to 60 sessions each, with the drawdown and daily limits checked against each trade's intraday low. Intervals are Wilson 95% intervals of the simulation. A simulation of the past, not the chance that anyone passes.

7Limitations

• Bands are built from 5-minute bars, not ticks.
• Mean reversion loses on trend days, and the stop's distance moves with the day's volatility.
• One market (MES).
• A backtest on the past, with fills simulated from one-minute bars; live fills, queues and outages differ.

8Conclusion

No edge after costs: profit factor 0.68 out of sample. Nothing on the grid clearly helps. Its in-sample best, entry band (sigma) 3.5 and stop beyond the close (sigma) 2, did better out of sample, profit factor 1.12 against 0.68 for the defaults, but over only 35 trades, t = 0.5: too few, or too noisy, to call it a fix. On this reading of the rules, with these costs, the historical record gives no reason to expect a profit.

9References

1. Berkowitz, S. A., Logue, D. E., and Noser, E. A. (1988). The Total Cost of Transactions on the NYSE. Journal of Finance, 43(1), 97-112.