Backtesting Mean Reversion Strategies: Essential Metrics and Timeframe Selection
Mean reversion trading rests on the statistical premise that asset prices oscillate around a long-term equilibrium and tend to return to that mean after deviating from it. Backtesting this family of strategies requires more nuance than testing trend-following systems, because mean reversion edges are often small, regime-dependent, and dangerously sensitive to transaction costs and parameter drift. The two pillars that determine whether a backtest reflects reality are the metrics chosen to evaluate performance and the timeframes selected for signal generation and execution. Getting either wrong can turn a losing system into an apparent winner on paper.
Core Metrics Beyond Net Profit
Net profit alone is meaningless without context. The first metric every mean reversion backtest should report is the profit factor—gross profits divided by gross losses. Values above 1.5 are generally acceptable, but mean reversion strategies with profit factors below 1.3 often collapse once realistic slippage is applied. Pair this with win rate and average win/loss ratio; mean reversion typically produces high win rates (60–80%) with small average wins and occasional large losses. A high win rate can mask catastrophic tail risk, so the win/loss ratio must be examined alongside the maximum single loss.
Maximum drawdown and drawdown duration are equally critical. Mean reversion systems often experience prolonged drawdowns when the market enters a trending regime. Reporting maximum drawdown in percentage and in absolute currency terms, plus the longest recovery period, reveals whether the strategy is psychologically and financially survivable. The Calmar ratio (annualized return divided by maximum drawdown) and the MAR ratio (same calculation but using the full backtest period) help compare strategies across different timeframes.
Sharpe, Sortino, and Omega ratios adjust returns for risk. The Sharpe ratio penalizes all volatility, but mean reversion strategies often have negatively skewed return distributions—many small gains and rare large losses. The Sortino ratio, which only penalizes downside deviation, is more appropriate for these strategies. The Omega ratio goes further by considering the entire return distribution and is especially useful when returns are non-normal.
Tail risk metrics deserve dedicated attention. Value at Risk (VaR) at 95% and 99% confidence levels and Conditional VaR (CVaR) quantify the worst expected losses. Mean reversion strategies frequently exhibit fat left tails; ignoring CVaR can lead to overleveraging. Ulcer index measures the depth and duration of drawdowns, offering a more nuanced view than maximum drawdown alone.
Trade-level statistics include average holding period, number of trades, and profit per trade after costs. A strategy with a 0.2% edge per trade will be destroyed by 0.1% slippage and 0.05% commission. Turnover and cost sensitivity analysis—re-running the backtest with doubled transaction costs—should be standard practice.
Statistical significance must be established. The t-statistic of average trade returns, Monte Carlo permutations of trade order, and bootstrap confidence intervals for Sharpe and profit factor help distinguish skill from luck. A mean reversion strategy with 200 trades and a t-stat of 1.2 is not statistically robust.
Timeframe Selection: Signal, Execution, and Regime
Timeframe selection in mean reversion backtesting operates on three levels: the signal timeframe (e.g., 5-minute, hourly, daily), the execution timeframe (tick or 1-minute for entries/exits), and the regime timeframe (weekly or monthly for filtering). Confusing these leads to look-ahead bias and unrealistic fills.
Intraday timeframes (1-minute to 15-minute) offer more trades and faster mean reversion, but they are dominated by microstructure noise, bid-ask bounce, and high transaction costs. Backtesting a 1-minute mean reversion strategy on OHLC data without tick-level bid-ask spreads will overstate returns by 30–70%. These timeframes require order book data and realistic latency assumptions. They also suffer from non-stationarity; a strategy that works in the first hour of trading may fail in the last hour.
Hourly to 4-hour timeframes balance statistical significance with tradability. They capture intraday overreactions while avoiding the worst microstructure noise. However, they are sensitive to session boundaries (e.g., Asian, London, New York) and overnight gaps. Backtests must align signal generation with the correct session and account for rollover costs in forex or futures.
Daily timeframes are the most common for mean reversion in equities and ETFs. Daily closing prices reduce noise, and transaction costs are a smaller fraction of the edge. However, daily mean reversion strategies often rely on overnight holds, exposing the trader to gap risk. Backtesting must include dividends, splits, and borrow costs for short positions. The lookback period for the mean (e.g., 20-day simple moving average) and the entry threshold (e.g., 2 standard deviations) must be optimized on a rolling basis to avoid curve fitting.
Weekly and monthly timeframes produce very few trades, making statistical significance difficult to achieve. They are better suited for portfolio-level mean reversion (e.g., pairs trading across sectors) where the law of large numbers applies across many instruments. Backtesting these requires survivorship-bias-free data and point-in-time fundamentals if the mean is based on valuation ratios.
Regime filters are not optional. Mean reversion fails during strong trends, regime shifts, and liquidity crises. Backtests should include a trend filter (e.g., 200-day moving average slope), a volatility filter (e.g., VIX or ATR percentile), and a liquidity filter (e.g., minimum average daily volume). The Hurst exponent or variance ratio test can quantify whether a timeframe exhibits mean-reverting or trending behavior. A strategy that works on the 15-minute timeframe during low-volatility regimes may invert during high-volatility regimes.
Data frequency and bar construction matter. Time bars (e.g., 1-hour candles) are convenient but can distort mean reversion signals because they sample irregularly in volume and volatility. Tick bars, volume bars, and dollar bars produce more stationary statistical properties and often improve backtest realism. Renko and range bars filter noise but introduce path dependency that is difficult to replicate in live trading.
Walk-forward and out-of-sample testing are mandatory. A mean reversion strategy optimized on 2010–2015 data must be tested on 2016–2020 data with no parameter changes. Rolling window optimization (e.g., 6-month in-sample, 3-month out-of-sample) reveals parameter stability. Combinatorial purged cross-validation with embargo periods prevents leakage from overlapping labels.
Slippage and market impact models must match the timeframe. For daily strategies, a fixed slippage of 0.05% per side is often reasonable for liquid large-caps. For intraday strategies, slippage should be modeled as a function of spread, order size relative to average volume, and volatility. Limit order fill assumptions are particularly dangerous; a backtest that assumes all limit orders fill at the mean will fail in live trading when the market runs away.
Benchmarking against a simple buy-and-hold or random entry strategy with the same holding period ensures the mean reversion logic adds value. Risk-adjusted alpha relative to factors (market, size, value, momentum) isolates the true mean reversion premium. Turnover-adjusted returns prevent favoring high-frequency strategies that generate illusory profits before costs.
Parameter sensitivity heatmaps should show Sharpe or profit factor across a grid of lookback periods and entry thresholds. A robust strategy shows a broad plateau, not a sharp peak. Monte Carlo resampling of trade returns with replacement generates confidence intervals for drawdown and terminal wealth. Stress testing against historical crises (2008, 2020, 2022) reveals tail behavior that average metrics hide.
Execution timeframe alignment ensures that signals generated on one timeframe are executed on a finer timeframe without look-ahead. For example, a daily mean reversion signal should be executed at the next day’s open, not the same day’s close. Time zone alignment for global instruments prevents mixing signals from different sessions. Holiday and half-day handling avoids phantom trades.
Machine learning enhancements to mean reversion—such as Ornstein-Uhlenbeck parameter estimation, Kalman filters for dynamic means, and regime-switching models—require even more careful backtesting. Overfitting risk grows exponentially with model complexity. Purged k-fold cross-validation and probability of backtest overfitting (PBO) metrics become essential.
Reporting standards for mean reversion backtests should include: equity curve with drawdown overlay, monthly return table, trade duration histogram, rolling Sharpe (12-month), rolling correlation to benchmarks, and a full cost sensitivity table. Out-of-sample degradation (in-sample Sharpe minus out-of-sample Sharpe) should be reported; values above 1.0 indicate overfitting.
Timeframe selection for pairs trading and statistical arbitrage differs from single-asset mean reversion. Pairs require cointegration testing on the same timeframe used for signal generation. The half-life of mean reversion (from an Ornstein-Uhlenbeck fit) dictates the optimal holding period and therefore the minimum timeframe. A half-life of 5 days cannot be traded on a 1-minute timeframe profitably after costs. Rolling cointegration with a 60-day window and a 5-day trading window is common, but the backtest must include re-estimation frequency and parameter drift.
Intraday seasonality affects mean reversion. The first 30 minutes and last 30 minutes of the trading day exhibit different mean-reverting properties than midday. Backtests should either avoid these periods or model them separately. Overnight vs. intraday mean reversion are distinct phenomena; mixing them in one backtest without separation produces misleading results.
Volume and volatility filters on the signal timeframe improve robustness. For example, only take mean reversion trades when volume is above its 20-period average and ATR is below its 80th percentile. Backtesting these filters requires conditional performance attribution—reporting metrics for filtered vs. unfiltered trades.
Short-selling constraints and borrow availability are critical for equity mean reversion. A backtest that shorts hard-to-borrow stocks without borrow fees and recall risk is unrealistic. Locate fees (0.1–3% annualized) and short squeeze risk must be modeled. For crypto, funding rates on perpetual swaps can erase mean reversion edges.
Latency and order types matter for intraday mean reversion. A strategy that assumes market orders fill at the signal price will fail. Limit order backtesting requires modeling queue position and adverse selection. Stop-loss and take-profit levels must be tested with intrabar price paths, not just OHLC, because a bar that touches both stop and target can produce different results depending on path.
Final metric checklist for any mean reversion backtest: net profit, profit factor, Sharpe, Sortino, Calmar, maximum drawdown, drawdown duration, Ulcer index, CVaR 95%, average trade, average holding period, number of trades, turnover, cost sensitivity, t-stat of average trade, Monte Carlo p-value, out-of-sample Sharpe, parameter sensitivity plateau width, and regime-conditional performance. Timeframe checklist: signal timeframe justified by half-life, execution timeframe finer than signal, regime timeframe for filters, session alignment, gap handling, bar construction method, walk-forward window, and stress test periods. Without both checklists, a mean reversion backtest is a hypothesis, not evidence.
