Mean reversion without catching every falling knife

Mean reversion works best when you’re selective. This guide shows how to frame a mean reversion trading strategy with four concrete filters—regime, extension, liquidity, and invalidation—so you avoid trying to buy every vertical drop or fade every breakout. Each filter maps to a measurable hypothesis you can backtest for stocks and futures. We also include a step‑by‑step build, instrumentation, and journal tags so your results are reproducible.

Mean reversion, as used in research—not just “buying dips”#

Define the setup: signal + regime + extension + liquidity + invalidation#

Signal (residual):

• Construct a stationary residual you can trade. Options include: (a) regression of an asset on market/sector factors with the residual as the series; (b) a pair spread; (c) a PCA factor‑neutralized idiosyncratic return. In practice, many statistical‑arbitrage approaches then model the residual as OU/AR(1) with rolling parameters Statistical Arbitrage in the U.S. Equities Market.
• Standardize to a z‑score using a rolling mean and standard deviation appropriate to your timeframe. OU approaches can estimate the mean and the mean‑reversion speed directly, informing dynamic bands Statistical Arbitrage in the U.S. Equities MarketOptimal Mean-Reverting Spread Trading: Nonlinear Integral Equation Approach.

Regime filter (do not fight a trend):

• Hypothesis: the residual’s reversion is more reliable in low‑trend, lower‑volatility states. Use a regime classifier and only allow entries when that state is active.
• Implementations: a two‑ or three‑state Markov‑switching model on returns or volatility; or a simpler momentum/volatility score with a threshold. Markov‑switching regime segmentation is a well‑established tool in economics and time‑series analysis and is directly applicable here A New Approach to the Economic Analysis of Nonstationary Time Series and the Business.

Extension filter (how far is far enough?):

• Distance rule: require the absolute z‑score to meet or exceed an entry threshold but cap participation if the absolute z‑score is above a maximum. Distance/z‑score methods are the baseline in pairs‑trading research Pairs Trading: Performance of a Relative-Value Arbitrage Rule by Evan Gatev, William N. Goetzmann, K. Geert Rouwenhorst :: SSRN.
• OU‑optimal bands: estimate the reversion speed and volatility of the residual and compute optimal entry/exit bands that account for costs; this avoids arbitrary entry/exit levels Optimal Mean-Reverting Spread Trading: Nonlinear Integral Equation Approach.

Liquidity filter (make fills credible):

• Hypothesis: realized edge is sensitive to trading costs and execution frictions. Require minimum daily dollar volume, cap quoted/effective spread, and down‑weight or skip assets failing these thresholds.
• Rationale: transaction costs and execution frictions are first‑order in statistical‑arbitrage design and should be modeled and constrained in testing Statistical Arbitrage in the U.S. Equities Market.

Invalidation (define “you’re wrong”):

• Price‑based: use a hard stop when the long entry’s z‑score becomes more negative than a stop threshold (and the mirror for shorts).
• Model‑based: use OU‑derived exit/stop regions that incorporate costs; the optimal policy defines where persistence outweighs expected reversion Optimal Mean-Reverting Spread Trading: Nonlinear Integral Equation Approach.
• Time‑based: if reversion hasn’t occurred within a fixed number of bars and the regime or reversion‑speed estimate deteriorates, exit. OU policies shift with the estimated speed Optimal Mean-Reverting Spread Trading: Nonlinear Integral Equation Approach.

Entry, exit, and invalidation hypotheses#

From rules to a test: plan and instrumentation#

1. Choose your universe and bar interval. For equities, define a liquid universe (for example, top ranks by dollar volume). For futures, verify the live contract specifications—multiplier, tick size, and session hours—on the exchange before testing or managing rolls.
2. Construct the residual. Options: factor‑neutral residual (regression or PCA) or a pair spread. Fit a rolling OU/AR(1) or estimate a rolling mean and standard deviation with a robust window Statistical Arbitrage in the U.S. Equities Market.
3. Standardize to a z‑score and precompute realized volatility per asset or residual. Log both raw and winsorized values for diagnostics.
4. Build the regime classifier. A two‑ or three‑state Markov‑switching model on returns or volatility is a concrete, testable option A New Approach to the Economic Analysis of Nonstationary Time Series and the Business. Store state probabilities and transitions; define a binary “allow entries” flag.
5. Specify the extension rule. Start with entry, exit, and maximum‑extension levels for a distance rule aligned with pairs‑trading convention Pairs Trading: Performance of a Relative-Value Arbitrage Rule by Evan Gatev, William N. Goetzmann, K. Geert Rouwenhorst :: SSRN, or compute OU‑optimal bands that incorporate your cost model Optimal Mean-Reverting Spread Trading: Nonlinear Integral Equation Approach.
6. Define liquidity gates. Set minimum daily dollar volume and maximum effective spread thresholds appropriate to your timeframe. Record the value used at each decision to avoid survivorship bias. Tie your size or pass/skip to this gate Statistical Arbitrage in the U.S. Equities Market.
7. Encode invalidation. Implement stop bands (price‑ or model‑based) and a time‑stop. Ensure stops are simulated with intrabar logic consistent with your bar type.
8. Simulate with realistic frictions. Include commissions and slippage consistent with your venue and size; statistical‑arbitrage models emphasize execution costs Statistical Arbitrage in the U.S. Equities Market.
9. Segment by regime and extension. Produce performance broken out by allowed/blocked regimes and by z‑score buckets to test whether the filter actually adds value.
10. Hold out data. Split sample into training, validation, and final test. Freeze parameters before the final test. Report both gross and net results and the fraction of signals filtered out.
11. Stress and sensitivity. Shock the estimated reversion speed, volatility, and cost inputs to see how OU‑band policies and fixed‑z rules behave under parameter error Optimal Mean-Reverting Spread Trading: Nonlinear Integral Equation Approach.

Where the idea can fail#

• Regime misclassification: If your classifier tags a trending phase as mean‑reverting, you will fade sustained moves. Markov‑switching models help, but they are probabilistic and can be wrong at transitions A New Approach to the Economic Analysis of Nonstationary Time Series and the Business.
• Structural breaks in the residual: Relationships that supported stationarity (for example, pair linkage or factor exposures) can change abruptly; extreme z‑scores may reflect a new equilibrium. OU‑optimal policies widen bands under higher volatility or slower reversion speed, but that also reduces trade frequency Optimal Mean-Reverting Spread Trading: Nonlinear Integral Equation Approach.
• Costs and slippage: If you do not model execution frictions realistically, they can consume expected edge. Statistical‑arbitrage research emphasizes modeling costs rather than assuming constant fills Statistical Arbitrage in the U.S. Equities Market.
• Over‑aggressive extension: Very large absolute z‑scores may look attractive but often coincide with regime shifts or breaks; using OU‑derived bands and stop regions helps bound this risk Optimal Mean-Reverting Spread Trading: Nonlinear Integral Equation Approach.

• Strategy family: mean-reversion, stat-arb, pairs
• Signal type: residual-z, OU-band, factor-neutral
• Regime tag: regime-low-trend, regime-trend, regime-transition
• Extension bucket: z-2to3, z-3to4, z-gt4 (customize bins)
• Liquidity tier: liquid-A, liquid-B, illiquid-skip
• Stop type: stop-price, stop-model, stop-time
• Outcome: tp-hit, stop-hit, time-exit, flip-regime
• Futures admin: contract-code, expiry, roll-rule-verified

FAQ#

Is mean reversion just “buying dips”?

Pairs trading and modern stat‑arb research typically trade stationary residuals—spreads or de‑trended series—not discretionary dip‑buys. Many studies standardize the residual and use entry/exit thresholds tied to that series Statistical Arbitrage in the U.S. Equities MarketPairs Trading: Performance of a Relative-Value Arbitrage Rule by Evan Gatev, William N. Goetzmann, K. Geert Rouwenhorst :: SSRN.

Do I need a Markov‑switching model to run the regime filter?

No. Any regime classifier that measurably separates trending from mean‑reverting conditions can work. Markov‑switching models are a well‑established way to represent regime changes in economic and financial time series and are commonly used for such segmentation A New Approach to the Economic Analysis of Nonstationary Time Series and the Business.

How do I pick z‑score thresholds without overfitting?

Distance‑based pairs trading often uses standardized thresholds (for example, enter when the spread’s z‑score exceeds a level and exit near zero) Pairs Trading: Performance of a Relative-Value Arbitrage Rule by Evan Gatev, William N. Goetzmann, K. Geert Rouwenhorst :: SSRN. If you calibrate an OU model, you can compute optimal entry/exit bands that reflect your estimated reversion speed, volatility, and costs, reducing arbitrariness Optimal Mean-Reverting Spread Trading: Nonlinear Integral Equation Approach.

Can I use this on futures as well as equities?

Yes, if your residual is well‑defined (for example, between related contracts or versus a fair‑value basis) and you verify liquidity and contract specifications. Always confirm multipliers, tick sizes, and session hours with the exchange before testing or placing orders.

How should I handle costs and slippage in the backtest?

Backtests should include commissions and slippage. Statistical‑arbitrage frameworks explicitly incorporate cost modeling; ignoring execution frictions can materially overstate net performance Statistical Arbitrage in the U.S. Equities Market.

Why include a time‑stop if I already have price stops?

Time‑stops recognize that reversion probability decays when the estimated mean‑reversion speed is low or the regime shifts. OU‑based policies adjust optimal actions to those parameters; a simple time‑stop is a practical proxy Optimal Mean-Reverting Spread Trading: Nonlinear Integral Equation Approach.

Sources#

• Statistical Arbitrage in the U.S. Equities Market — jeremywhittaker.com
• A New Approach to the Economic Analysis of Nonstationary Time Series and the Business — citeseerx.ist.psu.edu
• Pairs Trading: Performance of a Relative-Value Arbitrage Rule by Evan Gatev, William N. Goetzmann, K. Geert Rouwenhorst :: SSRN — papers.ssrn.com
• Optimal Mean-Reverting Spread Trading: Nonlinear Integral Equation Approach — arxiv.org