Run momentum when the market is trending; run mean reversion when it's range-bound. That's the whole verdict, but the useful part is how you measure "trending" versus "range-bound" without guessing.
A workable regime filter combines three numbers, utilizing insights from the best trend indicators for smarter trading to identify market momentum effectively. When ADX reads above 25 and Average True Range sits above its 60th percentile over the past year, you're in trending conditions and momentum has the edge. When ADX drops below 20, ATR is compressed, and the Hurst exponent falls below 0.45, price is behaving like a mean-reverting series and Bollinger Band or z-score setups tend to outperform. The middle zone, where signals conflict, is where most strategies bleed money through overtrading.
Momentum typically wins with a lower win rate but larger average winners, since it holds through trending continuation. Mean reversion often flips that, with higher win rates but smaller average gains per trade because it exits at the mean, not at extremes.
- ADX >25 + ATR above 60th percentile → favor momentum
- ADX <20 + Hurst <0.45 → favor mean reversion
- Conflicting readings → reduce size or sit out
- Always validate with walk-forward testing and realistic transaction costs before sizing real capital
Key Takeaways
Regime detection, not conviction, should decide whether a trader runs momentum or mean reversion on any given position.
| Point | Details |
|---|---|
| Use numeric regime filters | Combine ADX, ATR percentile, and Hurst exponent rather than trading on instinct about trend direction. |
| Match strategy to payoff shape | Expect momentum's low win rate and large winners versus mean reversion's high win rate and small gains. |
| Validate stationarity before pairs trades | Run an ADF and cointegration test on any spread before assuming it reverts to a mean. |
| Add hysteresis to regime switches | Require several bars of confirmation before reallocating between sleeves to avoid whipsaw costs. |
| Backtest with real frictions | Include commissions, slippage, and borrow costs, then validate with walk-forward testing before sizing capital. |
Table of Contents
- Momentum vs Mean Reversion: Quick Comparison
- Why Momentum and Mean Reversion Are Different Bets
- Indicator Stacks: What Actually Goes Into Each Signal
- How to Detect Trending vs Ranging Regimes
- Backtesting Rules That Separate Real Edges From Noise
- Combining Momentum and Mean Reversion in One Portfolio
- Stops, Sizing, and Execution: Making Signals Tradable
- Two Working Recipes: Momentum and Mean Reversion Pseudocode
- What Actually Matters Once You Get Past the Theory
- Sources
Momentum vs Mean Reversion: Quick Comparison
Choosing between the two isn't really about which strategy is "better." It's about which one fits the regime you're in, the timeframe you can commit to, and how much your broker's commission structure punishes frequent trading. A trader paying $0.65 per contract on low-volume names should think hard before running a mean reversion strategy that fires ten signals a week.
| Dimension | Momentum | Mean Reversion |
|---|---|---|
| Best for / regime | Trending markets, ADX >25 | Ranging markets, ADX <20, low Hurst |
| Win rate & payoff | Lower win rate, larger average winner | Higher win rate, smaller average gain |
| Holding period | Weeks to months | Hours to days |
| Turnover & cost sensitivity | Lower turnover, less cost-sensitive | High turnover, very cost-sensitive |
| Key indicators | Trailing returns, ADX, cross-sectional rank | Bollinger Bands, RSI, z-score, ADF test |
| Main failure mode | Momentum crashes during sharp reversals | Trading a broken mean during regime shift |
Read the turnover row first if you're retail. Mean reversion's edge often gets eaten alive by commissions and slippage on anything traded more than a few times a week, while momentum's lower trade frequency tolerates a wider cost basis. A trader facing $1 to $2 in round-trip friction per contract, common at many retail brokers on options, usually does better sticking to momentum setups on daily or weekly bars than chasing mean reversion signals that need dozens of round trips to pay off.
The holding period row matters just as much. If you can only check positions once a day, mean reversion's intraday z-score signals will decay before you act on them. Momentum's slower cadence forgives a slower reaction time.
Why Momentum and Mean Reversion Are Different Bets
Momentum and mean reversion aren't two flavors of the same idea. They rest on opposite statistical assumptions about how prices behave after a move.
- Momentum bets on persistence. It assumes recent winners keep winning because information diffuses slowly through a market, order flow clusters, and institutional buying pressure takes weeks to fully play out. That's positive autocorrelation in returns, and it's why momentum strategies tend to buy strength rather than dips.
- Mean reversion bets on stationarity. It assumes a price series (or a spread between two related instruments) has a stable long-run average, and that deviations from that average are temporary noise rather than the start of a new trend. This only holds while the underlying mean itself doesn't shift, which is exactly the failure mode documented in mean reversion research).
- The P&L shapes look almost like mirror images. Momentum produces a right-skewed distribution: frequent small losses, occasional large winners that carry the whole strategy. Mean reversion produces something closer to a left-skewed distribution: frequent small wins, occasional large losses when the "temporary" dislocation turns out to be permanent.
The practical consequence shows up in drawdown behavior. Momentum drawdowns tend to be gradual, a string of whipsaw losses during choppy transitions between trends. Mean reversion drawdowns tend to be sudden and severe, a single blown-up trade when a spread that "always" reverts simply doesn't. That's the tail risk lurking behind every high win rate mean reversion system, and it's also why position sizing has to differ between the two: momentum can tolerate looser stops because the thesis is patience, while mean reversion needs tighter stops because the thesis breaks catastrophically, not gradually.
Correlation with the broader market regime differs too. Momentum strategies tend to correlate with trend strength across the whole market, so multiple momentum positions often lose money at the same time during a regime flip. Mean reversion positions, especially pairs trades built on cointegrated instruments, aim for market neutrality, though that neutrality is only as good as the cointegration relationship holding up.
Indicator Stacks: What Actually Goes Into Each Signal
Concrete indicator recipes beat vague strategy descriptions, so here's what a functioning stack looks like for each side.
Mean reversion stack:
- Bollinger Bands set at a 20-period moving average with 2 standard deviation bands, which capture roughly 95% of price action under normal conditions, flagging the remaining 5% as statistical extremes worth fading
- RSI below 30 or above 70 as a secondary confirmation filter, reducing false signals from Bollinger Band touches alone
- A rolling z-score of price or spread relative to its trailing mean, sized so entries only trigger past 1.5 to 2 standard deviations
- Half-life estimation from an AR(1) fit on the spread, which tells you how long a reversion typically takes and directly informs position holding time
- For pairs specifically, a cointegration test and ADF check before trusting that the spread reverts at all, since correlation between two tickers tells you nothing about whether their price relationship is actually stationary
Momentum stack:
- Trailing returns over a 12-week lookback, a standard window that balances signal strength against noise
- Cross-sectional ranking, comparing an instrument's return against a universe of peers rather than judging it in isolation
- Percentage-change lookbacks at multiple horizons (4-week, 12-week, 26-week) blended into a composite score
- A smoothing filter, often a short exponential moving average, to reduce whipsaw from single-day outliers skewing the ranking
Pro Tip: Test your lookback windows across at least two distinct market cycles before trusting them. A 12-week momentum lookback tuned entirely on a single bull run will look brilliant in-sample and fall apart the first time the market chops sideways for two quarters.
Parameter choice isn't cosmetic. A 20 period Bollinger Band and a 50 period one produce meaningfully different trade frequency and drawdown profiles on the same instrument, and the only way to know which fits your universe is cross-validating across multiple out-of-sample windows rather than picking whatever number backtested best once.
How to Detect Trending vs Ranging Regimes
Regime detection is the hinge the entire momentum vs mean reversion decision swings on, and it needs to run on numbers, not gut feel.
Three metrics do most of the work:
- ADX (Average Directional Index): readings above 25 typically signal a trending market where directional strategies get paid; readings below 20 usually mean range-bound conditions better suited to reversion trades
- ATR percentile: comparing current Average True Range against its trailing distribution tells you whether volatility is expanding (often accompanies trend initiation) or contracting (often precedes range-bound chop)
- Hurst exponent: a value above 0.5 suggests trending, persistent behavior; a value below 0.5, and especially below roughly 0.45, suggests mean-reverting behavior in the series
No single metric is reliable alone. ADX lags at trend inflection points, ATR percentile says nothing about direction, and the Hurst exponent needs a long enough window to estimate reliably, which makes it slow to react to fresh regime changes.
The fix is combining them into either a majority-rule vote (two out of three metrics agreeing before switching sleeves) or a weighted composite score that blends all three into a single number with a threshold. Either method should include hysteresis: require the regime signal to hold for a minimum number of bars, say five to ten, before acting, rather than flipping strategies every time ADX ticks across 25. Without hysteresis, you end up whipsawing between momentum and mean reversion allocations exactly when both are performing worst.
The other trap is lookahead bias in the regime calculation itself. Every one of these metrics has to be computed only on data available at the time of the trading decision, using a rolling window that ends at the prior bar's close, never on a window that peeks into future price action to "confirm" a regime after the fact. That single mistake inflates backtested performance more than almost any other coding error in this kind of system.
Backtesting Rules That Separate Real Edges From Noise
A backtest that ignores costs is a work of fiction. Building a credible test of either strategy means running through a specific checklist before trusting any equity curve.
- Model realistic transaction costs. Include commissions, bid-ask spread as slippage, and for short positions in mean reversion pairs, the borrow cost on the shorted leg. A mean reversion strategy that looks profitable at zero cost frequently turns negative once you apply even modest per-trade friction, because its edge per trade is small to begin with.
- Run walk-forward validation, not a single in-sample fit. Split the data into rolling training and testing windows, re-optimizing parameters only on the training segment and scoring performance strictly on the untouched segment that follows. This is the single best defense against overfitting a lookback period to one lucky stretch of history.
- Check parameter stability across neighboring settings. If a 20-day z-score lookback works but 18-day and 22-day both collapse, that's a sign you found a coincidence, not an edge.
- Run Monte Carlo resampling on trade sequences to see the realistic distribution of drawdowns rather than trusting the single path your backtest happened to produce.
- Test for stationarity directly. An ADF test on the price series or spread, combined with a cointegration test for pairs, tells you whether the mean reversion assumption is statistically defensible rather than a pattern you're imposing on random noise. Estimate the half-life from that same series to set realistic holding periods and stop distances.
A strategy that survives walk-forward testing and still shows a stable half-life estimate across multiple out-of-sample windows has cleared a bar that eliminates most retail backtests before they ever reach a live account.
Skipping any one of these steps doesn't just weaken the result. It usually means you're measuring the noise in your own historical sample rather than a repeatable market inefficiency, and the failure only becomes visible after the strategy is already running with real money behind it.
Combining Momentum and Mean Reversion in One Portfolio
Running both sleeves at once, rather than picking one permanently, is usually the more robust design, provided the switching logic is disciplined.
- Regime-switch allocation with hysteresis and caps. Allocate capital to whichever sleeve the regime filter favors, but cap the maximum swing per rebalance (say, no more than a 30 percentage point shift in allocation in one period) so a noisy regime read doesn't whipsaw the whole portfolio.
- Residualization. Strip the trend component out of a price series first, using a rolling regression against a benchmark or a longer moving average, then trade the leftover residual as a mean-reverting series. This lets you run mean reversion logic even on an asset that's trending overall, because you're only trading the noise around that trend, not fighting the trend itself.
- Risk-parity blending. Weight the momentum and mean reversion sleeves by inverse volatility rather than equal dollar allocation, then rebalance on a fixed cadence, monthly is common, so neither sleeve's recent volatility spike dominates total portfolio risk.
Hysteresis deserves a second mention here specifically because portfolio-level switching amplifies the cost of getting it wrong. Flipping the entire book's allocation on a single noisy ADX reading doesn't just cost you a bad trade. It costs you transaction costs on unwinding one sleeve and building the other, twice, if the signal reverses again the following week. Building in a minimum dwell time before any allocation shift is the cheapest insurance in the entire framework.
Stops, Sizing, and Execution: Making Signals Tradable
A signal without an execution plan isn't a strategy, it's an opinion. Converting momentum and mean reversion signals into positions that actually protect capital takes different mechanics for each.
- Use ATR-based trailing stops for momentum. Since momentum positions can run for weeks, a fixed-dollar stop either gets hit too early on normal noise or doesn't adjust as volatility expands. An ATR multiple, commonly 2x to 3x the 14-day ATR, trails the position and lets winners run while still capping the downside.
- Use fixed, tighter stops for mean reversion. Because the entire thesis is that price reverts quickly, a reversion trade that keeps moving against you past a modest threshold is telling you the mean has shifted, not that you need to wait longer. A stop at roughly 1.5x to 2x the average holding period's typical adverse excursion works better than a trailing stop here.
- Scale position size to volatility, not dollar amount. Sizing every position identically regardless of the instrument's volatility means your riskiest positions dominate portfolio drawdown. Normalize size by recent ATR so each position contributes roughly equal risk.
- Slice larger orders with TWAP or VWAP execution rather than market orders, especially in less liquid names, to avoid moving the price against yourself on entry.
- Build an automated kill-switch. If the regime filter and the strategy signal disagree for more than a set number of consecutive bars, say the regime reads "ranging" while a momentum position is still open and losing, that mismatch should trigger a size reduction or full exit rather than waiting for a manual review.
Pro Tip: Set your kill-switch threshold before you go live, not after a bad week. Traders who wait to define "how wrong is too wrong" usually end up rationalizing the very drawdown the rule was supposed to prevent.
Reviewing options trade ideas with a consistent checklist before entry catches most of these sizing and stop mistakes before they cost real money, rather than after.
Two Working Recipes: Momentum and Mean Reversion Pseudocode
Concrete parameters beat abstract descriptions, so here are two compact recipes worth coding up and sanity-checking against your own data before scaling either one.
- Momentum sleeve: rank a universe by 12-week trailing return, go long the top decile and short (or avoid) the bottom decile, scale each position's size by the inverse of its 20-day ATR, apply a 2.5x ATR trailing stop, and rebalance the whole ranking monthly.
- Mean reversion sleeve: compute a 5-day rolling z-score on a cointegrated spread (validated first with an ADF test), enter when the z-score crosses beyond 2 standard deviations, size the position using the spread's estimated half-life from an AR(1) fit, and apply a fixed stop at roughly 3 standard deviations to protect against the mean itself shifting.
- Sanity checks before trusting either backtest: confirm the momentum sleeve's win rate lands meaningfully below 50%, often in the 35% to 45% range, with average winners several multiples larger than average losers. Confirm the mean reversion sleeve's win rate lands well above 55%, often 60% to 70%, with a much smaller average R per trade and a maximum drawdown concentrated in a handful of outlier trades rather than spread evenly across the sample.
If either sleeve's diagnostic metrics look inverted from this pattern, a high win rate momentum system or a low win rate mean reversion system, that's usually a sign of a lookahead bias in the regime filter or the entry logic, not a genuinely different edge.
NumPy's rolling statistics functions handle the z-score, ATR, and moving average calculations behind both recipes without needing custom implementations for basic statistical operations, which keeps the modeling focus on the strategy logic itself rather than infrastructure.
What Actually Matters Once You Get Past the Theory
The conventional treatment of momentum vs mean reversion spends too much time on which philosophy is "correct" and not enough on the fact that both are correct, in their own regime, and wrong everywhere else. That framing shift changes what you should spend your time building. It's not a better momentum indicator or a cleverer z-score. It's a regime filter you actually trust enough to act on without second-guessing it mid-trade.
Most retail traders skip the boring part, hysteresis, walk-forward validation, half-life estimation, and jump straight to indicator shopping. That's backward. A mediocre indicator stack on a well-validated regime filter beats a brilliant indicator stack that fires signals in the wrong market condition half the time.
If you take one thing from this piece, take the discipline of computing regime metrics on available data only, then giving the signal a few bars to prove itself before switching sleeves. That single habit prevents more damage than any parameter tweak to RSI or ADX ever will.
— Customer