Go Hybrid: Mean-Reversion Framework PLUS Technical Bias Filtering
Pairs trading remains one of the most widely used market-neutral strategies, built on the principle of exploiting mean reversion in a spread constructed from two related assets. However, traditional implementations—typically based on cointegration tests and z-score entry/exit thresholds—face well-documented limitations, including instability across regimes and declining profitability in modern markets (Tenyakov & Mamon, 2017). So in this blog post we suggest our own pair trading strategy enhancement: combine cointegration and z-score based trading signals with refined technical trading indicators to create a Relative Edge indicator.
Recent research suggests that augmenting classical pairs trading with additional filters and conditional signals can materially improve risk-adjusted performance. In particular, filtering approaches that remove low-quality trades and incorporate additional state information have been shown to significantly enhance performance (Ti et al., 2024).
More broadly, the literature shows that pure mean-reversion signals are insufficient on their own, and that hybrid approaches incorporating additional information outperform static strategies (Kwon & Satchell, 2022).
Enhancing a Classical Pairs Trading Framework
A natural extension is to combine statistical mean reversion signals with technical state filters that help determine when a spread is more likely to revert successfully. Our average trade window is about 20 days.
As Jared Mann would have said “Read the Price Charts!” OK, but we want the software to do that accurately and automatically.
This idea is supported by empirical evidence showing that technical analysis applied to spreads retains statistically significant predictive power, even after controlling for data-snooping bias (Psaradellis et al., 2018).
The key question therefore becomes:
How can we systematically and automatically incorporate technical information into a cointegration-based framework without introducing unnecessary complexity or overfitting?
Four Core Technical Dimensions
We propose evaluating each leg of a pair across four independent technical dimensions for the next 20 trading days:
- Trend direction
- Momentum acceleration
- Trend strength
- Volume confirmation
These dimensions capture complementary aspects of price behaviour and help assess whether a security is more likely to continue its trajectory or revert, aligning with broader findings that combining momentum and mean reversion signals improves performance under varying market conditions (Kwon & Satchell, 2022).
A Practical “20-Day Bias Engine”
As a starting point, we developed a simple and computationally efficient implementation that can be constructed using the following indicator stack:
- Supertrend (10,3) → Trend direction
- MACD histogram slope (12,26,9) → Momentum acceleration
- ADX + DI (14) → Trend strength
- On-Balance Volume (OBV) → Volume confirmation
These indicators are selected to minimise overlap while capturing distinct and relevant dimensions of market behaviour.
Bias Scoring System
Each asset is assigned a score between −4 and +4, representing its expected directional bias over the next ~20 trading days.
- +3 to +4 → Strong bullish bias
- +2 → Moderate bullish bias
- 0 to +1 → Neutral / weak bias
- Negative values → Bearish bias
Each indicator contributes +1, 0, or −1, producing a transparent and interpretable composite score.
Relative Edge: Pair-Level Signal Enhancement
To translate individual signals into a pair-level decision, we compute:
Relative Edge = Bias_A − Bias_B
Then, we can look to define a further edge by filteringg signals based on a compound new indicator with built-in rules such as:
- |Relative Edge| ≥ 2.5, to ensure there is sufficient divergence in relative technical strength
- Weak leg score ≤ +1, to avoid shorting a stock with a long bias
- Both legs have |score| ≥ 1, to avoid taking trades in the “chop” where volatility is low and there is insufficient movement to bring meaningful (profitable) mean reversion to fruition.
A trade is only executed when:
The technical bias aligns with the direction implied by the z-score signal and the compound indictor rules are met.
This effectively introduces a conviction filter, consistent with research showing that filtering low-probability trades improves overall strategy performance (Ti et al., 2024).
Example
Consider a +3.0sd short KO/long PEP PairTrade Finder® signal, but with the additional of the following Relative Edge based on the above:
- KO bias = −4
- PEP bias = +4
- Absolute Relative Edge = 8
This would represent the highest-conviction setup, where both statistical mispricing and technical positioning are aligned.
Limitations of Standard Technical Implementations
While intuitive, practical and likely quite familiar to retail traders, our initial framework still exhibits several structural limitations:
- Indicators are lagging transformations of price and volume
- Fixed parameters fail to adapt to changing volatility regimes
- Scores are not cross-sectionally comparable
- Equal weighting assumes uniform predictive power
These limitations are consistent with broader findings in the literature, where parameter instability and regime sensitivity are major drivers of performance decay in pairs trading systems (Tenyakov & Mamon, 2017).
Toward Institutional-Grade Pair Trading Strategy Enhancements
The retail framework above provides a practical and intuitive “20-day bias engine,” but each component is inherently scale-dependent and sensitive to volatility regime changes. As a result, while the signals are directionally useful, they are not directly comparable across assets or stable across time. So we are proposing the following enhancements to solve these limitations. The stack should still remain computationally efficient with high-grade retail hardware:
Supertrend (10,3) → Normalized Trend Slope (NTS)
The Supertrend indicator provides a binary or step-like representation of trend direction, but it is highly parameter-sensitive and dependent on absolute price volatility. In the institutional version, this is replaced with a 20-day linear regression slope divided by ATR(20), then z-scored over a 252-day rolling window.
This transformation replaces a rule-based trend flip with a continuous, volatility-adjusted measure of trend intensity. Dividing by ATR ensures that trend strength is evaluated relative to market noise, while z-scoring makes the signal comparable across assets and time periods.
MACD Histogram Slope (12,26,9) → Normalized Momentum Z-Score (NMZ)
MACD histogram slope captures momentum acceleration but remains anchored to price scale and can behave inconsistently across different volatility regimes. The institutional replacement reframes momentum as a standardized deviation from its own historical distribution.
Rather than measuring raw acceleration, NMZ expresses how extreme current momentum is relative to its own history. This removes asset-specific scaling effects and ensures that momentum signals reflect statistical rarity rather than absolute movement size.
ADX + DI (14) → Volatility-Adjusted Trend Strength (VATS)
ADX measures trend strength effectively but does not account for whether the underlying price environment is stable or noise-driven. The institutional version retains ADX as the core trend-strength engine but adjusts it using the ATR percentile rank:
ADX × (1 − ATR percentile rank)
This modification introduces a regime filter: when volatility is elevated relative to history, trend strength is discounted, reflecting the reduced reliability of directional signals in unstable environments. The result is a trend strength measure that is context-aware rather than purely directional.
On-Balance Volume (OBV) → Normalized OBV Flow (NOBF)
OBV aggregates cumulative buying and selling pressure but is path-dependent and difficult to compare across assets with different liquidity profiles or trading histories. The institutional transformation converts OBV into a z-scored signal over a rolling window.
This shifts the focus from absolute volume accumulation to unusualness of flow relative to historical norms. It allows the model to detect statistically significant accumulation or distribution events rather than raw volume trends, improving cross-asset comparability.
Together, these transformations preserve the original four-factor structure but elevate it into a statistically consistent framework. Each signal is redefined as either:
- a volatility-normalized measure (NTS, VATS), or
- a distribution-relative measure (NMZ, NOBF)
This transformation ensures that signals are comparable across instruments, resilient to regime changes, and more suitable for systematic portfolio construction. The approach aligns with literature in statistical arbitrage that emphasises normalization, volatility scaling, and adaptive filtering as key drivers of robustness in multi-asset signal design (Tenyakov & Mamon, 2017).
Integrating Technical Bias with Statistical Signals
Importantly, this technical layer does not replace the core mean-reversion engine.
Instead, it acts as a conviction filter and trade selector:
- Z-score identifies statistical mispricing
- Bias stack evaluates directional conviction
- Only trades satisfying both conditions are executed
This approach aligns with broader evidence that multi-signal and hybrid strategies outperform single-factor approaches, particularly by reducing false positives and improving trade quality (Kwon & Satchell, 2022).
Implementation in PairTrade Finder®
These enhancements form the basis of the next evolution of the platform.
Planned features include:
- A compound Relative Edge score for every signal
- Visual indicators showing alignment with trade signal direction, e.g. use a Green/Amber/Red colour with the score inside
- Watchlist integration for real-time monitoring
- Early exit triggers based on collapsing conviction
The objective is to preserve the simplicity of pairs trading while improving its robustness in modern markets.
Final Thought
Pairs trading remains a powerful framework—but naïve implementations are increasingly ineffective.
The edge no longer lies solely in identifying mean reversion, but in answering a more important question:
When is mean reversion most likely to succeed?
Happy Trading.
Geoff S.T. Hossie, CMT
📚 References
- Kwon, O.K. & Satchell, S. (2022) When does pairs trading outperform cross-sectional momentum? Journal of Risk and Financial Management.
- Psaradellis, I., Laws, J., Pantelous, A. & Sermpinis, G. (2018) Pairs trading, technical analysis and data snooping: Mean reversion vs momentum. Global Commodities Applied Research Digest.
- Ti, Y.-W., Dai, T.-S., Wang, K.-L., Chang, H.-H. & Sun, Y.-J. (2024) Improving cointegration-based pairs trading strategy with asymptotic analyses and convergence rate filters. Computational Economics.
- Tenyakov, A. & Mamon, R. (2017) A computing platform for pairs-trading via a blended Kalman-HMM filtering approach. Journal of Big Data.
