Global Biases Fuel U.S. Stock Returns

by

Author: Denis Avetisyan

New research reveals that trading signals derived from the Chinese market continue to predict American stock performance, even when accounting for traditional financial metrics.

A double-selection LASSO approach confirms that behavioral biases identified in one market are transferable and contribute to asset pricing in another.

Despite growing interest in behavioral factors, cross-market validation of short-term trading signals remains surprisingly limited. This study, ‘Cross-Market Alpha: Testing Short-Term Trading Factors in the U.S. Market via Double-Selection LASSO’, rigorously examines the transferability of factors originating from the unique microstructure of the Chinese A-share market to U.S. equities. Utilizing a double-selection LASSO approach, we demonstrate that 17 Alpha191 factors exhibit significant incremental explanatory power for U.S. stock returns, even after controlling for established fundamental characteristics. Do these findings suggest that behavioral biases represent a universal component of asset pricing, and can cross-market factor transferability consistently enhance investment strategies?

Unveiling the Pattern Zoo: Navigating Complexity in Asset Pricing

The landscape of asset pricing has dramatically evolved beyond foundational models like the Three-Factor Model, now populated by a burgeoning collection of variables – often referred to as the ‘Factor Zoo’ – each attempting to explain market anomalies. Initially, these factors aimed to improve predictive power, incorporating elements like momentum, quality, and size beyond simply market risk, value, and size. However, this rapid expansion has sparked concern among researchers, as the addition of numerous factors increases the risk of identifying spurious relationships that appear significant within a dataset but fail to hold up in real-world performance. The sheer volume of potential factors necessitates careful scrutiny, as statistical noise can easily be mistaken for genuine drivers of asset returns, ultimately hindering the development of robust and reliable investment strategies.

The pursuit of comprehensive asset pricing models has led to an increasing number of factors intended to capture subtle market dynamics, yet this expansion carries inherent risks. As models incorporate more variables to explain historical data, they become susceptible to overfitting – essentially memorizing noise rather than identifying true predictive signals. This phenomenon dramatically diminishes the model’s ability to accurately forecast future returns – a critical measure known as out-of-sample performance. Consequently, a more discerning approach is needed, one that prioritizes parsimony and statistical rigor over simply adding more factors, to ensure models generalize effectively and remain useful for investment decisions.

The increasing number of factors used in asset pricing models isn’t simply a matter of discovering more meaningful relationships; it’s fundamentally complicated by the challenges of statistical analysis in high-dimensional spaces. Traditional methods, designed for scenarios with fewer variables than observations, begin to falter as the number of potential factors grows. This leads to inflated statistical significance for factors that may, in reality, be driven by random noise or data mining – a phenomenon known as spurious regression. As the ‘Factor Zoo’ expands, the reliability of any single factor diminishes without employing advanced techniques designed to account for multiple hypothesis testing and the inherent complexities of analyzing datasets where the number of predictors approaches or exceeds the number of observations, thereby demanding a re-evaluation of statistical rigor in this field.

Before embarking on the creation of new factor libraries intended to capture market anomalies, a thorough understanding of existing methodological limitations is paramount. The pursuit of predictive power through increasingly complex models can easily lead to spurious correlations, where factors appear significant in historical data but fail to generalize to future market conditions. Traditional statistical techniques, particularly those reliant on hypothesis testing and p-values, struggle to effectively navigate the high-dimensional space created by numerous potential factors, increasing the risk of identifying relationships purely by chance. Rigorous selection techniques, including robust out-of-sample testing, cross-validation, and consideration of economic rationale, are therefore essential to distinguish genuinely informative factors from those arising from data mining or statistical noise, ultimately safeguarding against the pitfalls of overfitting and ensuring the reliability of any newly proposed asset pricing model.

The Alpha191 Library: A Microstructure-Driven Exploration

The Alpha191 Library consists of 191 distinct signals constructed from price, volume, and order-flow data, with all data originating from the Chinese A-share market. This market is notable for its substantial retail investor participation, currently representing approximately 80% of total trading volume. The library’s exclusive sourcing from this market provides a unique dataset, as retail investor behavior can significantly impact short-term price dynamics and create distinct market characteristics compared to markets dominated by institutional investors. The signals are designed to quantify these dynamics and capture patterns specific to this investor base.

The Alpha191 Library’s 191 factors are systematically categorized into three primary thematic domains to represent different facets of market dynamics. The Price Action domain focuses on immediate price changes and relationships, capturing short-term fluctuations. The Trend & Momentum domain identifies persistent price movements and the rate of change, indicating potential continuations or reversals. Finally, the Volume & Flow domain analyzes trading activity, including order book dynamics and the relationship between price and volume, to reveal insights into market participation and conviction. This organization allows for targeted factor selection and portfolio construction based on specific investment hypotheses and market conditions.

Alpha191 factors differentiate themselves from traditional investment factors based on fundamental analysis – such as those evaluating company earnings, assets, or debt – by focusing exclusively on directly observable market data. These factors are constructed using price, volume, and order flow, reflecting immediate trading activity rather than underlying economic conditions. This approach provides a distinct and complementary perspective, potentially capturing short-term dynamics and behavioral patterns not readily apparent in fundamental data. By utilizing quantifiable market behavior, the Alpha191 library offers a different lens for portfolio construction and risk management, and can be used in conjunction with fundamental factors to create more robust investment strategies.

The Alpha191 Library’s construction is predicated on the hypothesis of Behavioral Universality, which posits that consistent psychological biases among market participants drive predictable patterns in asset pricing. This theory suggests that irrational behaviors, stemming from cognitive and emotional factors, are not unique to any specific market but manifest consistently across diverse trading environments. The library’s factors are therefore designed to identify and exploit these universally-occurring behavioral effects, operating on the premise that these biases create temporary mispricings that can be captured through quantitative strategies. The intention is to create factors robust across markets, rather than those reliant on specific country or industry fundamentals.

Refining Signal Selection: From LASSO to Double-Selection

Dimensionality reduction in factor selection frequently utilizes techniques such as Least Absolute Shrinkage and Selection Operator (LASSO) and Elastic Net. These methods introduce L_1 and L_2 regularization terms to the regression objective function, penalizing model complexity and shrinking the coefficients of less important factors towards zero. This process effectively performs feature selection by eliminating irrelevant or redundant variables, thereby reducing the risk of overfitting, particularly when the number of factors exceeds the number of observations. By simplifying the model, LASSO and Elastic Net improve generalization performance on unseen data and enhance the interpretability of the results.

Standard Least Absolute Shrinkage and Selection Operator (LASSO) regression, while effective for variable selection and regularization, exhibits inherent bias when the number of predictors (p) approaches or exceeds the number of observations (n), a condition known as the high-dimensional setting (p > n). This bias arises because the penalty term in LASSO, designed to shrink coefficients, can inconsistently select relevant variables and suppress true signals. Specifically, in these scenarios, the estimated coefficients may not converge to the true underlying values, even with an infinite sample size. Consequently, more sophisticated techniques are required to mitigate this bias and ensure reliable factor selection in high-dimensional data, such as debiased LASSO or alternative selection procedures.

Double-Selection LASSO is a factor selection method designed to mitigate the bias inherent in standard LASSO implementations when applied to high-dimensional datasets. This technique operates in two stages: first, a preliminary LASSO regression identifies potentially relevant factors; second, a post-selection LASSO is performed using only the factors selected in the initial stage. This two-stage process effectively reduces the variance of the selected factors and improves statistical power by focusing the second regression on a smaller, more refined set of variables. The method’s efficacy stems from its ability to correct for the shrinkage bias of the initial LASSO, leading to a more reliable and stable factor selection process compared to single-stage LASSO approaches.

Principal Component Analysis (PCA) offers a method for dimensionality reduction when applied to the Alpha191 library, effectively decreasing the number of input features to a model. This simplification is achieved by transforming the original variables into a new set of uncorrelated variables, called principal components, ordered by the amount of variance they explain. By retaining only the components that capture a substantial portion of the total variance-typically 80-95%-the model’s complexity is reduced, mitigating overfitting and improving its ability to generalize to unseen data. This process not only decreases computational cost but also addresses potential multicollinearity issues among the original factors, leading to more stable and interpretable results.

Beyond Prediction: Implications and Future Directions

A newly developed library of 191 Alpha factors exhibits considerable promise for enhancing asset pricing models, potentially surpassing the explanatory power of conventional approaches. Rigorous testing reveals that these factors, initially derived from the China A-share market, don’t simply represent localized anomalies; a noteworthy seventeen factors demonstrated statistically significant performance when applied to the U.S. S&P 500, even after accounting for the influence of 151 established fundamental factors. This cross-market generalizability suggests the Alpha191 factors capture underlying economic rationales relevant beyond a single geographic context, offering a robust toolkit for investors and financial researchers seeking to refine portfolio construction and risk management strategies. The observed improvements imply that incorporating these factors could lead to more accurate predictions of asset returns and a deeper understanding of market dynamics.

The research reveals a surprising degree of cross-market applicability within a newly developed library of 191 investment factors, termed Alpha191. Originating from analysis of the China A-share market, the study found that 17 of these factors demonstrably explained variations in returns within the U.S. S&P 500, even after accounting for the influence of 151 commonly used fundamental factors. This suggests that information embedded within the Chinese market – regarding investor behavior, corporate strategies, or market microstructure – can offer predictive power in a distinctly different market environment. The retention of statistical significance for these 17 factors highlights a potential for broader, globally-applicable investment strategies and challenges the assumption that market-specific factors are entirely localized in their effect.

The study’s findings reveal a surprising degree of cross-market applicability within the Alpha191 factor library; a noteworthy 17 out of 191 factors, initially derived from the China A-share market, demonstrated statistically significant explanatory power for stock returns within the U.S. S&P 500. This robustness persisted even after accounting for the influence of 151 established fundamental factors, suggesting these Alpha191 factors capture unique information not already reflected in conventional asset pricing models. The consistent statistical significance-indicated by p-values below 0.05-highlights the potential for these factors to improve portfolio construction and risk management strategies beyond their origin market, offering a compelling argument for their broader implementation in global investment practices.

The potential of the Alpha191 factor library extends beyond initial demonstrations in the China A-share market and the U.S. S&P 500. Future investigations should systematically assess its performance across a broader range of global equity markets, including emerging economies and developed nations, to determine the universality of these predictive signals. Beyond simply replicating results, research could delve into the complex relationships between these factors, identifying synergistic effects or redundancies within the library. Understanding how different factor themes – such as value, momentum, or quality – interact could lead to the construction of more robust and diversified investment strategies, potentially uncovering hidden drivers of asset pricing and improving risk-adjusted returns. This deeper exploration of factor interplay represents a significant opportunity to refine investment models and enhance predictive power.

The study’s findings regarding the transferability of behavioral factors-specifically, the persistence of Alpha191 signals from the Chinese market into U.S. equities-highlights a core tenet of understanding complex systems. It suggests that underlying patterns of irrationality are not geographically bound, but rather represent universal aspects of human decision-making influencing asset pricing. This resonates with Paul Feyerabend’s assertion that “Anything goes.” The researchers demonstrate that even when rigorously controlling for established fundamental factors, these behavioral signals retain explanatory power, indicating that a purely rational model of market behavior is insufficient. The application of Double-Selection LASSO further emphasizes the need for flexible methodologies capable of capturing these nuanced, often unpredictable, relationships.

Beyond the Signal

The persistence of cross-market alpha, as demonstrated by the transferability of Chinese A-share trading signals to U.S. equities, subtly shifts the discourse. It isn’t merely about finding factors, but acknowledging the systemic, and potentially irrational, origins of those factors. The study suggests behavioral patterns aren’t localized anomalies, but rather fundamental components of price formation, echoing across disparate markets. A crucial question remains: are these signals truly predictive, or are they merely reflections of shared cognitive biases amplified by algorithmic trading?

Future research should move beyond simply identifying profitable signals and towards a deeper investigation of the underlying behavioral mechanisms. A fruitful avenue lies in exploring the interplay between these short-term trading signals and established fundamental factors. Do they represent genuine informational content, or simply exploit the mispricing created by behavioral overreactions to fundamental news? Furthermore, the limitations of the Double-Selection LASSO technique itself warrant attention – specifically, the potential for spurious findings in high-dimensional settings.

The apparent universality of these biases implies a degree of market inefficiency that is, perhaps, comforting. It suggests that the pursuit of alpha isn’t a zero-sum game predicated on superior information, but rather a continuous exploration of predictable irrationalities. The challenge, then, isn’t just to identify these patterns, but to understand their fragility and, inevitably, their eventual decay as markets adapt – a cycle as relentless as the price fluctuations themselves.

Original article: https://arxiv.org/pdf/2601.06499.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-01-13 23:13