Factor investing emerged as a method to apply scientific rigor to the stock market, aiming to uncover why certain stocks consistently outperform others. However, as time has progressed, results have often fallen short of expectations. Current research indicates that the challenge lies not with the data itself but with the underlying models interpreting this data. A recent study introduces the concept of a factor mirage, where models mistakenly equate correlation with causation.
The initial appeal of factor investing stems from a clear rationale: markets generally reward investors who assume specific, non-diversifiable risks, such as value, momentum, quality, and size. This theoretical framework has led to trillions of dollars being invested in various financial products based on these principles. Yet, empirical evidence presents a starkly different reality.
Table of Contents:
Performance of factor models
Analyzing performance metrics, the Bloomberg-Goldman Sachs US Equity Multi-Factor Index reveals troubling outcomes. Since 2007, it has recorded a Sharpe ratio of merely 0.17, statistically insignificant when accounting for costs. In simpler terms, factor investing has not delivered the anticipated benefits for investors. This underperformance has disappointed individual investors and diminished the credibility of fund managers relying on these models for their strategies.
Conventional wisdom attributes this failure to issues like backtest overfitting or p-hacking, where researchers sift through data noise until they discover what appears to be a legitimate alpha. While this explanation holds some merit, it fails to capture a more profound issue identified by recent analysis from the ADIA Lab, published by the CFA Institute Research Foundation: a systemic misspecification in how these models are constructed.
Understanding model misspecification
The majority of factor models are founded on traditional econometric principles, relying heavily on linear regression, significance testing, and two-pass estimators. This approach often leads to a conflation of correlation with causation. Econometric education instructs students to include any variable associated with returns, regardless of its role within the causal framework.
This methodological oversight can significantly distort coefficient estimates. Specifically, including a collider—a variable affected by both the factor and the returns—while omitting a confounder—a variable that impacts both—can skew results dramatically. Such biases can even reverse the sign of a factor’s coefficient, prompting investors to make poor decisions based on flawed data.
The implications of a factor mirage
The concept of the factor zoo is well-recognized; it refers to the myriad anomalies published in research that fail under real-world conditions. However, the ADIA Lab’s research highlights a more insidious issue: the factor mirage. This phenomenon is not simply a result of data mining but stems from models that are fundamentally misconfigured, even when adhering to established econometric guidelines.
Models that include colliders can present misleadingly high R² values and lower p-values, leading researchers to mistakenly favor these models over correctly specified alternatives. In scenarios where a collider is present, the return’s value is predetermined before considering the collider, resulting in illusory anticipated profits. Such methodological errors can have staggering financial implications, costing billions in potential earnings.
Real-world effects on capital allocation
For instance, if two researchers evaluate a quality factor, one may rightfully account for profitability, leverage, and size, while another might erroneously incorporate return on equity—a variable influenced by both profitability and stock performance. The latter’s model could misleadingly correlate high quality with strong past returns, appearing superior in backtesting but ultimately resulting in capital erosion in live trading.
The ramifications of model misspecification extend beyond individual portfolios. According to ADIA Lab’s findings, no portfolio can truly achieve efficiency without a careful understanding of causal factors. If foundational components are misidentified, even accurate estimates of means and covariances will lead to subpar portfolio performance. This indicates that investing is not merely a problem of forecasting; introducing complexity does not inherently enhance model accuracy.
Shifting towards causal reasoning
The initial appeal of factor investing stems from a clear rationale: markets generally reward investors who assume specific, non-diversifiable risks, such as value, momentum, quality, and size. This theoretical framework has led to trillions of dollars being invested in various financial products based on these principles. Yet, empirical evidence presents a starkly different reality.0
The initial appeal of factor investing stems from a clear rationale: markets generally reward investors who assume specific, non-diversifiable risks, such as value, momentum, quality, and size. This theoretical framework has led to trillions of dollars being invested in various financial products based on these principles. Yet, empirical evidence presents a starkly different reality.1
The initial appeal of factor investing stems from a clear rationale: markets generally reward investors who assume specific, non-diversifiable risks, such as value, momentum, quality, and size. This theoretical framework has led to trillions of dollars being invested in various financial products based on these principles. Yet, empirical evidence presents a starkly different reality.2
