Factor investing has been recognized as a method to apply a scientific framework to the stock market, aiming to uncover the reasons behind the outperformance of certain stocks. However, after years of underwhelming results, analysts are increasingly acknowledging that the primary issue lies not within the data itself, but in the construction of these models. Recent research reveals that many factor models mistake correlation for causation, a phenomenon known as the factor mirage.
The premise of factor investing is based on the idea that markets inherently reward specific risks that cannot be diversified away, including value, momentum, quality, and size. This principle has driven the allocation of trillions of dollars into financial products designed around these factors, yet the outcomes suggest a different narrative.
Table of Contents:
Analyzing the disappointing results of factor investing
The statistics present a concerning picture. The Bloomberg–Goldman Sachs US Equity Multi-Factor Index has demonstrated a Sharpe ratio of just 0.17 since 2007, accompanied by a t-statistic of 0.69 and a p-value of 0.25. These statistics indicate that the performance of factor investing strategies is statistically indistinguishable from zero when not accounting for costs. In practical terms, factor investing has failed to provide the expected value to investors. For asset managers depending on these models, this underperformance represents years of disappointing results and a substantial loss of credibility.
Understanding the underlying problems
The common explanations for suboptimal investment outcomes often cite backtest overfitting and p-hacking. These terms refer to researchers selectively analyzing data until they identify what seems to be actionable alpha. Although these reasons have some merit, they do not fully encompass the complexities involved. New research from the ADIA Lab, published by the CFA Institute Research Foundation, highlights a more significant concern: systematic misspecification of models.
Many factor models rely on a conventional econometric framework that utilizes linear regressions and significance tests, which can blur the line between mere correlation and genuine causation. Standard econometrics literature often suggests incorporating any variable correlated with returns, irrespective of its actual role within the causal framework. This practice can lead to considerable methodological inaccuracies.
Implications of model misspecification
The incorporation of colliders—variables influenced by both the factor and the returns—into a financial model can lead to significant biases in coefficient estimates. Such biases may reverse the sign of a factor’s coefficient, potentially misguiding investors into purchasing assets they should sell and vice versa. Even when all risk premiums are accurately estimated, a misspecified model can still yield consistent losses.
The phenomenon known as the factor zoo illustrates the many published anomalies that fail to withstand out-of-sample testing. However, the factor mirage presents a more subtle and insidious challenge. This issue arises not from data mining, but from models that are incorrectly specified, despite following the econometric principles typically taught in academic environments.
How colliders can create illusions of success
Models that incorporate colliders frequently exhibit inflated R² values and reduced p-values. This can mislead researchers into believing their models are accurately specified. In these scenarios, the return value is established before the collider. As a result, the increased correlation produced by the collider cannot be effectively monetized. Consequently, the profits indicated by academic studies may prove to be illusory, with potential real-world implications reaching billions.
Consider two researchers working to estimate a quality factor. One researcher effectively controls for metrics such as profitability and leverage. The other, however, includes return on equity, a variable influenced by both profitability and stock performance. This addition can create a misleading impression, suggesting a correlation between high quality and strong past returns. Although the second model may show favorable results in backtesting, live trading often uncovers a statistical illusion that can gradually deplete capital.
Moving towards causal reasoning in finance
The importance of accurate causal models in investment portfolios
The consequences of model misspecification in investment strategies are significant. Research from ADIA Lab highlights that no investment portfolio can achieve optimal efficiency without accurately specified causal factor models. When the underlying factors are incorrectly defined, even the best estimates of means and covariances lead to subpar portfolio performance. This finding emphasizes that investing is not solely about making predictions; increasing complexity does not automatically enhance model effectiveness.
Addressing the challenges of factor investing requires a fundamental shift towards causal reasoning. This evolution in methodology is not exclusive to finance; other sectors, such as medicine, have moved from correlation-based methods to causal frameworks, thereby improving the reliability of treatment strategies. As finance aims to adopt a similar approach, the goal should not be scientific perfection but rather practical reliability.
A causal model that identifies the true sources of risk and return enables investors to allocate capital more effectively and offers credible explanations for performance outcomes. For investors, adopting this approach is crucial, as it encourages the development of strategies that endure in real-world conditions. These models clarify the reasons for their success, rather than simply claiming that they are effective.
To regain its scientific credibility, factor investing must advance from merely recognizing patterns to comprehending the underlying causal mechanisms. This transition marks a pivotal moment when quantitative investing evolves from being systematic to genuinely scientific.
