Factor investing has long been recognized as a method for systematically identifying stocks anticipated to outperform the market. This strategy relies on the premise that specific inherent risks—such as value, momentum, quality, and size—are rewarded by the market. However, emerging findings indicate that the underlying models used in this approach may be fundamentally flawed, resulting in disappointing outcomes for investors.
Recent studies suggest that the issue lies not in the data itself, but in the construction of these predictive models.
An analysis conducted by the ADIA Lab and published by the CFA Institute Research Foundation introduced the concept of a factor mirage, where models mistakenly confuse correlation with causation.
Table of Contents:
The allure of factor investing
Factor investing emerged from a theoretically sound premise: investors should yield higher returns by exposing their portfolios to specific systematic risks. This foundational idea has attracted trillions of dollars into various investment products designed around these factors. Nonetheless, the performance of these products has frequently fallen short of expectations.
Performance indicators
For example, the Bloomberg–Goldman Sachs US Equity Multi-Factor Index, which assesses the long-short strategy performance of classic factor models, has a disappointing Sharpe ratio of only 0.17 since 2007. This statistic indicates that the model’s returns, adjusted for risk, are statistically indistinguishable from zero when costs are considered. Such results have contributed to years of underperformance for many funds and have eroded investor confidence in these strategies.
The model misalignment
Common explanations for these failures often cite issues such as backtest overfitting or the practice of p-hacking, where researchers manipulate data to find statistically significant results. While these explanations are valid to some extent, they do not encompass the full scope of the problem. The research from ADIA Lab highlights a more profound issue: systematic misspecification of models.
Understanding model specifications
Most factor models are constructed using classical econometric methods, including linear regressions and significance tests. These techniques frequently misinterpret associations as causal relationships. For instance, the econometric literature suggests that all variables correlated with returns should be included in regressions, regardless of their causal roles. This practice can introduce significant biases, particularly when colliders—variables affected by both the factor and the returns—are included or confounders—variables influencing both—are omitted.
Such biases can distort the coefficients of a factor model, leading to inaccurate investment signals. Consequently, investors may find themselves purchasing securities they should sell or vice versa. Even if risk premiums are correctly measured, a misaligned model can result in persistent losses.
The dangerous allure of the factor zoo
The term factor zoo refers to the multitude of published anomalies that fail when tested out-of-sample. However, the work from ADIA Lab indicates an even subtler danger: the factor mirage. This phenomenon does not arise from data mining; rather, it stems from inherent flaws in model specifications, despite adherence to conventional econometric principles.
Misspecified models can produce misleadingly high R² values and lower p-values, which create a false impression of model reliability. When colliders are present, the model’s ability to predict returns becomes compromised, leading to misleading conclusions that can have significant financial consequences.
Practical implications for investors
To illustrate, consider two researchers attempting to determine a quality factor. One researcher appropriately controls for variables such as profitability, leverage, and size, while the other introduces return on equity, a variable influenced by both profitability and stock performance. The second researcher’s model may appear superior in backtests, but in real-world trading scenarios, it becomes a statistical illusion that can undermine capital.
Recent studies suggest that the issue lies not in the data itself, but in the construction of these predictive models. An analysis conducted by the ADIA Lab and published by the CFA Institute Research Foundation introduced the concept of a factor mirage, where models mistakenly confuse correlation with causation.0
A shift towards causal reasoning
Recent studies suggest that the issue lies not in the data itself, but in the construction of these predictive models. An analysis conducted by the ADIA Lab and published by the CFA Institute Research Foundation introduced the concept of a factor mirage, where models mistakenly confuse correlation with causation.1
Recent studies suggest that the issue lies not in the data itself, but in the construction of these predictive models. An analysis conducted by the ADIA Lab and published by the CFA Institute Research Foundation introduced the concept of a factor mirage, where models mistakenly confuse correlation with causation.2
Recent studies suggest that the issue lies not in the data itself, but in the construction of these predictive models. An analysis conducted by the ADIA Lab and published by the CFA Institute Research Foundation introduced the concept of a factor mirage, where models mistakenly confuse correlation with causation.3
