Portfolio managers often prefer the diversification advantages of using broad investment universes and more than a single source of alpha. The typical research workflow involves identifying and selecting the proper signals, then finding the optimal way of combining them into a composite alpha signal, which is used in the final portfolio optimization.
In this article we explore the benefits of more granular (bottom-up style) analysis in the research and generation of alpha signals for optimizing equity portfolios. We will group assets according to their sectors classification in this analysis, but the points made can naturally extend to other grouping categories (e.g., country or finer separations of industries/subindustries) as long as enough assets are present in each group to perform meaningful analysis.
One caveat of using a diverse investment universe is that there can be a large variation of how an asset’s price depends on different signals. For example, say a portfolio manager wants to invest in value but there are sectors where book to price is the best predictive metric and others where sales to price is a better value metric.
This problem gets exacerbated when using more than one signal to combine into a composite alpha. Moreover, many signals have biases in the means and dispersions across different sectors. Various sector neutralization techniques are frequently utilized to varying success to mitigate the latter effect. However, the former is not addressed as often.
An approach that accounts for both issues is the more drastic separation of sectors from the very beginning of the alpha research workflow. That separates the universe into a set of smaller sub-universes, each containing the constituents of the respective sector.
For each, we perform a separate analysis to select and optimize the components of the composite alpha signal, and only in the end do we aggregate the alpha signals for each sub-universe into the alpha signal for the entire universe. In the following section this method is explored through a case study. A schematic representation of the method is shown below:
Figure 1. Schematic representation of the proposed method
In this example we use the top 1,000 U.S. public companies by market cap over the period January 2018 – December 2023 as an investment universe, using monthly frequency data. We emulate a completely from-scratch research process, with in-sample analysis of the first three-year period to generate the alpha signal and an out-of-sample back-test of the aggregated signal in the second half of the available period. We start with the following 112 potential alpha signals:
Table 1. The starting set of potential alpha signals
First, we use the sector classification to separate the universe into 11 sectors. From here onward, all analysis is performed on each sector separately. The signal scores are winsorized and normalized—this in itself is one of the simpler methods of sector neutralization that is often used on its own.
Next, we use the first three years of data (January 2018 – December 2020) for in-sample signal selection. This process involves two steps: 1. reduction of multicollinearity and 2. selecting the best forward return predicting signals.
We achieve the first step by using the Variance Inflation Factors (VIFs) method. VIF is a measure of multicollinearity, or how strongly a signal is correlated to a linear combination of the rest of the signals. The VIFs of all signals are calculated, and the signal with the highest VIF is dropped. Then the process is iteratively repeated until all remaining signals have a VIF below a desired threshold.
This reduces the number of signals from 112 to sets of 60 - 80 linearly independent ones in each sector. At this point there are already differences in the signals selected that form the best linearly independent basis in which to search for an optimal composite signal in each sector. Then a multilinear model of forward asset returns from signal scores is used to find the subset of signals that best predicts returns without overfitting.
For this purpose, optimize for the Akaike Information Criterion (AIC) using a bidirectional stepwise regression with Monte Carlo randomizations of the starting points to avoid sub-optimal selection of signals. The following signals were chosen by sector, and we do the same analysis over the whole universe for comparison:
Table 2. The final selection of alpha signals for each sector, including a selection over the whole universe
The variety of selected signals across the different sectors is a strong indicator supporting the initial claim that signals have different predictive power across different sectors.
Signal returns for each date (and sector) are calculated using a light multivariate cross-sectional regression—including the signal score as well as beta and volatility risk factor exposures. We then run an optimization maximizing the Information Ratio (IR) to find optimal weighting of signals for each date in the out-of-sample period (January 2021 - December 2023).
The IR calculation is based on an expected return estimate that uses a three-year rolling time window preceding each date and ex-ante risk from the FactSet US Equity Risk Model. A couple of simple constraints are also applied: 25% double sided turnover limit, and an upper limit on each signal weight of 50% (for diversification).
With the calculated optimal weights, we generate the optimized composite alpha signal for each sector and recombine them into the complete universe. Then, the out-of-sample backtest (January 2021 - December 2023) is performed to evaluate how effective the final aggregated composite alpha signal is. We also perform similar analysis over the whole universe for comparison. Below are the results, sorted by IR ratio:
Table 3. Out-of-sample performance of different composite alpha signals
The proposed strategy (MAX_IR_Composite_by_sector) outperforms the alternatives. The second-best performer (MAX_IR_universe_non_neutralized_composite) is the same analysis and optimizations performed only on the whole universe. Skipping the optimization step and just taking an equal-weighted composite of the selected signals produces alphas that perform worse than their optimized counterparts.
All data used was sourced by FactSet, and analysis was performed using the tools available in the FactSet Programmatic Environment (FPE)—specifically, the Backtest, SignalSelector, and OptimalWeightsEngine packages.
References
Svallin, J., Mitov, G., Margaritov, E., Radev, N., Gum, B., Bilarev, T., and Stefanov, I. 2023. “Optimal Signal Weights” FactSet Whitepaper.
Svallin, J., Mitov, G., Radev, N., and Atanasov A. 2024. “Signal Selector” FactSet Whitepaper.
This blog post is for informational purposes only. The information contained in this blog post is not legal, tax, or investment advice. FactSet does not endorse or recommend any investments and assumes no liability for any consequence relating directly or indirectly to any action or inaction taken based on the information contained in this article.