A Practical Approach to Weighting Signals

A range of alpha signals can help investors discern whether a stock or a security is likely to outperform the market, and each signal has its own unique strengths and weaknesses. While an individual alpha signal can be effective on its own, thoughtfully combining signals is worth consideration.

In this article we discuss how effective signal weighting within a multi-signal strategy can offer a more robust, diversified, and adaptable approach to investment decision-making. This potentially leads to improved performance and risk-management outcomes.

Introduction

We will primarily focus on the Maximum Information Ratio, also referred to as Max IR, which is an approach designed to maximize return per unit of risk. For those interested in other weighting methodologies, download our Optimal Signal Weights white paper for detailed coverage of alternatives.

1. Signal weighting: The most obvious rationale is weighting signals according to an objective that reflects the philosophy of your firm’s investment team. We developed several approaches that are well documented in our white paper, including Max Information Ratio - Risk-Adjusted IC Based, Min Variance, Risk Parity, Max Return, and Max Information Ratio (Max IR). The Max IR methodology provides a comprehensive framework that integrates both expected return forecasts and risk modeling, with sufficient flexibility to enforce investment constraints.

2. Risk management: Even if an investment team prefers Equal Weighting or simple heuristics (such as 30% Value, 40% Quality, and 30% Sentiment), knowing the risk contribution continuously through a reporting framework is paramount since risk drift from the backtest and “alpha eating”^[1] can otherwise become issues.

3. Alternative hypothesis: If Equal Weighted or manual weighting is preferred, having an approach that highlights the difference from both a research and risk-management standpoint proves helpful.

The schematic below illustrates ways to evaluate different signal weighting objectives:

Signal Weighting

As outlined in our white paper, the purpose of using a weighting method other than Equal Weighting is to achieve a risk or IR outcome consistent with a firm’s investment philosophy (e.g., strong aversion to drawdowns). In many cases, Equal Weighting is a difficult benchmark to match when looking only at summary statistics. However, when one analyzes the results more carefully, it becomes clear there are many drawbacks.

Equal-weighted strategies often experience significant drawdowns, which are reflected in metrics such as Expected Tail Loss. This measure represents the average loss beyond a specified Value at Risk threshold, providing insight into potential adverse outcomes.

Another statistic that is often overlooked is turnover. For both Equal Weight and Risk Parity approaches, turnover can be quite high. By contrast, the turnover for Max IR can be calibrated to match the turnover of the Equal Weighted approach, while still delivering superior risk-adjusted properties through careful management of model constraints.

In this example, we used the FactSet Quant Factor Library (“QFL”) since it is point-in-time, comprehensive (close to the entire “Factor Zoo^[2]”), and broad (75,000 securities through time). We selected 12 robust and defensible signals.

A distinction is made between alpha signals and risk factors. Specifically, factors with positive Information Ratios are treated as signals for return generation, while factors like Volatility are treated as factors used for risk management. These were then evaluated in three different ways.

One important observation is that pure Max IR, like most optimized objectives, tends to amplify estimation error out-of-sample unless explicit constraints are imposed:

• Universe: US Large Cap (Financials and REITs were excluded from the universe as these sectors require tailored signals that were not part of this study).

• Structure: 12 Signals in 3 Composites (see appendix for details).

• Period: The evaluation period stretches back 156 months (from January 31, 2010, to December 31, 2025).

• Equal Weight: 1/12 or 8.33% weight for each signal.

• Risk Parity: Equal Risk Contribution. Each signal's risk contribution is 8.33%, but weights can range from 0% to 100%. Each month, the weights are recalibrated based on the Signal Ex-Ante covariance matrix.

  • The Signal Ex-Ante covariance matrix will be estimated using Factor Mimicking Portfolios (“FMP”) and asset level risk model data.^[3]

• Max IR: Each month, the weights are recalibrated based on performance over the previous 36 months and the asset level risk model data.

  • Return Estimation: The signal expected returns will be estimated by multiplying the historical Information Ratio (IR) of each signal by its Ex‐Ante risk estimate. The historical Information Ratios are based on the returns of the signals over the previous 36 months.

  • Risk Estimation: The Ex‐Ante risk estimates are calculated based on asset level data from the associated risk model and the factor mimicking portfolios. Evaluation Methodology: The approach for calculating performance of the signals utilizes FMP where the portfolios are scaled to have a Gross Exposure of 200% (100% Long and 100% Short) to ensure comparability across signals.

Table 1 shows the summary statistics.

Table 1: Backtest of Three Alphas

The results in Table 1 are sorted by the annualized Information Ratio, but this should be only one parameter to consider when evaluating performance. For example, the annualized volatility column provides important context. It shows the Risk Parity approach is the most sensitive to risk, since risk is the only objective under consideration, while Max IR demonstrates awareness of both risk and return. The Equal Weighting method does not directly control either the risk or return tradeoff.

Other informative statistics in the table include turnover, the length of the longest drawdown (an indicator of career risk), and the Expected Tail Loss. Over this 13-year period from January 2013 to December 2025, Max IR appears to be the preferred approach, particularly when considering the longest drawdown. In asset management, limiting the severity of long drawdowns is crucial to maintaining the confidence of stakeholders and the ability to continue managing capital.

From the perspective of career risk, it is useful to evaluate how combinations of alpha signals perform over time, rather than relying solely on overall summary statistics. In this analysis, both 24-month and 36-month rolling windows (these statistics are derived from the backtested alphas) are considered, although the primary construction of alpha signals is based on 36 months.

Examining the IR over these rolling periods helps to determine whether the experience of following the strategy remains acceptable through different market regimes. Even if a strategy displays superior statistics on average, extended drawdowns during periods of widespread factor underperformance (i.e., Factor Fimbulwinter^[4]) may make it unattractive for long-term adoption.

Chart 1: 36-Month Rolling IR

 Chart 2: 24-Month Rolling IR 

Risk Management

Since most quantitative strategies use an optimizer and emphasize stock selection (alpha) in the objective function, it is not uncommon for the risk budget to comprise at least 60% to 70% specific risk. This is the portion that cannot be diversified away.

Therefore, knowing the alpha’s characteristics for purpose of portfolio optimization is paramount. By leveraging the FMP framework and the risk model early (to scale both return and risk and constrain style factor exposure), the stock selection is more likely to reflect the type of portfolio that is eventually optimized.

Chart 3: Max IR Risk Contribution by Signal Composite

Chart 4: Max IR Signal Risk Contribution

Chart 5: Max IR Risk Budget and Expected Alpha Exposure

Chart 6: Style Factor Exposure

Alternative Hypothesis

Even if the philosophy relies on backtests, correlation matrices, and a heavy dose of human oversight, it might be wise to continuously review how different the production process is from a tool such as the Optimal Weights Engine.

All the previous charts can be tweaked and augmented to capture the details that are most important. For example, if one is concerned with how much risk comes from Size, one can compare different processes to the live one. In this case, Equal Weight is utilized as a proxy for the live version.

Chart 7: Size Exposure

Summary

Effective signal weighting is essential for building robust multi-factor portfolios. This analysis has shown that methods such as Equal Weight, Risk Parity, and Max IR produce different outcomes in risk, return, and turnover. Max IR, especially when combined with appropriate constraints, stands out for offering superior risk-adjusted performance and smaller drawdowns across a long-term backtest.

Risk management is critical at every stage of the weighting process. Careful use of position and risk constraints prevents excessive concentration while maintaining flexibility. Tools like the FactSet Optimal Weights Engine help practitioners implement and monitor these methods efficiently, allowing integration with any risk model and providing transparent reporting.

Continuous evaluation of weighting approaches, along with regular review of risk exposures and drawdowns, enables investment teams to respond to market changes and avoid extended periods of underperformance, which can pose career risk. By combining thoughtful optimization, thorough risk controls, and ongoing monitoring, signal weighting can significantly enhance portfolio durability and long-term investment outcomes.

Download Our White Paper

Access your free copy of our Optimal Signal Weights white paper, which digs into a top-down approach in our flexible framework to combine alpha signals into a multi-signal strategy. The paper highlights tailored FactSet Programmatic Environment (FPE) functionality, including:

• Capabilities to leverage asset-level information and enforce constraints

• How to modify objective and constraint sets to suit your requirements

• The importance of a flexible framework for solving diverse optimization models

• How FPE accompanies you throughout every step in developing frameworks based on systematic signal research

The article includes thoughtful input from Todor Bilarev and Nikolay Radev at FactSet.

^[1] "Alpha is Volatility Times IC Times Score" by Richard C. Grinold, The Journal of Portfolio Management, Summer 1994

^[2] Feng, G., Giglio, S., and Xiu, D. (2020). "Taming the Factor Zoo: A Test of New Factors." The Journal of Finance, 75(3), 1327–1370. Available at: https://onlinelibrary.wiley.com/doi/10.1111/jofi.12883

^[3] J. Menchero and J.H. Lee. Efficiently combining multiple sources of alpha. The Journal of Investment Management, 13(4):71–86, 2015

 ^[4] Hoffstein, C. (2018). "Factor Fimbulwinter." Newfound Research. Available at: https://blog.thinknewfound.com/2018/06/factor-fimbulwinter/ 

Appendix

Signals:

1. Sentiment
   1. 252D Return Momentum (log normalized)
   2. Price Target Estimate Revisions (75D)
   3. Sales Estimate Stability
   4. EPS Revisions
2. Quality
   1. Cash Flow Return on Invested Capital
   2. Gross Profit Return on Assets
   3. Net Margin
   4. Asset Turnover Change
3. Value
   1. Book-to-Price
   2. Earnings Yield
   3. EBIT-to-Enterprise Value
   4. Operating Cash Flow Yield

Signal Transformation:

1. Winsorized: Gaussian at +/- 4 standard deviations, 3 passes
2. If NA replace with Industry Group median value
3. All signals standardized using median and standard deviation.
4. Neutralized Quality and Value signals

Optimal Weights Engine Instructions

1. Signals
   1. +/- 100% of Equal Weight, i.e., 0.0% to 16.7%
   2. Composites: 50% to 150% of Equal Weight, i.e., 16.7% to 50.0%
2. Style Factors:
   1. +/- 0.70 risk factor exposures (weight x loading)
3. Turnover: 8% per month after first month.
4. Multivariate regression
   1. Orthogonalize the signals with respect to QFL 250-Day Beta and 250-Day Volatility (log normalized)
   2. Scale positions 1/Stock Specific Variance from Risk Model to reduce impact from outliers and size
   3. The regression includes an intercept to account for the scale difference between the dependent variable (expressed in percent) and the independent variables (expressed as z-scores)
   4. Total Return is winsorized using Gaussian at +/- 4 standard deviations, 5 passes
5. Return Scaled to Risk Model
6. Risk calculated using Risk Model (applies to Risk Parity as well)

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.