- Among a flood of alternative data sources, consumer sentiment based on online citations stood out as adding explanatory power in the cross-section of returns.
- The consumer sentiment signals examined had little correlation with traditional style-factor exposures.
- Individual metrics and the equal-weighted combination had positive factor and quintile spread returns and positive information ratios.
As factor investing and technology have evolved, investors have been inundated with alternative data, much of it coming with great claims of adding value to the investment process. Among the many alternative data sources we have examined, data on consumer sentiment stood out as a diverse and potentially powerful metric beyond traditional factors.1
Spotlight on consumer sentiment
We analyzed a number of consumer-sentiment metrics based on changes in “web luminosity” (the number of citations a company receives on its products and brands in reviews, social media and other web content) and its relationship with that company’s fundamentals and stock performance.
For this analysis we used BrandLoyalties' “as-published” data, which begins in 2012.2 As of Dec. 31, 2018, the data covered over 1,500 global companies with approximately 1,100 in the MSCI ACWI IMI universe and close to 1,000 in the MSCI USA IMI. Sectors with more visibility among consumers, such as consumer discretionary, consumer staples and technology had higher coverage than those with less consumer presence, such as energy and utilities. Because the sample is dominated by U.S. companies, we used the MSCI Barra US Total Market Equity Trading Model (USFAST) for our analysis and restricted our sample to the MSCI USA IMI universe.
Coverage of BrandLoyalties data in MSCI indexes
Consumer sentiment factors were largely uncorrelated to traditional factors
To test how independent these consumer-sentiment metrics were, we calculated their cross-sectional correlation with the style-factor exposures of USFAST and with each other. In the exhibit below, we see that most of the correlations, including those with other sentiment measures, such as analyst sentiment (earnings estimate revisions) and short interest, were near zero.
Correlations of consumer sentiment metrics3 with each other and USFAST style factors
|Signal||Percent Change YOY||Slope Last 91 Days||Percentile Ranking|
|Percent Change YOY||1.00|
|Slope Last 91 Days||0.28||1.00|
Average cross-sectional correlation of BrandLoyalties signals with USFAST style factors and with each other. Sample period is 2012-2018 in the MSCI USA IMI universe.
Looking beyond correlation
Of course, low correlation to existing factors does not guarantee a metric will add explanatory power on top of other factors. To assess marginal explanatory power in the cross-section of returns, we evaluated the performance of the consumer sentiment metrics in both univariate and multivariate frameworks. The multivariate results show the unique contribution of a given factor over and above those of all other factors – style, industry and market. In the exhibit below, we display performance statistics of the BrandLoyalties metrics as well as an equal-weighted combination of the metrics averaged over 2012 to 2018.
Averaged performance of consumer sentiment metrics
|Quintile Portfolio Statistics||Multivariate Regression Statistics||Stability|
|Signal||Quintile Spread Return, %||Quintile Spread IR||Rank IC (%)||Factor Return, %||Factor Volatility, %||Factor IR||Mean |t||||t| > 2, %||Exposure auto-correlation|
|Percent Change YOY||0.85%||0.16||1.57%||0.75%||1.17%||0.64||0.98||11.26%||0.92|
|Slope Last 91 Days||1.98%||0.5||0.81%||0.61%||1.02%||0.6||0.91||10.99%||0.55|
Data from 2012 to 2018. Quintile spreads are calculated as the difference in equal-weighted top- and bottom-quintile portfolio monthly returns. Rank ICs are defined as the rank correlation of the month-end signal exposures with security returns in the subsequent month. The multivariate statistics are the results of weekly cross-sectional regressions in USFAST. Stability is defined as month-to-month auto-correlation in exposure.
While the t-stats and volatilities were relatively low, all the individual metrics, as well as the equal-weighted combination had positive factor returns, quintile spread returns and information ratios. These characteristics are more typical of alpha associated factors than traditional risk factors such as size and beta. Stability was generally moderate, in the 0.50 to 0.90 range that we label medium-term factors. Further, the individual metrics and the combination factor generated positive quintile spreads and factor returns. Those results are shown in the exhibit below.
Cumulative factor returns of consumer sentiment metrics were positive over time
In all, our results indicated that online consumer-sentiment signals provided additional transparency into sources of risk and return and could be a valuable new factor.
1 Factors constructed from BrandLoyalties data are included in MSCI FactorLab.
2 BrandLoyalties as-published data represent exactly the then-live data provided to their clients on any given historical day. BrandLoyalties also maintains a bi-temporal pro-forma data set, which starts in 2006, that back fills history of their full-coverage universe as of a given date. The back filled pro-forma data also restates the historical metrics using the latest available mix of brand information and corporate actions.
3 We evaluated three BrandLoyalties as-published metrics
- Percent Change YOY: year-over-year percentage change (to two decimal places) in the daily number of web-based consumer brand citations, averaged over the trailing 91 calendar days.
- Slope Last 91 Days: regression slope of the year-over-year change in the number of web-based consumer brand citations over the last 91 calendar days.
- Percentile Ranking: the percentile ranking (0-100) of the citation share growth of the brand names for each company on that given day. Percentile ranking is among the entire BrandLoyalties universe, calculated via an algorithm that weights both YOY change and 91-day slope of citation share.
Peering into peer selection: Quantifying company similarity