How to enhance sentiment strategies using MSCI Barra Risk Models

Executive Summary:

We move beyond excess returns to consider specific returns in a daily strategy which improves risk-adjusted performance driven by both higher returns and lower volatility.

Hedging Sentiment Signals with MSCI Barra Risk Models' Highlights

• For the U.S., Information Ratios (IR) of 3.8 and 6.1 are achieved for the Russell 1000 and the Russell 2000, respectively. We get IRs of 3.2 and 4.2 for their European equivalent.
• The predictive power of the signal is statistically significant and positive across the entire backtesting period.
• The prediction quantile analysis shows a desirable profile, providing higher returns for more extreme predictions.

USFAST Model	Large/Mid Cap		Small-Cap
Statistics	Specific Return	Excess Return	Specific Return	Excess Return
Annualized Return	12.9%	10.9%	32.7%	30.1%
Annualized Vol.	3.4%	4.5%	5.4%	6.1%
Max. Drawdown	2.3%	6.4%	2.8%	4.4%
Information Ratio	3.76	2.42	6.07	4.92

Experimental Stats. Comparison between MSCI specific returns and excess returns for a set of statistics. Results are shown for USFAST model for Large/Mid-Cap and Small-Cap. Source:RavenPack, MSCI, November 2017.

Details on the methodology used for the strategy and on the achieved results:

Excess returns: Ignores several risk factors

For simplicity, researchers often use excess returns when evaluating the efficacy of a trading signal. This is equivalent to assuming just a single risk factor (the market) with constant factor exposure (𝛽=1). However, in reality, several other risk factors exist and exposure not only differs across time but also across assets. This can lead to erroneous conclusions when it comes to signal performance, both in terms of returns and volatility.

To truly isolate the effects of a given trading signal, we need to account for the contribution from a broader set of risk factors and ensure that any observed performance cannot be explained by known risk premia.

Excess Returns vs. Specific Returns: Hedging a Sentiment Signal

In particular, we move beyond excess returns to consider the specific returns as defined by several distinct MSCI Barra risk models . This is done as a convenient shortcut to a factor-neutral portfolio implementation. In this context, we evaluate the effect of applying a regional model vs. a global model (USFAST/ USMED vs. GEM3), and faster vs. slower factors (USFAST vs. USMED).

To create our sentiment signal, we take a 1-day average of RavenPack’s Event Sentiment Score (ESS), filtering on highly relevant and novel events. The resulting sentiment indicator is then used as input to a simple linear regression model with the dependent variable being the MSCI specific returns .

Excess Returns vs. Specific Returns: Our main findings

• RavenPack sentiment adds consistent value in a daily strategy across both U.S. large/mid-caps and small-caps – after accounting for MSCI risk factors (see Figure 4). In particular, we achieve Information Ratios of 3.8 and 6.1 (before transaction costs), respectively, across the two market-cap portfolios.



• Sentiment models built on MSCI’s regional risk models (USFAST) outperform, in terms of Information Ratio, the global model (GEM3) by 45% across large and mid-caps and by 18% on small-caps (see Figure 1).



• The performance of the prediction quantiles shows the desired monotonic behavior - with more extreme predictions leading to more extreme returns and Information Ratios (see Figure 6).

Introduction

When evaluating the efficacy of a trading signal, for simplicity researchers often use excess returns. This is equivalent to assuming just a single risk factor (the market) with constant factor exposure (𝛽 = 1) – inspired by CAPM [1]. However, in reality, several other risk factors exist and exposure not only differs across time, but also across assets. This can lead to erroneous conclusions when it comes to signal performance, both in terms of signal return and volatility. To truly isolate the effects of a given signal, we need to account for the contributions from a broader set of risk factors and ensure that any observed performance cannot be explained by known risk premia [2, 3, 4, 5].

The use of news sentiment has recently been a broadly treated research subject [6, 7]. In particular, in our previous paper [8], we introduced a simple daily strategy based on average sentiment per company. We showed how companies with positive return predictions outperformed companies with negative return predictions across various investment horizons. In this paper, we aim to build upon this previous study by introducing a risk model to examine the performance of RavenPack sentiment signals after taking into account the contributions from the market, industries, and styles such as value, momentum, size and many others.

In particular, we move beyond excess returns to consider the specific returns as defined by several distinct MSCI risk models . This is done as a convenient shortcut to a factor-neutral portfolio implementation. In this context, we evaluate the effect of applying a regional model vs. a global model (USFAST/ USMED vs. GEM3 in the U.S. and EUE4DUK vs. GEM3 in Europe), and faster vs. slower factors (USFAST vs. USMED within the U.S.). Furthermore, it allows us to isolate the alpha generated by our sentiment signal.

As part of this research, we will focus on three different topics: (1) we will perform an analysis on the different signal implementations using the specific returns from the different risk models; (2) we will showcase the benefits of constructing signals based on specific returns including an analysis on the quantile performance; and (3) we will perform an evaluation of the signal speed or decay. As our findings are consistent and robust across both U.S. and Europe, we will focus primarily on the U.S. region, providing a summary of the European results in Appendix A.

The layout of this paper is as follows. Section 2 provides an overall description of the data used in this research. We explain how to build the reference indicator to predict specific returns provided by the risk models and outline the regression framework we implement in order to produce our trading signals. Results for the various risk-based signals are presented and discussed in Section 3. Finally, in Section 4, our general conclusions are provided.