Out-of- Sample Stock Return Predictability of Alternative COVID-19 Indices

ABSTRACT

We explore the predictive value of the various indices developed to capture COVID-19 pandemic for daily stock return predictability of 24 Emerging Market economies (based on data availability). We identify eight measures of COVID-19 indices, namely, the uncertainty due to pandemics and epidemics (UPE) index, Global Fear Index (GFI), COVID index, vaccine index, medical index, travel index, uncertainty index and aggregate COVID-19 sentiment index. We find that, out of the considered measures, the GFI consistently offers the best out-of-sample forecast gains followed by the aggregate COVID-19 sentiment index while the UPE index offers the least predictability gains. The outcome generally improves after controlling for oil price but the ranking of forecast performance remains the same and robust to multiple forecast horizons and alternative forecast evaluation methods. We infer that the relative predictive powers of the indices are proportional to the extent to which the indices truly measure the pandemic.

KEYWORDS:

• COVID-19 indices
• out-of-sample forecast evaluation
• emerging markets

JEL:

• C53
• D80
• G12

Acknowledgments

The authors wish to acknowledge the many helpful comments received from the Editor-in-Chief and two anonymous reviewers. However, usual disclaimer applies.

Disclosure Statement

No potential conflict of interest was reported by the author(s).

Notes

1. This connection also captures other financial markets such as foreign exchange market (see Narayan 2020a, 2020b; Salisu, Lasisi, and Olaniran 2021).

2. The start date for the data scope is informed by the discovery of the COVID-19 virus while the end data is given by the end date of the available COVID-19 indices.

3. The GFI data can be obtained from https://www.researchgate.net/publication/350947756_COVID-19_Global_Fear_Index_Data_Update_16042021.

4. The historical average is used as a baseline model here since it is an equivalent version of a random walk model for logged stock prices (see Bannigidadmath and Narayan 2015; Narayan and Gupta 2015; Phan, Sharma, and Narayan 2015; Narayan et al. 2016; Devpura, Narayan, and Sharma 2018; Salisu, Swaray, and Oloko 2019a, 2019b; Salisu and Akanni 2020).

5. Although, there are traditional predictors of stock returns, however, these factors are not considered in this paper for a number of reasons. First, the period for the predictability analysis is limited to the COVID-19 pandemic period and we use daily data for relevant variables in the predictive model. Since the pandemic indices used in this study only start from year 2020, the traditional factors with the highest frequency being monthly can only guarantee 20 (monthly) data points in order to align with the available data scope for the daily pandemic indices. This certainly will not offer meaningful outcomes. Second, should we use the data anyway, this will completely change the focus of the study on comparing the predictability of alternative COVID-19 pandemic indices. Third, the literature is already replete with studies involving the traditional predictors of stock return predictability (see Narayan and Bannigidadmath 2015; Narayan et al. 2016; Phan, Sharma, and Narayan 2015) and we do not intend to repeat the exercise in our study. Rather, we isolate the COVID-19 effect in the return predictability following the studies of Narayan and Sharma (2014) and Narayan and Gupta (2015) which also isolated the role of oil price ain return predictability. Fourth, we employ the heterogenous panel data techniques of Chudik and Pesaran (2015), Chudik et al. (2016) and Westerlund and Narayan (2016) which simultaneously capture both the unobserved common factor loading and heterogeneous factor loading in the estimation process, thereby circumventing any inherent bias associated with the omission of other predictors.

6. Simulations conducted by Pincheira, Hardy, and Muñoz (2021) show that the WCW is asymptotically normal and it is well-sized compared to the CW by keeping the distribution of the test statistic around zero mean and removing autocorrelation from the structure.