Lagged country returns and international stock return predictability during business cycle recession periods

ABSTRACT

This study examines stock return predictability in business cycle fluctuations across 17 developed countries and 26 developing countries over the period from January 1970 to December 2019. We uncover that lagged U.S. returns can be regarded as a reliable predictor only during recessions. The results remain robust after controlling for commonly used return predictors. Our empirical findings carry some implications for the role of leading markets, fundamental uncertainty, change in investors’ beliefs and dynamics of stock return volatility in economic downturns.

KEYWORDS:

• Return predictability
• recession
• lagged country returns
• business cycle

JEL CLASSIFICATION:

• G15
• G12

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

¹ The sample period for 26 developing markets is from January 1988 to December 2019, but the starting sample dates may vary for some markets due to the availability of data. Additionally, though Finland and New Zealand are developed markets, their data starting date is 1988; we thus put them in the group of developing countries.

² Rapach, Strauss, and Zhou (2013) show out-of-sample R²s in recession periods using graphs, while Siliverstovs (2017) tests the significance of out-of-sample R²s by dividing the sample into recession and expansion periods. As our focus is on in-sample predictability to demonstrate that the predictability mainly lies in the recession periods, we use the dummy variable approach.

³ In our additional test, we take into consideration the publication lag of the NBER recession indicator, which is assumed to be two months. We run the following regression model: $R_{i,t+1}^{own}={\alpha }_i+{\beta }_i{R}_t^{US}+{\gamma }_i{R}_t^{US}\times Re{c}_{t-2}+{\epsilon }_{i,t+1}$ to check the predictive power of lagged U.S. returns in recessions using a lag of two months. Our result is quite similar to the one presented in Panel A of Table 4. The results also hold for the assumption of one-month publication lag.

⁴ We also run Equation (3)(3) $R_{i,t+1}^{own}={\alpha }_i+{\beta }_i{R}_{i,t}^{own}+{\gamma }_i{R}_t^{US}+{\epsilon }_{i,t+1}$(3) by including two lagged country returns for the full sample. Untabulated results again confirm a weaker prediction.

⁵ Following most studies, we report Newey-West t-statistics using fixed lags (12 months in our paper). In a recent paper, Lazarus et al. (2018) recommend to use lags which vary with sample size (number of lags = 1.3T^1/2, where T is the time-series sample size; see Equation (3)(3) $R_{i,t+1}^{own}={\alpha }_i+{\beta }_i{R}_{i,t}^{own}+{\gamma }_i{R}_t^{US}+{\epsilon }_{i,t+1}$(3) in their paper) in Newey-West tests for the balance of size distortion and power loss. However, the fixed-b critical values used in their recommendation is non-standard, but can be derived using Table 1 in Kiefer and Vogelsang (2005). We also conduct such a test following their recommendation, and find very similar results (results are not shown here to conserve space).

⁶ Similarly, when the lagged own country returns are removed in Equation (8)(8) $R_{i,t+1}^{own}={\alpha }_i+{\beta }_i{R}_{i,t}^{own}+{\gamma }_i{R}_t^{US}+{\delta }_i{R}_t^{US}\times RE{C}_t+{\epsilon }_{i,t+1}$(8) , we also find that the coefficients on the interaction term ($R_t^{US}\times Re{c}_t$) remain strongly significant. Our results do not suffer from the multicollinearity issue.