Index Construction
To construct the index, we sum the daily word count for all the keywords, defined in this paper as EPU, and thereafter, we run a heteroskedasticity-consistent ordinary least squares (OLS) regression of EPU on day-of-the-week dummy variables to accommodate the possible distinct variations in public sentiments in the 5 days of the week.[1] To avoid dummy trap, we exclude the Friday dummy while the resulting constant and residuals from the OLS model are added to adjust the data for day-of-the-week effects:
; . (1)
where EPU is the daily sum of word count for each of the keywords; is for the week-day dummy variables for 4 days (for a 5-day week) since one dummy has to be excluded to avoid dummy trap; is the regression constant while is the disturbance term. The (day-of-the-week) adjusted EPU () is obtained as . In the final step, we compute the index as:
(2)
Table 1: Some useful highlights of the EPU index for Nigeria
|
|
PUNCH |
TRIBUNE |
PMNEWS |
THIS DAY |
THE GUARDIAN |
BUSINESS DAY |
DAILY TRUST |
LEADERSHIP |
|
No. of news articles downloaded |
361,170 |
199,569 |
234,409 |
230,135 |
163,786 |
236,057 |
435,430 |
1,707 |
|
No. of news articles containing at least a keyword |
94,312 |
69,651 |
52,662 |
90,712 |
151,767 |
92,956 |
112,028 |
1,580 |
[1] This is an important feature of daily frequency variables that connect with investors and practitioners which has to be incorporated into the construction of the index to make it more realistic (see for example, Narayan, Iyke, and Sharma, 2021). This phenomenon has been largely validated in the finance literature where it is assumed that the same information may have differing effects during the days of the week as the average daily return of the market is not the same for all the days of the week (see Drogalas et al., 2007).