Page 182 - 34-2

Order Choices, Order Execution Quality and Trading Volume: Evidence from Reductions in the Call Auction
Interval

observations being 5-minute frequency intervals (thus all variables are labelled with a
subscript of t), and market participants being separated into individual investors and
institutional investors. For order choices, y includes the proportion of order cancellations
t
(rCn ), most aggressive orders (rMg ), and larger orders (rSz ); the measures of order
t
t
t
aggressiveness (G ) and trend-following (Tr ); the number of new order submissions (N ).
t
t
t
For order execution quality of uninformed traders (represented by individual investors), y t
includes the probability of limit-order execution within the present call (rFc ) and present
t
day (rF ); measures of bargain (Sur ), ‘picking-off’ risk (PK ), ‘non-execution’ risk (NE ),
t
t
t
t
and comprehensive risk of limit-orders (MX ). For individual stock turnover, y includes
t
t
the change rate of stock turnover (gTV ) and the trade proportion of institutional investors
t
(rIns ).
t
It is worth noting that though the Difference-in-difference (DID) approach is
frequently adopted to control for short-run time trend of y so as to assess the “true” impact
t
of an event, it is difficult to utilize the DID approach for the present study because of the
lack of a “control group”, with the reductions of auction intervals being implemented in
all listed firms on the TWSE. To address this issue, I utilize the “daily time trend” measure
(denoted as ẑ ) to control for the potential common factors sharing the daily time trend or
t
market condition with each y measurement. Fortunately, the opening call auction (from 8:30
t
to 9:00) on the TWSE remains a single 30-minute call auction during the sample period,
providing us an opportunity to take out the information regarding such common factors.

Accordingly, I estimate the daily values of ẑ for each of y with the cross-day auxiliary
t
t
regression based on observations computed from the opening calls. These “daily time
trend” measures served as additional control variables in our primary regression models.

Last but not least, some regression models discussed in our work are unsuitable
to estimate their slope coefficients by means of traditional Least Square (LS) method.
More concretely, considering that rCn , rMg , rSz , rFc , rF , rIns are proportions and Sur t
t
t
t
t
t
t
has non-negative continuous values, these models are designed as Tobit regressions as
suggested by Chiu, Chung, and Wang (2017). Besides, considering that N has counted
t
data observations, its model is designed as a negative binominal regression as suggested
by Tseng (2019). The above cases are no longer linear and their slope coefficients are
estimated by means of the Maximum Likelihood (MLE) method. After controlling for
the potential disturbances (listed in Table 6), I focus on the estimates of coefficients on

174