Capítulo III Consideraciones para la Actualización de una
3.3 Proyecto de Iniciativa Municipal de Desarrollo Local
Having visualized the jump statistics of different characteristic buckets evolving over time, it
is beneficial to study the research question in a more rigorous way. Based on the auto-quote
implementation dates of NYSE stocks, we study the structural change around 2003 and the
heterogeneous jump properties across different characteristic quartiles. The fixed effect panel
data regression with auto-quote and characteristic quartiles/buckets dummy variables, as well
as their interaction terms is used to capture these properties of the jumps. The dummy variable
of auto-quote equals 0 before the introduction of auto-quote, which happened around early 2003,
and 1 after auto-quote. The dummy variable of the characteristic quartile/buckets equals 1 for
the stocks belong to the quartile/buckets, and 0 otherwise. The interaction term of auto-quote
and quartile/group dummies is 1 if and only if the stock belongs to the quartile/group and the
auto-quote has been implemented for the stock as well, otherwise the interaction term is 0.
The monthly aggregated dataset which contains 1379 NYSE stocks from January 1999 to De-
cember 2006. Note that as pointed out before, the NYSE listed stocks are the most stable ones
after 2003, compared with the stocks listed on other exchanges. Hence it is natural to expect
the before-after effects of auto-quote estimated based on NYSE stocks would be weaker than
those based on the whole CRSP universe.
around the event of NYSE’s auto-quote implementation:
Yi,t =αi+λt+βaIi,tauto+ 3 X j=1 βjIi,tj + 3 X j=1
γiIi,tauto∗j+controls , (4.6)
where the independent variable Yi,t is one of the variables that quantify the price instability:
the number of jumps (nJi,t), the number of transient/permanent jumps (ntransi,t /npermi,t ) and the
number of jump days (Di,tJ). The subscripts i and t denote that it is the observation of stock
iat month t. On the right hand side of equation (4.6), Ii,tauto is the indicator of auto-quote for
stock i at montht. The value of Ii,tauto is 0 for the periods prior to auto-quote of stock i, and
1 after auto-quote. Moreover, Ii,tj , j = 0,1,2,3, is the quartile indicator: for stocki in quartile
k, Ii,tj = 0 if j 6= k, and Ii,tj = 1 if j = k. The quartiles are grouped based on one of the
characteristics. In equation (4.6), as the quartile 0 is the base case, I0
i,t is not included in the regression. Regressors Ii,tauto∗j is the interaction terms of Ii,tauto and Ii,tj . Last but not least, to
test the robustness of the results, certain control variables are added in the regressions, such
as market cap, volume, inverse price, bid-ask spread percentage, bipower variation, etc., which
will be reported in tables 4.1 to 4.9.
Table 4.1 reports the fixed effect panel data regressions when the stocks are grouped into market
capitalization quartiles. The model specification is the same as equation (4.6), in which the
dependent variables are the number of jumps nJi,t, the number of permanent jumps npermi,t , the
number of transient jumpsntrani,t , and the number of jump daysDi,tJ . The results are presented
by column: for example, first two columns show the results withnJi,t as the dependent variable,
the next two show the results for ntran
i,t , and so forth.
In table 4.1, the coefficients for relevant variables are all significantly different from zero, and
the results are robust to control variables. The before-and-after effects of the NYSE auto-quote
implementation across different market capitalization quartiles can be evaluated based on these
coefficient estimates. To be more specific, the base case is the number of jumps of market
cap quartile 0 (capt0) before auto-quote, its value is set as zero for convenience (Please refer
to the jump statistics in the previous section for the scale of the baseline, which should be
smaller than the counterpart in the previous section, since we only consider the stocks with
coefficient of Ii,tauto is 5.077), in other words, nJ,capti,t 0 increases because of the auto-quote. For
market quartile 1 (capt1), before auto-quote,nJ,capti,t 1 =βi,tcapt1Ii,tcapt1 = 4.102, however, after auto-
quote,nJ,capti,t 1=βi,tautoIi,tauto+βi,tcapt1Ii,tcapt1+βi,tcapt1∗autoIi,tcapt1∗auto = 5.077+4.102−9.965 =−0.786.
Similarly, for market cap quartile 2 (capt2), before auto-quote, nJ,capti,t 2 equals 3.051, after the
auto-quote, it becomes 5.077 + 3.051−9.693 =−1.565. For market quartile 3 (capt3), before
auto-quote, nJ,capti,t 3 is 1.010 relative to the base case, after the auto-quote, it becomes 5.077 + 1.010−7.420 =−1.333. Similarly, we can obtain the following results for the number of transient
jumps ntransi,t : before auto-quote, the base case capt0 is set as zero, the number of jump days
for capt1, capt2, and capt3 are 2.964, 2.075, 0.496, respectively; after auto-quote, the number
of transient jumps become 4.05, −0.47, −1.08, −1.00 for capt0 to capt3, respectively. For the
convenience of comparison, the before-and-after effects are computed based on the coefficients
estimates, and compiled in table 4.9. Qualitatively speaking, the results are completely aligned
with our previous research: quartilecapt0 has the smallest market capitalization, and it is the
most stable bucket before auto-quote. However, while the other stocks become more stable
after auto-quote, the number of jumps for stocks in quartile capt0 increase considerably after
auto-quote.
Furthermore, by observing the results in the subsequent columns, we find that the majority of
the changes in nJi,t is because of transient jumps. To be more specific, for quartile capt0, nJi,t
increases by 5.077, out of which 4.046 is the transient jumps, only 1.031 is the permanent jump.
Considering that we may misclassify some of the transient-in-nature jumps as permanent, the
conclusion that the price instability deterioration is mainly due to the increased transient jumps
after auto-quote could be stronger than it appears there. Moreover, for quartilescapt1 tocapt3,
the transient jumps decrease considerably, which reflects the fact that most of those mid to large
market cap stocks actually become more stable.
Tables 4.2 to 4.8 report the fixed effect panel regressions with the stock buckets of volume, price,
effective spread, quoted spread, quote-to-trade ratio, market depth, bi-power variation, respec-
tively. For each stock characteristic, quartile/bucket 0 has the lowest value of the corresponding
characteristic, while quartile/bucket 3 has the greatest value of the characteristic. Similarly, as
the results of market cap quartiles, the estimates of relevant coefficients are significantly differ-
npermi,t ,ntrani,t , andDJi,t before and after auto-quote implementation are calculated and compiled
in table 4.9. The calculations are similar as shown previously for the market cap quartiles. To
avoid the repetitive discussion, we will not discuss tables 4.2 to 4.8 one by one, as we did for the
market cap quartiles. Let’s just point out that these results based on NYSE stocks provide us
with very similar empirical results as the previous results based on all CRSP common shares:
the low-priced stocks and the ones with the smallest market cap, trading volume, and market
liquidity, measured by effective/quoted spread and market depth, become more unstable after
2003, especially the drastic increase in the transient jumps. However, these stocks are the most
stable bucket cross-sectionally before 2003. In contrast, the stocks on the other end of the char-
acteristic spectra: the high-priced, large market cap and volume, high market liquidity stocks,
become stable after auto-quote.
The relative tick size (inverse stock price) and quote-to-trade ratio are the two characteristics
that particularly interesting to us, because they are related to the high frequency trading ac-
tivities, as discussed before. Table 4.3 shows the panel data regressions of price quartiles. The
four stocks buckets have the following price ranges: smaller than 5 dollars (prc0), between 5
and 25 dollars (prc1), between 25 and 50 dollars (prc2), and greater than 50 dollars (prc3).
The stock price is inversely related to relative tick size. The low-priced stocks with large rela-
tively tick size have more high frequency activities (Yao and Ye, 2014). The results compiled in
table 4.9 (third panel) indicate that the low-priced stocks (with large relative tick size) become
significantly unstable, especially in terms of the transient jumps; whereas the high-priced stocks
(with small relative tick size) experience less price jumps after 2003. Due to the fact that high
frequency trading increases with the relative tick size, one can conclude that for these low-priced
stocks, high frequency trading causes the price instability.
Furthermore, when the stocks are divided into buckets based on the quote-to-trade ratio, which
is a proxy for high frequency trading, we observe that the stocks in quartileqt ratio3, which has
the greatest quote-to-trade ratio, becomes very unstable after auto-quote. Before auto-quote
implementation, quartile qt ratio3 is the most unstable one (nJ
i,t = 3.624) compared with the other quartiles; after auto-quote implementation, nJi,t increases to 12.598, out of which more
than 10 increased jumps are transient. At the same time, the other quartiles with less high
of stocks with the most HFT activities are more prone to extreme price movements, especially
the transient jumps, while other stocks with less HFT activities become more stable after the
auto-quote implementation.
Theoretically, we can interpret the results as follows: the auto-quote reduces the latency con-
siderably thus algorithmic trading becomes much easier. After the latency reduction, the slow
market makers are more likely to meet high frequency bandits, relative to the chance of them
trading with liquidity traders, as discussed in Menkveld and Zoican (2017). Thus slow market
makers are more likely to be picked off than they are used to be. The slow market makers, who
have greater long-term risk-bearing capability, would be crowded out of the market by HFTs,
especially for the thinly-traded stocks. Compared with stock high volume and market liquidity,
the slow market makers are facing more severe adverse selection problem in the thinly-traded
stocks after the latency deduction. On top of this, the traditional designated market makers are
replaced by endogenous liquidity providers, who may synchronously withdraw from the market
when it is too risky to make the market, such as what has happened during the flash crash.
(Biais and Foucault 2014, Kirilenko, Kyle, Samadi and Tuzun 2017).
Based on the empirical evidence, high frequency trading can be a double-edged sword. While
the stocks with large market cap, volume, and market liquidity, and small relative stick size
have become more stable after the reduction of latency, which is consistent with the previous
empirical studies that suggest HFT is beneficial, the low-priced stocks with small market cap,
volume and liquidity are suffered from large surges of transient jumps that cannot be attributed
to information arrivals. The effects of high frequency trading can be destabilizing, which is in
line with the argument of Menkveld and Zoican (2017).
Interestingly, the bi-power variation, which is the continuous part of the total variation, is not
a relevant characteristic with respect to the jump properties: when stocks are grouped into
bi-power variation buckets, the variables nJi,t, npermi,t , ntransi,t , and DJi,t all uniformly decrease
across different quartiles. We also document that the quartile with smallest short interest and
analyst coverage become more unstable after auto-quote, while the other quartiles have the
fewer number of jumps and jump days, and the majority of the increase is because of transient
jumps. The regression results based on other characteristics, such as book-to-market ratio,
Another thing worth mentioning is that some results for the middle quartiles seem slightly
different from the diagrams before. The reasons are mainly two-fold: first of all, the diagrams
in the previous section are based on all of the stocks in the CRSP universe, which contains the
stocks listed on all the exchanges. However, the panel data regressions are merely based on the
NYSE stocks, due to the limitation that only the dates of NYSE auto-quote implementation
are available to us. Moreover, the diagrams of the previous section are only based on the jump
stocks, while the non-jump stocks are not included. The variables such as the number of jumps
are averaged among all the stocks with jumps, not among all the stocks.