See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/227623900

An empirical study of liquidity dynamics and resistance and support levels

Article  in  Quantitative Finance · December 2010

DOI: 10.1080/14697680902814258 · Source: RePEc

CITATIONS READS
3 120

2 authors:

Carla Gomes Henri Waelbroeck

Independent Researcher Portware, LLC
5 PUBLICATIONS   36 CITATIONS    75 PUBLICATIONS   725 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Volume prediction. Coarse-Grained Hawkes Model View project

Optimal Execution View project

All content following this page was uploaded by Henri Waelbroeck on 18 November 2016.

The user has requested enhancement of the downloaded file.

 An Empirical Study of
Liquidity Dynamics and Resistance and Support Levels

Carla Gomes* and Henri Waelbroeck

Pipeline Financial Group, Inc.
60 East 42nd St. Suite 624
New York, NY 10165

*Corresponding author e-mail: carla.gomes@pipelinefinancial.com

The authors thank Ian Dempsey for valuable comments. The opinions and inferences in
this article are solely those of the authors and do not necessarily reflect the view and
opinion of Pipeline Financial Group or Pipeline Trading Systems. This article is for
informational purposes only and is not soliciting any action.
 2

This study aims to get a depiction of liquidity dynamics with reference to resistance and
support price levels. The bulk of the existing literature on trade clustering focuses on
how trades tend to gather around prices that are psychological barriers or simply round
numbers, suggesting that some prices are indeed psychological references to traders and,
at least temporarily, turn into resistance and support levels. Our analysis looks into two
more general cases, where the presence of resistance and support levels may be either an
indicator of an unusual amount of liquidity available on one side of the market or,
instead, a reference price for traders due to prior accumulation of quantity.
After hypothesizing a set of potential key predictive drivers of volume traded at each time
point, we analyze market data from December 18th to December 28th, 2006 to fit
empirical models explaining number of shares traded. Our findings suggest that, at each
point in time, volume is correlated with quantity traded in the most recent prior instance
of current price, even more so if that event corresponded to a change in price direction
combined with a major clustering of volume.
 3

Introduction

The purposes of this empirical study are to examine clustering of liquidity around
reference price levels and implications for optimal trade execution. The bulk of the
current literature on trade clustering has focused on how trades tend to gather around
prices that are round numbers (Osborne, Niederhoffer, Harris) or that are psychological
barriers (Sonnemans, Donaldson and Kim). For example, Sonnemnans develops an
empirical strategy to test between the odd price hypothesis, according to which humans
attribute more weight to the first digit of each number, and the alternative hypothesis that
investors have target prices for their holdings. His findings suggest that prices can indeed
turn into psychological references to the traders and become resistance and support
levels. Donaldson and Kim find evidence that price levels at multiples of 100 are
psychological barriers to the Dow Jones Industrial Average and act, at least temporarily,
as support and resistance levels.

This study focuses on intraday fluctuations in liquidity, as measured by the number of

shares traded required to push a stock through a certain price level, and the associated
price movements. We do not consider resistance and support levels as asymptotic prices
at which trigger strategists buy or sell a stock (as in Krugman) but, instead, as reference
prices that can be crossed, although perhaps with more difficulty, if the number of shares
in the order flow imbalance is large enough to push the price through such levels (as in
Donaldson and Kim or Bertola and Caballero). The proposed estimation model of
liquidity dynamics is more general than those found in the existing literature since, for
each price level, we consider major prior events and associated quantities as potential
determinants of accumulation of liquidity. Resistance and support levels, in which an
unusual amount of liquidity is available on one side of the market, are a particular case of
reference historical price levels under consideration.

After proposing a set of potential key predictive drivers of liquidity at each price level,
we fit various estimation models explaining its fluctuations in order to estimate and test
the impact of each individual predictor.
 4

Data Coarse-Graining Methods

With the premise that traders will treat prices as equivalent if they differ only by a small
amount, we coarse-grain market data into buckets that include all prints within a price
interval defined by the mean and variance of the spread of each stock: δ= mean(spread)
+ stdev(spread). We measure the spread of each stock as the absolute price difference
between two consecutive prints, except when the prices are the same. In the latter case,
the spread is measured instead as the absolute difference between the current price and
the most recent dissimilar price.

In a preliminary step we exclude from the analysis all odd single prints (n) that are out of
line with adjacent prints i.e. they satisfy both conditions:
| Pn − Pn−1 |> δ AND | Pn +1 − Pn −1 |< δ (1)

For first and last prints in the day, the exclusion criteria are, respectively,
| Pn − Pn+1 |> δ (2a) and | Pn − Pn−1 |> δ (2b)

After filtering, we take the first print of each symbol on each trading day and include in
its bucket all subsequent prints n that satisfy the condition:
n ∈ bucket :| Max{Pn } − Min{Pn } |<= δ (3)

The first print that does not satisfy condition (3) closes the prior bucket and opens a new
bucket.

Price direction is defined as the sign of the price change from one bucket to the next. All
buckets are classified according to the price movement of their first print i.e. a bucket is
classified as an uptick (u) when it is started with a price increase, otherwise it is classified
as a downtick (d).
We then classify each bucket by type of event according to the price movement of its
first print and that of the subsequent bucket: if the bucket starts with an uptick and the
next price change is also an uptick, we classify the event as a double-uptick. Likewise, a
downtick that is followed by a downtick is classified a double-downtick. When price
 5

changes direction from an uptick to a downtick it is classified as a resistance level or, in

the reverse case, as a support level.

An example of data coarse-graining into buckets is displayed below for a stock with
spread+std (δ) equal to 0.04.

Table 1: Data Coarse-Graining

Pre-processing Data Buckets

Mean spread+ standard deviation of spread 0.040
Time Symbol Qty Price Type Side Quantity AvgPrice Event
09:32:11.405 XRM 400 9.68
09:32:43.700 XRM 300 9.62 d S 400 9.680
09:32:47.119 XRM 100 9.61 S
09:32:47.712 XRM 1100 9.6 S
09:38:14.225 XRM 100 9.57 d S 1500 9.605 D
09:39:55.635 XRM 1000 9.69 u B 100 9.570 S
09:50:53.572 XRM 200 9.69 B
09:50:54.059 XRM 100 9.7 B
09:55:04.682 XRM 200 9.7 B
10:09:11.762 XRM 300 9.7 B
10:16:25.310 XRM 100 9.7 B
10:24:25.737 XRM 100 9.7 B
10:29:57.762 XRM 100 9.7 B
10:38:29.590 XRM 100 9.7 B
10:46:46.116 XRM 1000 9.7 B
10:46:46.899 XRM 500 9.7 B
10:57:20.268 XRM 100 9.67 S
11:02:15.635 XRM 200 9.68 B
11:17:02.485 XRM 100 9.68 B
11:25:01.881 XRM 100 9.67 S
11:33:25.608 XRM 100 9.69 B
11:35:25.541 XRM 100 9.69 B
11:38:06.005 XRM 500 9.67 S
11:40:06.031 XRM 100 9.65 d S 4900 9.692 R
11:48:57.165 XRM 100 9.71 u B 100 9.650 S
11:53:09.153 XRM 200 9.68 S
12:11:27.695 XRM 100 9.68 S
12:11:35.103 XRM 100 9.74 u B 400 9.688 U
12:12:26.711 XRM 100 9.71 S
12:12:39.210 XRM 100 9.71 S
12:26:07.330 XRM 100 9.7 S
12:37:08.681 XRM 200 9.72 B
600 9.717
 6

Liquidity Dynamics

In this paper, we introduce a definition of liquidity corresponding to the above coarse-

graining of print data into buckets. Our approach differs from that of other authors who
look at the displayed order book for a measure of volume or number of orders available at
each price limit (eg. Kavajecz, Weber and Rosenow, Zovko and Farmer). Although the
displayed order book gives some information about supply and demand, we would argue
that depth of book information is unreliable and at times deliberately misleading. With
regards to the current interest in algorithmic trading programs, we suggest that using
depth of book information in a linear regression would make an algorithm vulnerable to
precisely the sort of manipulation that predatory traders would engage in to fake an
algorithm into making poor decisions. To avoid such problems, we will only consider
data about printed trades to model the liquidity vector, defining liquidity at price p as the
number of shares that trade in a bucket that covers this price. Our empirical analysis will
then circumvent difficulties stemming from the fact that the liquidity available might be
larger than the observed.

In Figure 1, we show the timeline of arrival of new information. With the price scale
discretized in ticks of size δ, x p is the liquidity at price p defined as the number of shares

that trade at prices within this bucket. The number of shares at price p (k ) , x p (k ) is

discovered at time k, as well as the price direction s(k ) , where s(k ) ∈ {− 1,1} according to
whether the price of the 1st print in new bucket is less or more than the average price of
prior bucket.
p ( k + 1) = p ( k ) + s ( k ) (4)
Based on this information, expectations can be formed about the (yet unobserved)
liquidity at price p ( k + 1) and about the subsequent price direction s ( k + 1) .
 7

Figure 1

Although our focus is not on price directionality, this may be a driver of liquidity at the
new price, so the dynamic model will necessarily explain price as well as liquidity.

We can further consider the liquidity at other price points, as the liquidity that one would
encounter should price reach that level. Liquidity Dynamics is a stochastic dynamic
model for the time evolution of the liquidity vector given the arrival of new information.
As previously discussed in the present study, this new information will be limited to print
data.

In Figure 2, we show the components of the liquidity vector at time k-1, E[ x p | k − 1] ,

which is the expected number of shares at prices around all levels of p given information
available at time k-1. The arrival of new information at time k, prompts an update of the
expected number of shares at each price level p, E[ x p | k ] . In particular, we will assume
 8

that observations at price p(k) do not affect expectations of liquidity at prices other than
p(k) and p(k)+s(k).

Figure 2

We propose a model to explain the persistence in price direction at time k+1 based on the
two prior price movements and both the predicted and actual number of shares traded in
the most recent segment k.
Price movements will be expressed as event types describing the two most recent non-
null price changes at each price level, with a vector of indicator variables at each price
level p, a p (k ) ={U, D, R, S} for, respectively, double uptick, double downtick, resistance

and support:
⎧⎛ 1 + s(l − 1) ⎞⎛ 1 + s (l ) ⎞ ⎛ 1 − s(l − 1) ⎞⎛ 1 − s (l ) ⎞ ⎛ 1 + s(l − 1) ⎞⎛ 1 − s(l ) ⎞
a p (k ) = ⎨⎜ ⎟⎜ ⎟, ⎜ ⎟⎜ ⎟, ⎜ ⎟⎜ ⎟
⎩⎝ 2 ⎠⎝ 2 ⎠ ⎝ 2 ⎠⎝ 2 ⎠ ⎝ 2 ⎠⎝ 2 ⎠
(5)
⎛ 1 − s(l − 1) ⎞⎛ 1 + s (l ) ⎞⎫
,⎜ ⎟⎜ ⎟⎬
⎝ 2 ⎠⎝ 2 ⎠⎭

Where l is most recent occurrence of p at time k, i.e. p(l)=p, k>=l

 9

( s(k + 1).s (k ) + 1)
= θ (α 0 + α1 ( x p ( k ) (k ) − E[ x p ( k ) | k − 1]) + β 1 a p ( k ) (k ) + λ1 E[ x p ( k )+ s ( k ) | k ]
2
+ ε 1 (k + 1))
(6)

Where ε 1k is a stochastic noise term representing exogenous market effects and θ is the
step function
θ ( x < 0) = 0 and θ ( x >= 0) = 1
(7)

We turn next to modeling liquidity. The expectations about liquidity in the next bucket
p(k+1) are affected by the technical event type {U, D, R, S}, the difference between
observed and expected liquidity at p(k), and the observed liquidity being very large as
defined by a threshold liquidity value x*. The actual number of shares at price p at time
(k+1) will deviate from the predicted quantity by a random term:
⎧ E[ x p | k − 1] + ε a (k + 1), p ≠ p (k ), p(k ) + s (k )
⎪
⎪
⎪ r r r
⎪ E[ x p | k − 1] + β 2 .z p (k ) + γ 2 .( x p (k ).a p (k ) − E[ x p | k − 1].a p (k − 1)) +
⎪ r r
x p (k + 1) = ⎨+ λ 2 ⋅ (θ ( x p (k ) − x*).a p (k ) − θ ( x p (k − 1) − x*).a p (k − 1)) + ε b (k + 1), p = p(k ) (8)
⎪
⎪
⎪ r
⎪ E[ x p | k − 1] + β 3 ⋅ a p (k ) + γ 3 .( x p (k ) − E[ x p | k − 1]) ⋅ a p (k )
⎪+ ε (k + 1) p = p(k ) + s(k )
⎩ c

Where:
r r r
z p (k ) = a p (k ) − a p (k − 1) (9)

x * is a pre-determined reference number of shares traded that we set to two times the
average bucket volume for the symbol in question.
x p (k − 1) is number of shares at price p at time k-1 or its most recent observation at that

price.
 10

Data

We analyze market data from NASDAQ and NYSE listed securities for the period
between December 18th and December 28th, 2006, excluding after-hours trading due to
the lower liquidity levels and frequency of trades at those hours. We restrict the universe
of stocks to those with an average volume-weighted price over 1 dollar and an average
daily number of executed shares over 400,000 to ensure a set of comparable stocks where
liquidity dynamics is more likely to occur. The resulting subset includes 1,519 stocks and
over 2 million buckets.

All selected stocks are pooled in the same estimation model. This requires a correction
for heterogeneity, where instead of looking at the number of executed shares, we consider
the adjusted volume in each bucket to the average traded volume per bucket in each
symbol on the same trading day. Additionally, we use a logarithmic transformation to
address fitting problems associated with the asymmetry of the distribution of shares
traded in a given day.

Table 2 displays the frequency of each type of event as well as the average number of
executed shares per event (x), the average number of shares per bucket relative to the
average bucket volume of the stock on each specific date x/avg(x)) and its logarithm
q=Ln(x/avg(x)). The results suggest that price movements are more likely to change
direction from one bucket to another than to persist. When price movements persist, the
number of executed shares is higher on average than at turning points. This finding is
consistent with the fact that turning points reflect one-sided liquidity that was not
exhausted, whereas double upticks and downticks are persistent price movements driven
by a higher than average number of executed shares. Our estimation models explore this
evidence more thoroughly by looking at the fluctuations in volume within each type of
event and testing its correlation with prior clustering at a similar price.
 11

Table 2: Type of Event and Executed Shares

Freq Quantity (x) x/Avg(x) Ln(x/Avg(x))

U 23% 10,400 1.071 -0.528
D 23% 10,158 1.046 -0.551
R 27% 8,838 0.92 -0.814
S 27% 9,276 0.964 -0.763

In an analogous process to the construction of bins, we build the set of explanatory

variables on shares traded and event type considering two prices to be similar when the
module of the difference between the two is smaller than the spread plus its standard
deviation. As hypothesized in our model, it is likely that volume traded in each bucket
will not only be affected by the immediately preceding event and the associated volume,
but also by other prior events at similar historical price levels and the corresponding
number of shares traded. The list of explanatory variables is the following:
● Most recent event type (from the two most recent price changes): Ak-1(T)
where T ε {U, D, R, S} is an indicator variable for double uptick, double downtick,
resistance and support, respectively. For example, Ak-1(U) is equal to 1 if the last event
type was a double uptick and equal to 0 otherwise.
● Quantity traded in the preceding bucket interacted with the respective
event type qxAk-1(T) where T ε {U, D, R, S}. This term allows for quantity traded in the
immediately prior bucket to have a different impact on current number of shares traded
depending on whether that event was a double uptick, a double downtick, a resistance or
a support level.
● Event type around last price similar to current price Aprice(T), where T ε
{U, D, R, S}. Aprice(U) is equal to 1 if event was a double uptick and equal to 0 otherwise.
The reference case is that where current price has not been visited in the past 24 hours.
● Interaction of the quantity traded in most recent bucket around current
price with associated event type q.Aprice(T), where T ε {U, D, R, S}.
● Whether there was an extraordinary number of shares traded around
current price at any instance within the prior 24 hours (BigQ). We consider volume to be
extraordinarily high if quantity is strictly larger than two times the average per bucket for
 12

that symbol on that specific day. The indicator variable for very high volume around
current price is interacted with an indicator variable for event type of its most recent
instance. BigQ.and.A (T) where T ε {U, D, R, S}. For example, BigQ.and.E(U) is strictly
positive if it is the case that current price has been visited within the prior 24 hours and
the last instance of that type of event was a double uptick.
● Whether current price is in the neighborhood of the maximum or
minimum volume-weighted prices of the prior trading day buckets. Max and Min are
indicator variables for each case.
● Whether current price is in the neighborhood of the first and last volume-
weighted price of the prior trading day buckets. Open and Close are indicator variables
for each case.
● Whether current price is in the neighborhood of the whole dollar or 50
cents. Dollar and Halves are indicator variables for each respective case.

Table 3 displays the sample means by current type of price movement. In the case of
indicator variables, the mean corresponds to the frequency of the event represented by the
variable in question.

The findings on the four subsets of the data display many similarities. For example, no
matter what the price movement is, the preceding event is slightly more likely to be a
turning point (resistance or support level) than a double tick (an uptick or downtick) and
the associated quantity is lower on average than that observed on double ticks. At double
upticks and resistance levels, the most recent price event - if it occurred - is most likely to
be found at a resistance level, where the average number of shares traded is lower than at
the double downtick. For double downticks and support levels, the most recent event at
the same price is a support level, where the average number of shares traded is lower than
at the uptick. Over 70% of the market data bins have an average weighted price around
which a significant amount of volume was traded at some point in the previous 24 hours.
Surprisingly, only about 11% of the bins take place at either the open, close, maximum or
minimum price levels of the prior trading day. 15% have a volume weighted price around
either the whole dollar or 50 cents.
 13

Table 3: Table of means by type of event

Event at time k Uptick Downtick Resistance Support

E k-1(U) 0.49 --- 0.45 ---
E k-1(D) --- 0.48 --- 0.44
E k-1(R) --- 0.52 --- 0.56
E k-1(S) 0.51 --- 0.55 ---
q*E k-1(U) -0.215 --- -0.266 ---
q*E k-1(D) --- -0.229 --- -0.271
q*E k-1(R) --- -0.344 --- -0.51
q*E k-1(S) -0.322 --- -0.474 ---
E price(U) --- 0.43 --- 0.379
E price(D) 0.425 0.377
E price(R) 0.473 0.491
E price(S) --- 0.469 --- 0.489
qxE price (U) --- -0.247 --- -0.223
qxE price (D) -0.254 -0.227
qxE price (R) -0.393 -0.455
qxE price (S) --- -0.365 --- -0.428
BigQ.and.E (U) 0.145 0.217 0.152 0.213
BigQ.and.E (D) 0.21 0.139 0.207 0.146
BigQ.and.E (R) 0.202 0.129 0.205 0.137
BigQ.and.E (S) 0.143 0.219 0.153 0.222
Open 0.023 0.022 0.024 0.023
Close 0.035 0.034 0.036 0.036
Min 0.034 0.035 0.035 0.035
Max 0.02 0.018 0.02 0.019
Dollar 0.074 0.075 0.079 0.079
Halves 0.144 0.145 0.151 0.151

These same variables are included in a linear regression as predictors of the logarithm of
the number of shares traded relative to the daily average per bucket for the specific
symbol. For each particular type of event, only two immediately prior types of event are
possible. For example, a double uptick can only be preceded by another double uptick or
 14

by a support level. For this reason, and since we are including a constant in the model,
only one of the events is represented by the indicator variable associated with prior price
movement. The estimated coefficient corresponds to the differential impact of the event
represented by the indicator variables as compared to the chosen reference group. As for
the interaction of event type with associated quantity, we should be able to identify both
interactions of lagged indicator variables with number of shares because the latter are not
the same for both subgroups.

In the estimation, we calculate the Huber/White “sandwich” estimators of variance,

which are robust in the sense that they give accurate assessments of the sample-to-sample
variability of the parameter estimates even when the model is misspecified, such as when
there are minor problems about normality, heteroscedasticity, or some observations that
exhibit large residuals.

Results

Tables 4 through 7 display the results of the least squares estimation models. The
explained variable is the logarithm of the number of shares traded in each bucket relative
to the daily average for the respective symbol [q=Ln(x/avg(x)]. All regressors referring to
the number of shares traded are divided by the daily average per bucket and transformed
to logarithms as well. For example, if number of shares in the immediately prior bucket is
x0, we take Ln (x0/avg(x)).

Under this specification, the estimated coefficients associated with each indicator
variable represent the impact of the respective event in terms of percent change in current
number of shares relative to the daily bucket average.

To estimate the marginal impact of the number of shares from a prior event such as, for
example, a support level, we set Et-1(S) equal 1 for all cases where prior event is a support
level and equal to 0 for all other cases. If the number of shares at the prior bucket is equal
 15

to x1 we multiply Et-1(S) by Ln(x1/avg(x)). The estimated coefficient on Et-1(S)xq

represents the percent change in current number of shares relative to the daily bucket
average associated with a 1% change in x at the prior support level.

The estimation results suggest that virtually all proposed determinants of volume are
statistically significant regardless of the type of current price movement. The impact of
the number of shares traded in the immediately prior bucket has a statistically significant
effect on current volume. This effect is higher if the prior event was a double tick rather
than a turning point and if current event is either a resistance or support level. When the
prior event is a double tick, the estimates range between .27 and .33, whereas in the cases
where a prior event was a turning point, estimates of the marginal impact of number of
shares vary between .16 and .28. Additionally, the estimated average impact of a turning
point relative to a double tick is negative ranging between -0.164 to -0.032.

The impact of shares traded in the most recent bucket at a price similar to current price
also varies with type of prior event. For double upticks and resistance levels, the average
impact of trades at resistance levels is the largest (.22 and .31), whereas for double
downticks, trades at support levels have, on average, the largest impact (.2). A 1%
increase in number of shares per bucket relative to the daily average is associated with an
increase in current volume between .15 and .2 %.

The estimated impact of a major trade clustering at a prior price level close to the current
price is always significant, but larger when the execution occurred in the neighborhood of
a resistance or support price level. If the current event is a double uptick or a resistance
level, a major clustering has the highest impact if it occurred at a resistance level (.16 and
.2). If the current event is a double downtick or a support level, a major clustering has the
highest impact if it occurred at a support level (.3).

The reference prices of the prior trading day are not always statistically significant and,
when significant, do not seem to be associated with a larger number of shares traded.
 16

As in the previous literature, we find that volume is higher at prices around multiples of
50 cents (halves), with point estimates ranging between .05 and .09, and even higher
around the whole dollar, which is the combined effect of a multiple of 50 cents and the
additional effect of the round dollar. The resulting total estimate of the effect of the whole
dollar ranges between .08 and .14.

Table 4: Linear Regression. Explained variable: qk=Ln [xk/avg(xk)] – Double Upticks

(1) (2)

Coeff. SE Coeff. SE

E k-1 (S) -0.032 0.003 -0.014 0.003

qxE k-1 (U) 0.272 0.002 0.309 0.002
qxE k-1 (S) 0.166 0.002 0.214 0.002
E price (D) 0.087 0.009 --- ---
E price (R) 0.218 0.009 --- ---
qxE price (D) 0.2 0.002 --- ---
qxE price (R) 0.15 0.002 --- ---
BigQ.and.E (U) 0.14 0.005 0.163 0.005
BigQ.and.E (D) 0.141 0.005 0.24 0.005
BigQ.and.E (R) 0.196 0.005 0.318 0.005
BigQ.and.E (S) 0.089 0.005 0.111 0.005
Open 0.001 0.011 -0.002 0.011
Close -0.032 0.009 -0.037 0.009
Max 0.039 0.011 0.041 0.012
Min -0.035 0.009 -0.036 0.009
Dollar 0.062 0.008 0.066 0.008
Halves 0.075 0.006 0.086 0.006
Constant -0.536 0.008 -0.556 0.004
R2 0.13 0.093
N 475,412
Note 1: Coeff. is point estimate and SE is standard error
Note 2: Italics represent estimates that are not statistically significant at the 5% significance level
 17

Table 5: Linear Regression. Explained: qk=Ln[xk/avg(xk)] DoubleDownticks

(3) (4)

Coeff. SE Coeff. SE

E k-1 (R) -0.034 0.004 -0.017 0.004

qxE k-1 (D) 0.267 0.002 0.302 0.002
qxE k-1 (R) 0.161 0.002 0.21 0.002
E price (U) 0.062 0.009 --- ---
E price (S) 0.197 0.009 --- ---
qxE price (U) 0.197 0.002 --- ---
qxE price (S) 0.145 0.012 --- ---
BigQ.and.E (U) 0.145 0.005 0.24 0.005
BigQ.and.E (D) 0.128 0.005 0.15 0.005
BigQ.and.E (R) 0.103 0.005 0.127 0.005
BigQ.and.E (S) 0.189 0.005 0.304 0.005
Open -0.017 0.011 -0.021 0.011
Close -0.017 0.009 -0.025 0.009
Max -0.013 0.012 -0.01 0.012
Min -0.001 0.009 -0.007 0.009
Dollar 0.059 0.008 0.066 0.008
Halves 0.081 0.006 0.089 0.006
Constant -0.544 0.008 -0.573 0.004
R2 0.123 0.089
N 471,013
Note 1: Coeff. is point estimate and SE is standard error
Note 2: Italics represent estimates that are not statistically significant at the 5% significance level
 18

Table 6: Linear Regression. Explained variable: qk=Ln[xk/avg(xk)] - Resistance

(5) (6)

Coeff. SE Coeff. SE

E k-1 (S) -0.162 0.004 -0.128 0.004

qxE k-1 (U) 0.33 0.002 0.365 0.002
qxE k-1 (S) 0.281 0.002 0.337 0.002
E price (D) 0.228 0.01 --- ---
E price (R) 0.303 0.01 --- ---
qxE price (D) 0.168 0.003 --- ---
qxE price (R) 0.199 0.002 --- ---
BigQ.and.E (U) 0.126 0.005 0.174 0.005
BigQ.and.E (D) 0.119 0.005 0.224 0.005
BigQ.and.E (R) 0.159 0.005 0.295 0.005
BigQ.and.E (S) 0.105 0.005 0.147 0.005
Open -0.038 0.011 -0.041 0.012
Close -0.032 0.009 -0.041 0.009
Max -0.008 0.012 0.008 0.012
Min -0.023 0.009 -0.026 0.009
Dollar 0.033 0.008 0.037 0.008
Halves 0.061 0.006 0.071 0.006
Constant -0.715 0.01 -0.653 0.004
R2 0.157 0.13
N 551,917
Note 1: Coeff. is point estimate and SE is standard error
Note 2: Italics represent estimates that are not statistically significant at the 5% significance level .
 19

Table 7: Linear Regression. Explained variable: qk=Ln[xk/avg(xk)] -Support

(7) (8)

Coeff. SE Coeff. SE

E k-1 (R) -0.164 0.004 -0.133 0.004

qxE k-1 (D) 0.321 0.002 0.358 0.002
qxE k-1 (R) 0.279 0.002 0.335 0.002
E price (U) 0.205 0.011 --- ---
E price (S) 0.198 0.002 --- ---
qxE price (U) 0.174 0.003 --- ---
qxE price (S) 0.198 0.002 --- ---
BigQ.and.E (U) 0.124 0.005 0.228 0.005
BigQ.and.E (D) 0.125 0.005 0.172 0.005
BigQ.and.E (R) 0.123 0.005 0.168 0.005
BigQ.and.E (S) 0.163 0.005 0.295 0.005
Open -0.011 0.011 -0.015 0.011
Close -0.021 0.009 -0.028 0.009
Max -0.003 0.012 -0.004 0.013
Min -0.019 0.009 -0.022 0.009
Dollar 0.027 0.008 0.03 0.008
Halves 0.05 0.006 0.058 0.006
Constant -0.639 0.01 -0.593 0.004
R2 0.157 0.129
N 551,174
Note 1: Coeff. is point estimate and SE is standard error
Note 2: Italics represent estimates that are not statistically significant at the 5% significance level

Tables 8 and 9 display the estimation results for empirical models that account for
possible censoring at resistance and support levels. Assuming a normally distributed error
term, the estimates are obtained through a maximum likelihood estimator where all
observations at resistance and support levels are considered to be right-censored i.e. the
number of shares available at the given resistance or support price level is greater or
equal to the number of shares executed. The rationale for this strategy is that while we
want to estimate liquidity available in each bucket, turning points in price direction most
 20

likely correspond to a peak in liquidity available at the resistance or support price level
which may not be observed in its entirety because it was not exhausted.

In a similar way, the number of shares at prior resistance and support levels, which are
included in the estimation models as explanatory variables, may not reflect the total
liquidity available at each price level. Because of that, the estimated impact of the
number of shares at these turning points may be lower since the quantity that is observed
possibly does not reflect the whole underlying liquidity available at the price point in
question.

In order to identify the parameters of the full distribution, censored observations have to
be pooled with non-censored observations. To address this issue, and to account for
possible differences related to whether the market is moving up or down, we estimate two
separate models: one for number of shares traded at double upticks or resistance levels
and another for number of shares traded at double downticks or support levels.

The estimates of the intercepts of these models are much higher than those of the linear
regression, whereas some of the estimates of the impact of the most recent events at
similar historical prices are lower. Nevertheless, all conclusions regarding the sign and
relative importance of predictors still hold.

The number of executed shares at the prior event around the current price is, again,
statistically significant. The estimated impact of the prior bucket ranges between
-0.109 and 0.181.

The impact of a 1% change in the number of shares in prior trades at a price similar to the
current one ranges between 0.157 and 0.194. The estimates are always slightly higher for
shares traded at resistance levels when impacting volume at double upticks or resistance
levels and at support levels when the current event is a double downtick or a support
level. Similarly, the point estimates of the impact of a prior major trade clustering at a
price level around current price is slightly higher when prior trades occur at resistance
 21

levels and current bucket is a double uptick or resistance level (0.125), or when prior
trades occur at a support level and current price event is a double downtick or a support
level (0.134).

As in the least squares regression model, we do not find a large impact of reference prices
from the prior trading day.

Volume is significantly higher at prices around multiples of 50 cents (halves), with point
estimates of .047 and .048 and even higher around the whole dollar, with estimates of
.077 and .07.
An increase in quantity traded in the immediately prior bucket is associated with a
statistically significant increase in current volume and this effect is, on average, lower
(0.22 and 0.23%) if prior event is a turning point instead of a double tick.
 22

Table 8: Censored Regression. Explained variable: qk=Ln [xk/avg(xk)] Double

Uptick or Resistance

(1) (2)

Coeff. SE Coeff. SE

E k-1 (S) -0.225 0.004 -0.151 0.004

qxE k-1 (U) 0.392 0.002 0.415 0.002
qxE k-1 (S) 0.319 0.002 0.376 0.002
E price (D) 0.141 0.01 --- ---
E price (R) 0.181 0.01 --- ---
qxE price (D) 0.123 0.007 --- ---
qxE price (R) 0.119 0.006 --- ---
BigQ.and.E (U) 0.09 0.005 0.103 0.005
BigQ.and.E (D) 0.09 0.005 0.181 0.005
BigQ.and.E (R) 0.125 0.005 0.242 0.005
BigQ.and.E (S) 0.119 0.005 0.057 0.005
Open -0.036 0.011 0.04 0.009
Close -0.035 0.009 -0.045 0.009
Max -0.022 0.012 0.02 0.012
Min -0.027 0.009 -0.033 0.009
Dollar 0.022 0.008 0.024 0.009
Halves 0.048 0.006 0.058 0.006
Constant 0.297 0.009 0.239 0.004
N 1,027,309
Note 1: Coeff. is point estimate and SE is standard error
Note 2: Italics represent estimates that are not statistically significant at the 5% significance level
 23

Table 9: Censored Regression. Explained variable: qk=Ln [xk/avg(xk)] Double

Downtick or Support

(1) (2)

Coeff. SE Coeff. SE

E k-1 (R) -0.222 0.004 -0.151 0.004

qxE k-1 (D) 0.377 0.002 0.401 0.002
qxE k-1 (R) 0.319 0.002 0.375 0.002
E price (U) 0.109 0.01 --- ---
E price (S) 0.137 0.01 --- ---
qxE price (U) 0.163 0.012 --- ---
qxE price (S) 0.191 0.002 --- ---
BigQxE (U) 0.101 0.005 0.191 0.005
BigQxE (D) 0.09 0.005 0.102 0.005
BigQxE (R) 0.073 0.005 0.084 0.005
BigQxE (S) 0.134 0.005 0.244 0.005
Open -0.013 0.011 -0.017 0.011
Close -0.031 0.009 -0.037 0.009
Max -0.023 0.012 -0.025 0.012
Min -0.011 0.009 -0.016 0.009
Dollar 0.02 0.008 0.023 0.008
Halves 0.047 0.006 0.054 0.006
Constant 0.335 0.009 0.251 0.004
N 1,022,187
Note 1: Coeff. is point estimate and SE is standard error
Note 2: Italics represent estimates that are not statistically significant at the 5% significance level

Tables 10 and 11 display the results of logistic regressions predicting the price
movements. The explained variable is the probability of reversal of price direction. The
set of explanatory variables includes immediately prior price movements, deviation of
actual quantity from what was predicted qk - E[qk| k-1] and expected quantity E[qk+1| k] at
the new price.
 24

The estimates displayed in the tables are odds ratios. In the case of indicator variables,
the odds ratio compares the probability of a price reversal when the variable is equal to 1
versus the case when it is equal to 0. For continuous variables, the odds ratio represents
the increased likelihood of a price change when the explained variable increases by one
unit. The logistic regression does not have an equivalent to the R-squared that is found in
a least squares regression. In the tables below, we display instead the McFadden’s pseudo
R2, obtained by computing 1 – ratio of log likelihoods between the full model and the
constant-only model. The McFadden's R2 can be as low as 0 but can never equal 1. The
R-squared for the parallel linear regression is equal to 0.009.

The results show a very slight correlation in price movements, since upticks are slightly
more likely to be followed by upticks than downticks. A reversal in price direction is
slightly more likely to be followed by another change in price direction.

We find that the higher the expected number of shares to be executed, the more likely is a
price movement to persist. To a lesser extent, a price movement is also more likely to
persist if there is a greater excess over the predicted number of shares.

Table 10: Logistic Regression Explained: P[Reverse]

(3)
ALL
O.R. SE
Up 0.99 0.003
Change 1.095 0.003
E[qk+1| k] 0.762 0.003
qk - E[qk| k-1] 0.967 0.002
N 2,038,970
Pseudo R2 0.0067

Table 11 displays estimating models for the probability of a price reversal for the subset
of upticks separated from the downticks. The conclusions from the pooled regression still
hold.
 25

Table 11: Logistic Regression Explained: P[Reverse]

(1) (2)
U/R D/S
O.R. SE O.R. SE
Change 1.092 0.005 1.098 0.005
E[qk+1| k] 0.765 0.004 0.759 0.004
qk - E[qk| k-1] 0.966 0.002 0.967 0.002
N 1,021,856 1,017,114
Pseudo R 2
0.0065 0.0068

In Figure 3 we check for arbitrage opportunities. First, we subdivide the data according to
the ratio of probabilities associated with negative and positive returns. Then, for each
segment, we calculate the respective ratio of positive to negative returns. The no arbitrage
condition, requires that gains from higher positive returns be compensated by lower
associated probabilities . This way the expected profit remains equal to zero and there are
no opportunities for arbitrage. Our findings suggest that is not the case since, although
the ratio of returns is equal to 1 when the ratio of probabilities is also close to 1, as
positive returns increase the associated probability increases as well.

In Figure 4 we check for the size of the potential expected profit. We find that it rises as
the uncertainty about the probability of reversal lowers but with exception of two outliers
caused by very thin data at the extremes, the maximum profit is 11 basis points.

To check the robusteness of all results we re-ran the analysis for the 12 trading day period
between January 16 and January 31. The estimates were virtually the same.
 26

Figure 3

Figure 4
 27

Discussion

This paper estimates a general model of liquidity dynamics where, at each price level,
major prior events and associated number of shares traded are potential determinants of
current liquidity.

In general, we find that there is a strong correlation between the numbers of executed
shares at different points in time as long as they occur around the same price. These
findings are consistent with the hypothesis that price levels at which there is
accumulation of volume become future reference points at which traders are willing to
buy or sell stocks. In addition, when price direction changes, the number of shares traded
is an even stronger predictor of subsequent volume at that same price. We interpret this
result as evidence that a resistance (support) price might be an indicator of a significant
amount of liquidity on the supply (demand) side.

In the empirical model for price direction, we find a strong impact of expected number of
shares traded on the subsequent price movement: the higher the number of shares
executed, the more likely is the price movement to persist. The associated price
predictability corresponds to a short term profit opportunity of 5-10 basis points, in line
with reasonable estimates of trading costs and adverse selection costs for a short-term
statistical arbitrage strategy.

References

Bertola G. and Caballero R. J. (1992) “Target Zones and Realignments” The American
Economic Review 82(3) pp 520-536.

Donaldson, R.G. and Kim, H. Y. (1993) “Price Barriers in the Dow Jones Industrial
Average” The Journal of Financial and Quantitative Analysis 28(3) pp 313-330.
 28

Harris L. (1991) “Stock Price Clustering and Discreteness” Review of Financial Studies
4(3), 389 - 415.

Kavajecz K. A. (1999) “A Specialist's Quoted Depth and the Limit Order Book” The
Journal of Finance 54(2) , Pages: 747-771.

Krugman P. (1987) “Trigger Strategies and Price Dynamics in Equity and Foreign
Exchange Markets” NBER Working Paper Series Vol. wp. 2459 pp-.

Niederoffer V. (1966) “A New Look at Clustering of Market Prices” Journal of Business

39(2), 309-313.

Osborne, M.F.M. (1965) “The Dynamics of Stock Trading” Econometrica 33(1), 88-113.

Sonnemans J. (2003) “Price clustering and natural resistance points in the Dutch stock
market: a natural experiment”. Tinbergen Institute Discussion Paper No. 03-043/1.
Available at SSRN: http://ssrn.com/abstract=417321 or DOI: 10.2139/ssrn.417321.

Weber P. and Rosenow B. (2005) “Order Book Approach to Price Impact” Quantitative
Finance 5(4): pp: 357 – 364

Zovko I. and Farmer J.D. (2002) “The Power of Patience: a Behavioural Regularity in
Limit-Order Placement” Quantitative Finance 2(5), pp: 387 – 392

View publication stats