The price impact of informed and uninformed trades : an Alex Vermogensbank case study

(1)

The price impact of

informed and uninformed

trades.

An Alex Vermogensbank Case study

Dhr. P.F.A. Tuijp Msc Master Thesis Finance

07-07-2014

(2)

(3)

2

Little is known about the empirical magnitude of the price impact of informed versus uninformed trades by institutional investors. This thesis uses a proprietary dataset to shed light on this matter. Researchers assume institutional traders possess superior information regarding the market. The literature on the price impact of trades is extensive. The most common conclusion is only informed traders impact prices permanently. Whether the assumption that institutional traders are informed is accurate, still remains a question to be answered.

To get an answer to these questions this research aims to determine which type of trade has what price impact. It aims to determine the informational part of a price change resulting from an (un)informed trade. By using a proprietary dataset of an institutional investor, in which it is known which trades are informed and which trades are uninformed, I aim to address the informational content of price impact in more detail. More specifically, to what extent does the price impact of informed trades differ from the impact of uninformed trades?

Former research to determine the price impact of a trade is extensive. Kyle (1985) and Glosten and Milgrom (1985) wrote the first main papers in which price impacts resulting from asymmetric information were modelled. Hasbrouck (1991) was the first to find evidence the informational content of a trade influenced the price of a trade. Multiple researchers, like Barclay and Warner (1993), Chakravarty (2001) and Chan and Lakonishok (1993), conclude institutional traders are informed traders. The more recent strands of literature of debate on whether the price impact function is linear or concave. At the moment the conclusions of Lillo, Doyne Farmer and Mantegna (2003), Toth, Lempérière, Deremble, de Lataillade, Kockelkoren and Bouchaud (2011) and Mastromatteo, Toth and Bouchaud (2013) the price impact curve is concave are leading. This paper investigates whether an institutional investor its trades move price. It also aims to differentiate between informed and uninformed trades of an institutional investor. An institutional investor could alter its trading strategies if they know their uninformed trades will impact prices, just because they are an institutional investor.

(5)

4 The paper uses the first three months of 2014 as sample period. The sample consists of stock listed in the Eurostoxx-50 index. The proprietary dataset contains the informed and uninformed trades of Alex Vermogensbeheer and the market in these stocks. In order to determine the price impact the paper uses the method of Chan and Lakonishok (1993). Furthermore a number of panel regressions of algorithmic and redemption trade dummies on price changes aim to differentiate between the price impacts of informed and uninformed trades. The first two sets of panel regressions aim to determine the price impact of the individual trades on prices within a short and a long time interval. The third set of panel regressions aims to determine the price impact of the individual trades on the subsequent trades. The goal of the third panel regression is to determine the price impact of blocks of trades in the fifteen minute time intervals. The fourth and fifth panel regression aim to determine the price impacts of blocks of trades on the price change in 15 minute time intervals and price change from daily open to closing price respectively.

The results of the method of Chan and Lakonishok (1993) in this paper provide evidence in favor of the short run liquidity hypothesis. The sales experience a price decrease in the time leading up to the trade and the purchases experience a price increase leading up to the time of the trade, because of liquidity constraints. According to the evidence provided by this method, the algorithmic trades do not contain any informational content. These results in combination with the performance of Alex Vermogensbeheer with respect to the benchmark indicate the ability to pick the right stock did not exceed the market. The results on an individual trade level do indicate algorithmic sales and purchases exhibit an informational price impact on the next trade. For sales the informational price impact is to 0.0023% above the average market impact and for the purchases it is 0.0036% above the average market impact. The evidence on blocks of trades is not in favor of the hypotheses that algorithmic trades exhibit informational price impact.

These mixed results do not support a clear conclusion. The individual algorithmic sales and purchases and the blocks of trades of Alex Vermogensbeheer do not influence prices in a short and long interval or in 15 minute and daily intervals. This result could have different reasons. On the one hand the trades could be perfectly camouflaged among uninformed trades as the theory of Glosten and Milgrom (1985) suggests. The second reason could Alex Vermogensbeheer performs the trades better than the average price of the intervals.

(6)

5 However, the regressions on the volume weighted average price do not confirm this theory. A third line of reasoning is Alex Vermogensbeheer is not informed. This would indicate the stock picking of Alex Vermogensbeheer in the sample period did not resemble the informed abilities of an informed trader. The results of the management of the assets by Alex Vermogensbeheer, summarizes in table 3, show they underperform with respect to the benchmark.

Two other possible explanations this paper offers are more general for models that use technical analysis. The Euro Stoxx 50 index moved “horizontal” during the time period, so the results could also indicate technical analysis in general cannot identify the right time to buy or sell a certain stock. This means an asset manager should use other methods to time their trades. If trades do not contain an informational content, asset managers may execute algorithmic trades in the same way as redemption trades. This means they do not have to camouflage their trades amongst trades without informational content.

For further research it would be interesting to investigate the informational content over a longer time period. Specifically, if the sample period contains multiple different periods of bull, bear and stable markets to differentiate between the informational price impacts in the different market conditions. Furthermore, like Toth et al. (2011), suggest it would be interesting to measure the order flow of the order book and try to measure the price impact function of institutional investors. The price impact function would show whether the blocks of trades of Alex Vermogensbeheer have a decreasing price impact over increasing trade size. If the price impact function is concave, it would be useful to use another model to estimate the informational content of the price impact.

In the following section the previous literature will be reviewed. The third section will elaborate on the dataset and on the methods used. The fourth section contains the analysis and the results. The sixth section contains the discussion and the conclusion.

2. Related literature

Many studies focus on the price impact of informed and uninformed trades. The review will be structured chronologically. The first paragraph focuses on efficient markets. Subsequently the second paragraph reviews the early research on informed and uninformed trading and asymmetric information. The literature in the third paragraph tries to identify

(7)

6 the informed trader and its price impact in equity and option markets. The fourth and final part of the review assesses the literature on the shape off the price impact function.

2.1 Efficient markets

According to Fama (1970) markets are strong form efficient if prices reflect all available information. The theory of efficient markets implies a trader with superior private information is not able to profit from the information, because it is already incorporated in market prices. However, empirical results do not show support for weak form efficient markets. For example, Basu (1977) shows information in P/E-ratios is not fully incorporated in prices and an excess return may be earned in the markets. After establishing doubt about whether markets are strong form efficient, multiple strands of literature try to determine the reasons why market prices do not reflect all the information.

2.2 Asymmetric information and the informational price impact

Asymmetric information implies a trader has more information than the market maker. In his paper Kyle (1985) asks the question how this information of an insider is incorporated in prices. First he models uninformed traders to provide camouflage for informed traders, this enables informed traders to profit from their private information. From the model he concludes private information gradually and linearly incorporates in prices and at the end of trading the price reflects all information. Glosten and Milgrom (1985) define an informed trader as a trader with slightly more information than the market maker. They define an uninformed trader as a trader who only trades for liquidity purposes. The market maker faces an adverse selection problem, because informed traders will pick off the market maker. They lose on trades with informed traders. He posts a bid-ask spread to recover the informational losses by trading with the liquidity traders. From their model, they conclude only informed traders push prices.

Hasbrouck (1991) argues in a market with agents with different information the trades of these agents contain information about the price of the security. Hasbrouck (1991) models the informational content of a trade as the impact of the trade innovation of a series of trades in a vector autoregressive model. He concludes the price impact of a trade depends on the informational content of the trade, the size of the trade and the prevailing spread at the time of the trade. He also found asymmetric information influences are more significant for price change, if firms have smaller market values.

(8)

7 2.3 Identifying informed traders

The early research on informed trading assumes informed traders exist and have superior information. The following related research tries to identify the informed traders in the market and how and when they trade. For example Barclay and Warner (1993) tested the cumulative price change in each trade size category of the NYSE. They hypothesize informed traders trade in the medium sized trade category to hide their trades amongst other trades. These trades exhibit the largest price impact. They call it the stealth trading hypothesis. Informed traders do not use too large trade sizes, because market makers notice the trades are informed and adjust prices accordingly. Informed traders do not use too small trade sizes, because transaction costs are too high. Barclay and Warner (1993) find confirmation for the stealth trading hypothesis, i.e. medium sized trades are responsible for the largest cumulative price impact.

Chakravarty (2001) links the price impact of a specific trade size to an investor category; individual investors or institutional investors. He defines individual investors as uninformed traders and institutional investors as informed traders. He confirms the stealth trading hypothesis of Barclay and Warner (1993) that price movements are attributable to medium sized trades and institutional investors, i.e. informed traders, account for these trades.

Chan and Lakonishok (1993) use a dataset of 37 money management firms with more than one million transactions. The money management firms labeled each trade as a purchase or a sell, therefore Chan and Lakonishok (1993) could determine the price change of the individual transactions. They offer three possible explanations for price changes. They argue difficulties in finding a counter party to trade results in short run liquidity costs. These costs result in temporary price changes. Their second explanation is the imperfect substitution hypothesis, i.e. no perfect substitute exists for a stock, and thus the seller should offer a discount to induce buyers to buy the shares. This hypothesis would be consistent with permanent price changes. Chan and Lakonishok (1993) also expect permanent price changes if the trades are convey private information. The informational price impact depends on the side of the informed trader.

They conclude the buy transactions of institutional traders are associated with a price increase of 0.34% and the sell transactions with a decrease of 0.04%. According to

(9)

8 them the difference between buy and sell impact is caused by the informational difference between the two. The choice to buy a stock from the whole universe of securities could imply a larger informational advantage over selling a stock from your own portfolio. Chan and Lakonishok assume all trades of institutional traders contain an informational content. The evidence does not show whether all trades of institutional traders exhibit the same price impact.

Chakravarty, Gullen and Mayhew (2004) use an adjusted version of the Hasbrouck-model to determine whether informed traders trade in the option markets. They conclude on average 17% of the price discovery occurs in the option markets. Suggesting informed traders both trade in stock and option markets. Anand and Chakravarty (2007) also use the Hasbrouck-model to investigate whether informed traders in option markets camouflage their trades using certain trade sizes. They conclude informed traders in the option market prefer to cut their large trades into medium sized trades for high trading volume contracts, i.e. liquid contracts. For illiquid contracts they prefer to use small sized trades. These results imply informed traders account for an informational and a volume effect in the choice of their trades.

Gomes and Waelbroeck (2013) investigate price reversion after executing informed and uninformed meta-orders. From a sample of institutional trades they define informed trades as non-cash flow trades and uninformed trades as trades for cash flow reasons. They find prices revert partially after executing informed meta-orders. They find prices revert completely after executing an uninformed order. They conclude informed meta-orders impact prices through their informational content, defined as the remaining “permanent” price impact after a meta-order. They conclude uninformed meta-orders do not exhibit informational price impact.

Blau, Van Ness and Van Ness (2009) found overall trading volume is larger at the beginning and the end of a trading day. Furthermore, these researchers found the price impact function on time of day of large trades has the shape of a parabola and the price impact function on time of day of smaller trades has the shape of a hyperbole. They argue smaller trades exhibit larger price impact during the middle of the day, because informed traders disguise their trades in smaller fragments. They argue larger trades exhibit larger

(10)

9 price impact during the beginning and the end of the day, because the informed traders are able to trade in larger fragments during times of high trading volume.

To conclude, researchers find institutional traders cut their trades in medium sized transactions to camouflage their trades among the market. Besides they use the option markets and use the beginning and the ending of a trading as convenient places and moments to execute their trades. Furthermore researchers, like Chan and Lakonishok (1993) and Gomes and Waelbroeck (2013), attribute price changes caused by institutional traders to the informational content of their trades. The evidence does not show whether all trades by institutional traders are informed and whether the informational content is the cause of the price impact.

2.4 The price impact function

The correlation between a trade and the price of the next trade is called the price impact of a trade. Kyle (1985) was the first researcher to show information evolves gradually into prices, resulting in a linear price impact function. In their paper Huberman and Stanzl (2004) assume liquidity is constant and quasi-arbitrage does not exist in a viable market. Quasi-arbitrage is a strategy with endless expected profits. In line with Kyle (1985) they show the price impact function has to be linear to trade size to ascertain viable markets.

A large strand in recent literature concludes the price impact function is concave. Toth, Lempérière, Deremble, Lataillade, Kockelkoren and Bouchaud (2011), for example, propose a model to show large meta-orders should be cut into pieces to be handled by the market liquidity. A meta-order is a bundle of orders resulting from a single trading decision. From the model they show the price impact function is concave, depending on the square root impact law. They attribute the concavity to a memory effect in the order book resulting from diffusive prices. These results imply traders need to account for these liquidity effects when they decide on their trading strategy.

Mastromatteo, Toth and Bouchaud (2013) review the paper by Toth et al. (2011). They also find that, if the last trades of a meta-order only have about 40% of the price impact of the first trades, a memory effect must exist in the market that extends over the normal time to execute a meta-order. They further conclude executing meta-orders aggressively using market orders or passively using limit orders does not change the shape of the price impact

(11)

10 curve. This is very counterintuitive for single orders, but is empirically proven for meta-orders by Mastromatteo et al. (2013).

The main goal of the paper by Doyne Farmer, Gerig, Lillo and Waelbroeck (2013) is to understand how splitting an order affects the impact. They model multiple competitive traders who receive the same information at the same time and must determine their trade size of the orders to submit. They conclude the final price after a meta-order is executed equals its average transaction price. Furthermore they find the market impact should increase as the square root of the trade size of the meta-order.

Another paper of Lillo, Doyne Farmer and Mantegna (2003) investigates the relation between market capitalization and price impact. They conclude, for the same trade size, stocks with higher market capitalization have a smaller price impact than smaller market capitalization stocks. They plotted price change against a normalized trade size for stock sorted by market capitalization on a double logarithmic scale. The results show an upward curve with a steeper slope for small transactions than for large transactions, i.e. a concave function with a steeper slope.

A different explanation for the concavity of the price impact function is offered by Weber and Rosenow (2005). They state a concave price impact function would imply that one should execute large trades instead of cutting them up into smaller fragments. They tested this and found a virtual price impact which is four times stronger than the actual price impact. It is attributed to an “anticorrelation” between returns and order flows. They interpret the anticorrelation as an indication that the numbers of limit sell orders increase if prices are rising and decrease if prices are falling. Weber and Rosenow (2005) thus conclude the dynamics of the order flow causes the actual price impact to be lower than projected.

Initially, researchers thought information gradually incorporated in prices. The recent strands of literature show the price impact function is concave. The evidence of the papers shows cutting a large meta-order in pieces is necessary when accounting for market liquidity, but the use of limit orders instead of market orders does not cause the shape of the price impact function to differ. Furthermore, evidence shows the final price after a meta-order is executed is its average transaction price and higher market capitalization stock experience smaller price impact than smaller market capitalization stock. Weber and

(12)

11 Rosenow (2005) claim one should execute large trades instead of cutting them into smaller pieces, because order flow dynamics induce the actual price impact to be lower than the expected price impact. This conclusion causes the evidence to be inconclusive on how to execute a meta-order.

2.5 Conclusion of the review

To conclude the review, markets appear not to be efficient as Fama (1970) assumed. Hasbrouck (1991) concludes asymmetric information influences the price forming. The early stage of research of Kyle (1985) and Glosten and Milgrom (1985) conclude informed trades are the only traders to push prices and the price impact function is linear. Barclay and Warner (1993), Chakravarty (2001), Gomes and Waelbroeck (2013) and Chan and Lakonishok (1993) find institutional investors are informed investors and they cut their trades in medium sized proportions, which have a significant price impact. Chakravarty et al. (2004) and Anand and Chakravarty (2007) find similar results for informed traders in the option markets. These results imply institutional investors only execute informed trades and that these trades impact prices. However Alex Vermogensbank, an institutional investor, also performs trades for liquidity purposes, i.e. trade based on uninformed reasons. Toth et al. (2011), Mastromatteo et al. (2013) and Lillo et al. (2003) find contradicting results to the assumption of Kyle (1985) and findings Huberman and Stranzl (2004) with respect to the linear shape of the price impact function. They all find a concave price impact function, implying a larger price effect for small trades. Weber and Rosenow (2005) state a correction mechanism of limit orders in the order book causes the actual price impact of market orders to decrease, causing the concave function. The most important result for this research of the literature review is that informed trades influence prices, but it does depend on the definition of informed trading.

3. Data and research method

This part will elaborate on the method to answer the questions, therefore the first part explains the role of Alex Vermogensbank and states three hypotheses, subsequently part three describes the data and summarizes summary statistics are shown and the last part describes the employed methods.

(13)

12 3.1 Hypotheses

Alex Vermogensbank is an institutional investor, which manages assets for their customers. A customer deposits a minimum of €25.000 in an account and grants Alex Vermogensbank authority over the composition of their portfolio. The bank uses a computerized model to determine which shares to buy and which ones to sell. The resulting trades will be used as informed trades, from now on earmarked as algorithmic trades. Customers may withdraw money from the account at any time, at no cost. Therefore if they extract part of the portfolio, Alex Vermogensbank will sell an even proportion of each stock of the portfolio for liquidation purposes. The resulting trades are not based on a signal of the computerized model and will be earmarked redemptions. This paper assumes these trades to be uninformed. The service of Alex Vermogensbank is called Alex Vermogensbeheer.

To recall from the first part, the aim of the paper is to determine which trade has what impact on price. First a method by Chan and Lakonishok (1993) will be used to determine the different kinds of price impact in the sample of which the results will be shown in section 4.1 and be evaluated using the following two hypotheses.

Hypothesis 1: Sales (purchases) exhibit a negative (positive) temporary price impact.

Hypothesis 2: Algorithmic sales (purchases) exhibit a negative (positive) permanent price impact.

Hypothesis 3: Redemption sales do not exhibit permanent price impact.

The first hypothesis results from the short run liquidity hypothesis, which accounts for the temporary price impact. Holthausen, Leftwich and Mayers (1987)argue for a seller initiated trade, if liquidity is low and it is hard to find potential buyers in the market, a seller will receive a lower price. The price paid is below the previous price. The same reasoning goes for a buyer initiated trade; if liquidity is low the buyer will pay a higher price.

The second hypothesis and third hypothesis follows from the imperfect substitution hypothesis, also defined by Holhausen et al. (1987). This hypothesis accounts for the permanent price impact in case of imperfect and perfect substitution. A stock with insufficient substitutes experiences a permanent price impact, so sellers initiating a trade have to offer discounts to buyers. The same argument leads to buyers paying a premium in

(14)

13 buyer initiated trades. In this sample of Eurostoxx-50 close substitutes do exist. Therefore in this sample the permanent price impact captures the informational content of the price impact if close substitutes do exist. A sale implies the stock being overvalued and a purchase implies the stock to be undervalued according to the trader.

Subsequently panel regressions will be used to determine the informational content of the price impacts in the sample. Therefore section 4.2, section 4.3 and section 4.4 will test the following hypotheses.

Hypothesis 4: Algorithmic sell trades of Alex Vermogensbeheer negatively impact prices. Hypothesis 5: Redemption sell trades of Alex Vermogensbank do not impact prices. Hypothesis 6: Algorithmic buy trades of Alex Vermogensbank positively impact prices. 3.2 Sample and summary statistics

The research uses tick data and interval data from the Eurostoxx-50, which is the leading Blue-chip index for the Eurozone (Euro Stoxx 50, n.d.). In collaboration with my employer, Alex Vermogensbank, I retrieve the tick data from Bloomberg. From a private database of Alex Vermogensbank I retrieve the interval data. The sample period will consist of the first of January 2014 up to March 31st 2013. The tick-data is only available from 90 days backwards; therefore I chose this time frame. The trades of Alex Vermogensbeheer will be earmarked using their database. The trades performed after a buy or sell signals are defined as algorithmic trades. The trades performed because a customer wants to liquidate part of the portfolio are defined as redemption trades. Therefore it is inevitable the sample of Alex Vermogensbeheer its trades does not contain uninformed purchases of stock.

A tick test divides the sample in buys and sells, which is part of a method constructed by Lee and Ready (1991). The tick test as described by Lee and Ready (1991) classifies a trade using the preceding price changes. A trade classifies as an uptick if the price is higher than the preceding price and it classifies as a downtick if the price is lower than the preceding price. The trade is a zero tick if the price did not change, in such a scenario the last price change is evaluated and the trade is a zero-uptick (zero-downtick) if the last change was an increase (a decrease) in price. The orders after an uptick or a zero-uptick are recorded as buys. The trade classifies as a sell order if it occurs after a downtick

(15)

14 or a zero-downtick. Lee and Ready (1991) use the combination of quotes and the tick test to increase the precision of the model. Relative to the tick test alone, this method increases precision in determining the characteristics of an order. For this thesis it was not possible to download the bid-ask spread, because the Bloomberg Terminal could not handle downloading such amounts of data. Therefore this paper only uses the tick test as described above.

Graph 1.A. shows the algorithmic and redemption sales in the sample. The algorithmic sales show some spikes of trades in the sample, but still each batch of trades is executed across several days. The redemption sales are fairly randomly distributed over the sample period. Graph 1.B. shows the algorithmic purchases, these are also fairly distributed over the sample period. Just like Huberman and Stranzl (2004) this research assumes the price impact will incorporate in prices linearly. Therefore linear panel regressions seem reasonable.

Table 1 shows the characteristics of the trades in the sample. The tick data contains 3.164 observations in which trade size is zero. These observations are removed. The sample consists of 14.573.456 trades, of which 12.409 are trades of Alex Vermogensbeheer. The total trade size of Alex Vermogensbeheer is 7.260.320 stock, this amounts to €216.866.208. The table shows the average trade size of algorithmic trades is larger than the average trade size of uninformed trades. The median of informed trades is also higher than the uninformed trades. One possible reason for this fact is the uninformed trades are smaller, because they result from liquidation requests and the positions in the portfolios, which are liquidated, are small on average. However, like Toth et al. (2011) and Chakravarty (2001), early research shows informed traders should cut their trades to camouflage them between uninformed trades. These summary statistics seem to suggest the informed trades on average do not resemble the uninformed trades in terms of trade size. Section 4.1 elaborates more on this subject.

Furthermore the maximum trade size of 21.200.000 does not seem to be an outlier. The stock ENEL is valued at €3, so the trade costs are about €60.000.000 which is high, but seems plausible, because trades of similar size were performed in the sample period in ENEL. It was, however, not a trade of Alex Vermogensbeheer.

(16)

15 Tr ad e s iz e P ri ce Tr ad e s iz e P ri ce Tr ad e s iz e P ri ce Tr ad e s iz e P ri ce M e an 812 44 .1 4 € -643 48 .3 6 € 643 48 .3 6 € SD 2, 17 9 22 .4 6 € -16 ,8 13 33 .0 4 € 16 ,8 05 33 .0 3 € O b se rv ati o n s 6, 57 0 6, 57 0 6, 68 7, 55 4 6, 68 7, 55 4 6, 69 4, 12 4 6, 69 4, 12 4 M ax imu m 65 ,0 04 95 .4 2 € -9, 68 9, 10 2 19 7. 95 € 9, 68 9, 10 2 19 7. 95 € M e d ia n 274 55 .4 5 € -191 51 .3 8 € 191 51 .3 9 € M in imu m 1 3. 86 € -1 3. 12 € 1 3. 12 € Tr ad e s iz e P ri ce Tr ad e s iz e P ri ce Tr ad e s iz e P ri ce Tr ad e s iz e P ri ce M e an 418 57 .6 2 € 62 57 .0 0 € 644 48 .2 6 € 643 48 .2 7 € SD 1008 16 .1 4 € 473 15 .7 7 € 17 ,0 06 33 .2 7 € 17 ,0 00 33 .2 7 € O b se rv ati o n s 4385 4, 38 5 1, 45 4 1, 45 4 7, 87 3, 49 3 7, 87 3, 49 3 7, 87 9, 33 2 7, 87 9, 33 2 M ax imu m 28994 67 .7 4 € 17 ,8 50 15 7. 50 € 21 ,2 00 ,0 00 19 7. 80 € 21 ,2 00 ,0 00 19 7. 80 € M e d ia n 197 63 .2 3 € 32 55 .7 7 € 192 51 .0 7 € 192 51 .1 2 € M in imu m 1 9. 83 € 1 9. 84 € 1 3. 12 € 1 3. 12 € P a n el B : S a le s Ta b le 1 . S u m m ar y st at ist ic s Pa n e l A su m m a ri ze s th e p u rc h a se s in th e s a m p le . P a n e l B s u m m a ri ze s th e s a le s in th e s a m p le . R e d e mp ti o n s A lg o ri th mi c A le x A le x M ar ke t M ar ke t To ta l A lg o ri th mi c R e d e mp ti o n s To ta l P a n el A : P u rc h a se s

(17)

16 0 50 1 0 0 1 5 0 2 0 0 N u mb e r o f tra d e s 0 5 0 0 1 0 0 0 1 5 0 0 N u mb e r o f tra d e s

01jan2014 01feb2014 01mar2014 01apr2014 Date

Algorithmic Redemptions

These graphs show the number of algorithmic and redemption sales each day.

Graph 1.A. Algorithmic and redemption sales

0 2 0 0 4 0 0 6 0 0 8 0 0 N u mb e r o f tra d e s

01jan2014 01feb2014 01mar2014 01apr2014 Date

This graph shows the number of algorithmic purchases each day.

(18)

17 3.3 Research method

This section explains the employed methods. The first section, 3.3.1, will elaborate on the method of Chan and Lakinoshok and section 3.3.2 on a set of panel regressions for individual trades. Section 3.3.3 focuses on a panel regression to evaluate the price impact of blocks of trades.

3.3.1 Total, temporary and permanent price impact

The method of Chan and Lakonishok (1993) calculates the percentage return from starting price of the interval to trade price, trade to closing price of the interval and percentage return from opening to closing. They respectively correspond to the total, temporary and permanent price impact of the trade, as defined in Holthausen, Leftwich and Mayers (1987). To determine whether Alex Vermogensbeheer its trades differ from market trades, the price of the trades is compared to the impact of the Volume Weighted Average Price (VWAP) in line with the method of Chan and Lakonishok (1993). The definition for sales is the natural logarithm of the trade price minus the VWAP. The definition for purchases is the natural logarithm of the VWAP minus the trade price. A result of these definitions is that a trader performs better than the market if he buys below VWAP and sells above the VWAP.

3.3.2 Individual trades

The second, the third and the fourth method will be used to determine the informational content of the price impact of the trades of Alex Vermogensbank using panel regressions. When using a panel regression one is able to control for time fixed effects and entity fixed effects. Therefore, this paper uses panel regressions instead of ordinary least squares regressions. The first set of panel regressions will be performed to determine the informational content of the price impact on a short interval level. The second set of panel regressions will be performed to determine the informational content of the price impact on a long interval level. The third set of panel regressions will be used to determine price impact on the next trade. The panel variable is the companies in the Eurostoxx-50 in which Alex Vermogensbeheer traded during the sample period. Each trade has its own time period.

3.3.2.1 Panel regressions on interval level

The first two sets of panel regressions will be performed on a short and long interval level using the price impact definitions of the method by Chan and Lakonishok (1993). The

(19)

18 beginning of the short interval will be the trade before the current trade and the end of the interval will be the tenth trade after the current trade. The start of the long interval will be the trade preceding the current trade and the end of the interval will be the hundredth trade after the current trade. In order to interpret the constant in the model as the average price impact of a trade, the variable trade size is defined as the difference between the trade size of the current trade and average trade size. Using this definition enables me to interpret the coefficient on the dummy variables as the difference in price impact between the algorithmic trades and market trades and the trades for liquidity motives and the market trades. The price impact on a short interval level will be tested for sales using equation (1), (2) and (3) and purchases using equation (4), (5) and (6). The equations (7), (8) and (9) test the informational content for sales on a long interval level and (10), (11) and (12) on for purchases. According to Chan and Lakonishok (1993) these respectively correspond to the total, temporary and permanent price impact.

Evidence will be in favor of the third hypothesis, from section 3.1, if the coefficient on the dummy variable algorithmic is negative and significant. The fourth hypothesis will be rejected if the dummy variable redemptions is significantly different from zero. The evidence will be in favor of the fifth hypothesis if the coefficient on the variable algorithmic is positive and significant.

For sales on a short interval: Total price impact:

(1)

Temporary price impact:

(2)

Permanent price impact:

(20)

19 For purchases on a short interval:

Total price impact:

(4)

(5)

(6)

For sales on a long interval: Total price impact:

(7)

(8)

(9)

For purchases on a long interval: Total price impact:

(21)

20 Temporary price impact:

(11)

(12)

Using the following variables;

- Pi,t is the price of the current trade, Pi,t+10 is the price of the tenth trade after the

current trade, Pi,t+100 is the price of the hundredth trade after the current trade and

Pi,t-1 is the price of the trade preceding the current trade.

- Algorithmic is one if the trade is executed following a signal of the model and zero otherwise.

- Redemptions is one if the selling trade is executed following a liquidation order and zero otherwise. Only for sell orders uninformed trades are executed.

- The entity fixed effect is a control variable for an effect fixed over time for each stock, but varying between different stock.

- The time fixed effects is a control variable for an effect fixed over all stock, but varying over time.

- The variable trade size is a control variable to control for the trade size of the specific trade minus the average trade size.

3.3.2.2 Panel regression on trade level

The second set of panel regressions, equation (13) for sales and equation (14) for purchases will be used to determine the informational price impact on the next trade. Alex Vermogensbeheer performs sell orders after informed and uninformed signals. Therefore the equation for sales contains dummies for both signals. The situation when both dummies are zero amounts to the market trades. Alex Vermogensbeheer does not perform uninformed buy orders. The definition of the control variable trade size is the same as in the regressions used in the previous regression methods. Therefore equation (14) tests whether

(22)

21 informed buying orders differ from the market buying orders using a dummy for trades performed because of the algorithmic model.

The results will be analyzed using the following conclusions. The evidence will be in favor of the third hypothesis, if the coefficient on the dummy variable is negative and significant. In such a case the sales of Alex Vermogensbeheer appear to exhibit informational negative price impact in the sample. The fourth hypothesis states uninformed trades do not impact prices. This hypothesis will be rejected if the coefficient on redemptions is significantly different from zero. The fifth hypothesis states the buy orders will positively impact price change if the coefficient on the dummy algorithmic is positive and significant. In that case the evidence is in favor of the fifth hypothesis.

For sales: (13) For purchases: (14)

- Pi,t is the price of the current trade and Pi,t+1 is the price of the next trade.

- Algorithmic is a dummy variable which is one if the trade is initiated by a signal from the model and zero otherwise.

- Redemptions is a dummy variable which is one if the trade is initiated by a liquidation order of the customer and zero otherwise.

(23)

22 - The variable trade size is a control variable which controls for the impact of the trade

size of each trade minus the average trade size. 3.3.3 Blocks of trades

To determine the effect of blocks of trades of Alex Vermogensbeheer, the fifteen minutes interval data and tick trade data is used. Two sets of panel regressions aim to determine the price impact on blocks of trades of Alex Vermogensbeheer. During the sample period the trades by the market and by Alex Vermogensbeheer are recorded. On each day the same fifteen minute time intervals are chosen and the opening price, closing price and trading volume of those intervals is recorded. Furthermore, the trade volume of the sales and purchases following the signals of the model and the sales following liquidation orders of Alex Vermogensbeheer are recorded.

3.3.3.1 Blocks of trades in 15 minute intervals

The first set of panel regressions determines the price impact of the three types of trades in the 15 minutes intervals the panel regression equations (15) and (16) are used. Equation (15) is used for sales and equation (16) is used for purchases.

- Pend is the price at the end of a 15 minute interval and Pstart is the price at the start of

the 15 minute interval. The difference between the natural logarithms is the percentage change in price between the start and the end of the interval.

- The variable algorithmic as the percentage of informed Alex’ trades per stock in the interval.

(24)

23 - The variable redemptions as the percentage of uninformed Alex’ trades per stock in

the interval.

Whether the trades are informed and uninformed in the sample and exhibit price impact on the prices in the market, depends on the coefficients on the variables algorithmic and redemptions. Evidence will exist in favor of the third hypothesis, if in the panel regression equation (15) the variable on the coefficient algorithmic is negative and significant. It could imply the algorithmic sales exhibit price pressure; because the block trades are informed independent of fixed effects. The fourth hypothesis will be rejected if in the same panel regression equation the coefficient on the variable redemptions is significantly different from zero. In case the coefficient would be non-zero and significant, the uninformed sell orders could influence prices on the basis of being informed. To infer evidence exist in favor of the hypothesis informed purchases exhibit an positive price impact independent of their trade size, the coefficient on the variable algorithmic in panel regression equation (16) should be positive and significant.

3.3.3.2 Daily price impact of blocks of trades

The second set of panel regressions on the price change between the daily open and closing price uses the daily block of trades of Alex Vermogensbeheer to determine the price impact. The panel regression for sales in equation (17) uses the daily percentage of sales resulting from the algorithms and the daily percentage of sales resulting from the liquidation orders to determine the price impact of the two types of trades. The panel regression for purchases in equation (18) aims to determine the price impact of the purchases resulting from the model through the daily percentage of informed purchases. The panel variable is the stock of the Eurostoxx-50 and the time variable are the days between the first of Januari 2014 and the 31st of March 2014.

The evidence favors the third hypothesis if the coefficient on the variable algorithmic in panel regression equation (17) is negative and significant. This would indicate with each percentage increase in daily percentage informed sales the price would drop a certain

(25)

24 amount over the day. The evidence rejects the fourth hypothesis, if the coefficient on the variable redemptions in panel regression equation (18) is significantly different from zero. This would indicate prices would change due to a one percentage increase in uninformed sales. The evidence supports the fifth hypothesis, if the coefficient on the variable algorithmic in panel regression equation (18) is positive and significant. It would provide support for the daily percentage of informed purchases to impact prices positively.

- Pclose is the price of the last sale or purchase of the day and Popen is the price first sale

or purchase of the day. The difference between the natural logarithms is the percentage change in price over the day.

- The variable algorithmic as the percentage of informed Alex’ trades per stock on each day.

- The variable redemptions as the percentage of uninformed Alex’ trades per stock each day

4. Results

The results of each method will be presented in individual sections. In section 4.1 the results of the method of Chan and Lakonishok (1993) are presented. The section 4.2 and 4.3 present the results of the panel regressions based on the short and long interval for

(26)

25 individual trades. The section 4.4 elaborates on the results of the panel regression for individual trades on trade level. Section 4.5 summarizes the results of panel regressions on blocks of trades based on 15 minutes intervals and daily open en close prices.

4.1 Total, temporary and permanent price impact

Table 2.A. and table 2.B. summarize the price impact of the sales and purchases respectively. The price impact for both sales and purchases the trades of Alex Vermogensbeheer are shown first and subsequently the trades of the market.

Table 2.A. Sales: total, temporary and permanent price impact.

Panel A and B summarize the price impact for informed and uninformed sales and panel C summarizes the price impact of the market sales. The price impacts are average price changes in percentage. The significance levels show whether the price impact

is significantly different from zero. *** Indicates 1% significance, ** indicates 5% significance, * indicates 10% significance

Panel A: Informed sales

Price change % TOTAL TEMPORARY PERMANENT VWAP

mean -0,004 *** -0,008 *** 0,005 *** -0,080 *** SD 0,026 0,071 0,072 0,513 Observations 5839 5839 5839 4385 max 0,245 0,389 2,153 2,095 median 0,000 0,000 0,000 -0,061 min -0,778 -2,153 -0,778 -2,406

Panel B: Uninformed sales

mean -0,017 *** -0,019 *** 0,002 0,016 SD 0,025 0,048 0,051 0,549 Observations 1454 1454 1454 1454 max 0,111 0,192 0,228 3,125 median -0,016 -0,018 0,000 0,033 min -0,193 -0,228 -0,279 -3,377

Panel C: Market sales

mean -0,011 *** -0,005 *** -0,006 *** -0,018 SD 0,024 0,050 0,052 0,607 Observations 7873492 7873416 7873415 7873493 max 0,000 6,336 9,106 3,650 median 0,000 0,000 0,000 -0,017 min -9,235 -9,106 -8,551 -6,215

The temporary price change for an algorithmic sale in panel A of table 2.A. is -0.008% and significant. I expected to find a negative temporary average price impact for algorithmic trades. The evidence in the sample does provide evidence for the first hypothesis, i.e. the algorithmic sales exhibit a negative temporary price impact. The temporary price impact of the redemption sales of Alex Vermogensbeheer in panel B of table 2.A. is -0.019% on

(27)

26 average. This suggests the redemption sales impact prices negatively. Evidence exist is in favor of the first hypothesis that sales impact prices negatively. Panel A of table 2.B. shows the temporary average price impact of Alex its purchases is 0.004% and it is significant. The first hypothesis states algorithmic purchases exhibit temporary price impact. Therefore evidence exists in favor of the first hypothesis. The evidence suggests the short run liquidity hypothesis of Holthausen et al. (1987) holds for sales and purchases, no matter the reason of the sale or purchase.

Table 2.B. Purchases: total, temporary and permanent price impact.

Panel A summarizes the price impact for informed purchases and panel B summarizes the price impact of the market purchases. The price impacts are average price changes in percentage. The significance levels show whether the price impact is significantly

different from zero. *** Indicates 1% significance, ** indicates 5% significance, * indicates 10% significance

Panel A: Informed purchases

mean 0,000 0,004 *** -0,004 *** 0,174 *** SD 0,037 0,053 0,060 0,840 Observations 6570 6570 6570 6570 max 1,498 0,624 1,498 2,973 median 0,000 0,000 0,000 0,143 min -0,353 -0,393 -0,624 -4,016

Panel B: Market purchases

mean 0,013 *** 0,006 *** 0,007 *** 0,006 *** SD 0,025 0,050 0,052 0,613 Observations 6687554 6687491 6687491 6687554 max 8,807 8,978 3,423 9,267 median 0,000 0,000 0,000 0,007 min 0,000 -3,390 -6,558 -6,203

The permanent price impact of algorithmic sales is 0.005% and significant and for redemption sales it is insignificant. I expected algorithmic sales to exhibit negative and significant price impact, because if a trader is informed about the value of a stock the sale negatively change prices and this change will be permanent. I did not expect the redemption sales to impact prices, because the informational content of redemption sales should be zero and price change should not be permanent. Therefore, this indicates evidence in favor of the fourth hypothesis. The permanent price impact for algorithmic purchases is negative and significant. I expected purchases to exhibit positive permanent price impact. These results are not in line with the imperfect substitution hypothesis and the private information hypothesis for algorithmic sales and purchases. The results imply sales result in a positive price change and purchases in a negative price change.

(28)

27 I suspect Alex Vermogensbeheer did not pick the right stock during the sample period. If they did not pick the right stock, the informational content of the trades could be zero or negative. Therefore, further research shows the return of Alex Vermogensbeheer was lower than the benchmark as shown in table 3. The benchmark uses a comparable strategy using the same universe of stocks; this could indicate Alex Vermogensbeheer picked the wrong stock.

Table 3.Results Alex Vermogensbeheer 2014

This table shows the results of Alex Vermogensbeheer from January 2014 up until May 2014. The column Stoxx TR shows the results of the benchmark of Alex Vermogensbeheer,

Month Defensive Cautious Offensive Speculative Very Speculative STOXX TR

January -1,10% -1,20% -2,60% -2,80% 1,50% -1,70% February 2,40% 4,10% 5,00% 4,90% 6,50% 5,00% March -0,60% -1,10% -1,20% -1,50% -1,30% 0,40% April -1,40% -3,40% -2,10% -1,40% -3,50% 0,80% May 0,40% 1,90% 2,00% 2,10% 0,80% 2,50% Total result -0,40% 0,10% 0,90% 1,10% 3,70% 7,10%

The summary statistics of table 1 show the average trade size for algorithmic trades is larger, than the average trade size for redemption trades. This observation may indicate it is easier for market participants to identify algorithmic trades compared to redemption trades. The price difference with respect to the volume weighted average price for algorithmic sales is -0.080% and significant. The price difference with respect to the volume weighted average price of the algorithmic purchases is 0.174% and significant. This indicates the broker of Alex Vermogensbeheer performed the algorithmic sales at a lower price than the volume weighted average price and the algorithmic purchases at a lower price than the volume weighted average price. This indicates the market is able to identify the algorithmic sales, but it is not able to identify the algorithmic purchases. Differentiating between the informational content of purchases and sales is outside the scope of this research. It could also indicate the broker of Alex Vermogensbeheer exerts more effort hiding the purchases among uninformed trades than the sales.

To conclude evidence exists in favor of the first and third hypothesis, but it does not exist in favor of the second hypothesis. It indicates the short run liquidity hypothesis of Holthausen et al. (1987) holds, i.e. Alex Vermogensbeheer faces liquidity constraints. The evidence does not indicate the algorithmic trades of Alex Vermogensbeheer contain

(29)

28 informational content. This could be a result of picking the wrong stock. The results do provide evidence the redemption trades do not contain informational content.

4.2 Individual trades on short interval level

Table 4 summarizes the results of the panel regression of the price impact of trades in the short interval, i.e. the interval ending with the tenth trade after the current trade. Panel A summarizes the regression results for sales and panel B the results for purchases. These will be evaluated using the six hypotheses stated in section 3.1.

The coefficient on the dummy variable algorithmic in the panel regression on the temporary price impact is insignificant for sales and for purchases. The coefficient on redemptions in the panel regressions on temporary price impact is negative and significant. Based on this evidence the first hypothesis is rejected for algorithmic trades and evidence is in favor for the first hypothesis for the redemption trades. So for the redemption trades the short run liquidity hypothesis holds, but for the algorithmic trades it does not hold.

In the panel regression on permanent price impact, the coefficient on algorithmic sales is 0.0112% and significant. It implies for each sale being performed, because of a signal of the model, the price of the tenth trade after it is 0.0112% higher. This does not provide evidence in support of the third hypothesis. It states the price would be lower, because the trade was initiated by a signal of the model. The panel regression results on permanent price impact for purchases from table 4 panel B shows a similar result. The coefficient on algorithmic purchases is -0.0092%, which implies the purchases initiated by the model push the price of the tenth consecutive trade to be 0.0092% lower. This does not support the fifth hypothesis. For purchases I expected the model to push the price upward, because of the informed character of the trade. The dummy on redemption sales is 0.0111%. This does not provide evidence in support of the fourth hypothesis. It states redemption sales do not impact prices.

The coefficients on the algorithmic dummies in the regressions on total price impact also do not provide evidence in support of the third and the fifth hypothesis. As stated in section 4.1 and shown in table 3, I suspect Alex Vermogensbeheer did not pick the right stock during the sample period. Alex Vermogensbeheer underperformed with respect to the benchmark. A second reason why the results do not provide evidence for the hypotheses could be Alex Vermogensbeheer tries to execute sales when the prices are on a short run high and purchases on a short run low. Due to the length of the interval, i.e. the end of the

(30)

29 interval is the tenth trade; the results could be biased by this. The regression on the volume weighted average however does not show any evidence in favor of this reasoning. The results from the regression on the long interval could prove otherwise.

To conclude, the results of the panel regressions on a short interval level do not show evidence Alex Vermogensbeheer its algorithmic sales or purchases exhibit informational price impact. One of the reasons could be the stock picking ability of the model of Alex Vermogensbeheer did not perform better than the market. Another could be the brokers of Alex outperform the market traders, but the evidence does not confirm this last theory.

Table 4. The percentage price change of individual trades on short interval level.

Panel A shows the price impact of individual sales. Panel B shows the price impact of individual purchases. The standard deviations are shown underneath the coefficients in parentheses. *** Indicates 1% significance, ** indicates 5% significance, *

indicates 10% significance

Panel A: Sales

Price Change % Total Temporary Permanent VWAP

Algorithmic 0,0076 *** -0,0037 0,0112 *** -0,0665 (0,0016) (0,0025) (0,0022) (0,1308) Redemptions -0,0029 * -0,0140 *** 0,0111 *** 0,0314 (0,0016) (0,0019) (0,0019) (0,0199) Trade size -0,0225 *** -0,0108 -0,0115 0,0058 (0,0053) (0,0072) (0,0097) (0,0138)

Constant -2,65E-06 *** -5,78E-06 *** 2,99E-06 ** -0,0183 ***

(9,37E-07) (1,44E-06) (1,30E-06) (0,00007)

Fixed effects Yes Yes Yes Yes

Clustered Yes Yes Yes Yes

Panel B: Purchases

Algorithmic -0,0087 *** 0,0005 -0,0092 *** 0,1619 (0,0008) (0,0015) (0,0017) (0,1439) Trade size 0,0365 *** 0,0149 0,0211 *** 0,0815 *** (0,0060) (0,0081) (0,0043) (0,0154) Constant 0,00001 *** -0,00002 *** 0,00003 *** 0,0062 ***

(7,92E-07) (1,51E-06) (1,70E-06) (0,0001)

Fixed Effects Yes Yes Yes Yes

(31)

30 4.3 Individual trades on long interval level

Table 5 summarizes the results of the panel regressions on the long interval level, i.e. the interval ending with the hundredth trade after the current trade. This section uses the five hypotheses of section 3.1 to evaluate the results of the regressions of equation (7) to (12). The panel regression on the temporary price impact in panel A for sales shows the coefficient on the variable algorithmic is -0.0743% and significant. The algorithmic sales of Alex Vermogensbeheer temporarily impact prices downward with 0.0743% with respect to the average sale of the market. The evidence is in favor of the first hypothesis. The panel regression on the temporary price impact in panel B for purchases shows the coefficient is 0.0292% and significant. According to this result the algorithmic purchases temporarily push prices upward with 0.0292% with respect to an average market purchase. The evidence for purchases is also in favor of the first hypothesis. It indicates that for algorithmic sales the short run liquidity effect is more pronounced. According to the short run liquidity

hypothesis of Holthausen et al. (1987) this result suggests that for sales liquidity is tighter for Alex Vermogensbeheer than for purchases. The coefficient on the variable redemptions in the regression on temporary price impact is -0.0264% and significant. This evidence is in favor of the third hypothesis. The temporary price impact for redemption sales in this interval is smaller than the temporary price impact of algorithmic sales. This could indicate the market identifies the algorithmic sales as informed and the redemption sales as

uninformed.

However, the regression on the permanent price impact for sales shows a significant coefficient of 0.0819% on the variable algorithmic and a 0.0236% on the variable

redemptions. This indicates the algorithmic sales push price up instead of down suggesting these sales are not informed. The redemption trades do exhibit price impact, therefore based on the evidence these trades seem to push prices. These results are not in line with the second and third hypotheses; therefore these will be rejected based on this evidence. The coefficient on the permanent price impact on purchases is -0.0379% and significant. This is not in line with informed purchases pushing prices upward. Therefore, the fourth hypothesis will be rejected.

The total price impact on sales and purchases does not provide evidence in favor of the second hypothesis, but it does provide evidence in favor of the third hypothesis. The

(32)

31 coefficient on the variable redemptions is insignificant on a 5% significance level. This

indicates they do not push prices.

Table 5. The percentage price change of individual trades on a long interval

Panel A shows the price impact of individual sales on a long interval. Panel B shows the price impact of individual purchases on a long interval. *** Indicates 1% significance, ** indicates 5% significance, * indicates 10% significance

Panel A: Sales

Algorithmic 0,0076 *** -0,0743 *** 0,0819 *** -0,0665 0,0016 0,0257 0,0260 0,1308 Redemptions -0,0029 * -0,0264 *** 0,0236 *** 0,0314 0,0016 0,0043 0,0051 0,0199 Trade size -0,0225 *** -0,0065 -0,0159 0,0058 0,0053 0,0141 0,0165 0,0138 Constant -2,65E-06 *** -0,000039 ** 0,000036 ** -0,0183 *** 9,37E-07 0,00001 0,00001 0,00007

Fixed effects Yes Yes Yes Yes

Clustered Yes Yes No Yes

Panel B: Purchases

Algorithmic -0,0087 *** 0,0292 *** -0,0379 *** 0,1619 0,0008 0,0079 0,0076 0,1439 Trade size 0,0365 *** 0,0197 0,0163 0,0815 *** 0,0060 0,0169 0,0148 0,0154 Constant 0,0000102 *** -0,00013 *** 0,00014 *** 0,0062 ***

7,92E-07 7,71E-06 7,44E-06 0,0001

Fixed Effects Yes Yes Yes Yes

Clustered Yes Yes Yes Yes

To summarize, the evidence of the regressions on a long interval do provide evidence in support of the short run liquidity hypothesis for sales and purchases. It shows the algorithmic sales experience the strongest liquidity constraints with respect to market trades. Against the expectations, the results do not show the algorithmic trades of Alex Vermogensbeheer contain an informational content in its price impact. In section 4.2 I suggested the lack of informational price impact on a short run interval could have been explained by the ability of Alex Vermogensbeheer to trade on short run highs and lows. The

(33)

32 results of the regression on a long interval do not support this conclusion. The remaining reasoning is that the individual trades of Alex Vermogensbeheer do not exhibit informational price impact. The evidence for this conclusion lies in the underperformance with respect to the benchmark in table 3.

4.4 Individual trades on trade level

In table 6 the results of the panel regression on the individual trades are shown. Panel A and B show the results of regression equation (13) and (14) from section 3.3.2.2 and will be used to evaluate the fourth, fifth and sixth hypothesis. The panel regressions were performed with and without fixed effects and clustered standard errors.

Panel A shows the effect of a sell order of Alex Vermogensbeheer initiated by a signal of the model or a liquidation order. The panel regression measures the price impact of the independent variables on the price change from price at the time of the trade to the next trade. The coefficient on the variable algorithmic is negative and significant in the first three regressions. The evidence is in favor of the third hypothesis, i.e. it suggests a sell order pushes prices downward. The coefficient indicates that for each stock sold after an informed signal the next price will decrease with 0.0023% with respect to the average market sales. Table 1 shows the average price of a sold stock by Alex is €57.62, so based on these estimates, selling one stock on a signal on average would result in a price drop of €0.0013 with respect to the market sales. At the same time the coefficient on the variable redemption is positive and significant throughout all the panel regressions. This does not indicate the evidence is in favor of the fourth hypothesis. The evidence indicates the uninformed sales of Alex Vermogensbeheer cause the subsequent price to be higher. The evidence suggests Alex Vermogensbeheer its algorithmic sales push prices down and its redemptions sales push prices up.

Panel B shows the regression results of equation (8). In the regressions the coefficient on the dummy variable indicating whether a trade is performed after a signal of the model is positive and significant. This indicates buying a stock after a signal causes the price to rise with 0.0036% with respect to an average market purchase. Based on the result of these regressions the evidence supports the sixth hypothesis. This suggests the purchases of Alex Vermogensbeheer exhibit an informational price impact with respect to the market purchases.

(34)

33 To conclude the results of the regression on trade level do provide evidence to support the fourth and the sixth hypothesis. The results suggest prices of the subsequent trade are 0.0023% lower if the sale is performed after a signal than the price following a market sale. The price following an algorithmic purchase of Alex Vermogensbeheer seems to be 0.0036% higher than the price following a market purchase.

Table 6. The percentage price change of individual trades

Panel A shows the price impact of individual sales and panel B shows the price impact of individual purchases. *** Indicates 1% significance, ** indicates 5% significance, * indicates 10% significance

Panel A: Sales Price Change in % 1 2 3 4 Algorithmic -0,0023 *** -0,0023 *** -0,0023 ** -0,0023 * (0,0004) (0,0012) (0,0004) (0,0012) Redemptions 0,0090 *** 0,0090 *** 0,0090 *** 0,0090 *** (0,0007) (0,0010) (0,0007) (0,0010) Trade size 0,0028 *** 0,0028 *** 0,0028 0,0028 (0,0005) (0,0042) (0,0005) (0,0042)

Constant 6,35E-07 6,35E-07 6,33E-07 6,33E-07

(9,16E-06) (4,88E-06) (9,16E-06) (6,02E-07)

Fixed effects No No Yes Yes

Clustered No Yes No Yes

Panel B: Purchases Price Change in % 1 2 3 4 Algorithmic 0,0036 0,0036 *** 0,0036 *** 0,0036 *** (0,00032) (0,00058) (0,00032) (0,00058) Trade size -0,0090 -0,0090 -0,0090 *** -0,0090 (0,00060) (0,00591) (0,00060) (0,00591)

Constant -1,83E-06 -1,83E-06 -1,83E-06 -1,83E-06 ***

(1,00E-05) (5,54E-06) (1,00E-05) (5,69E-07)

Fixed Effects Yes No Yes Yes

Clustered Yes Yes No Yes

4.5 Blocks of trades

Table 7 summarizes the results of the panel regressions on the block trades of Alex Vermogensbeheer. In panel A the two regressions aim to determine the price impact of blocks of trades of Alex Vermogensbeheer in 15 minute time intervals.

The price impact of informed and uninformed trades : an Alex Vermogensbank case study