Download Decomposing Financial Instrument Price Process: Permanent vs. Transitory Components and more Lecture notes Innovation in PDF only on Docsity!
Implementation Shortfall with Transitory Price
Effects
Terrence Hendershott, Charles M. Jones, and Albert J. Menkveld
March 18, 2013
Regulators and some large investors have recently raised concerns about temporary or transitory volatility in highly automated financial markets.^1 It is far from clear that high-frequency trading, fragmentation, and automation are contributing to transitory volatility, but some institutions complain that their execution costs are increasing. In this chapter, we introduce a methodology for decomposing the price process of a financial instrument into its permanent and transitory components, and we explore the insights from applying this method- ology to execution cost measurement. Among other things, our methodology allows an institutional investor to accurately measure the contributions of tran- sitory price movements to its overall trading costs. The methodology is partic- ularly applicable to an investor that splits a large order into small pieces and executes it gradually over time. The importance of transitory price impact has been well-known in the aca- demic literature since early work on block trading (e.g., Kraus and Stoll (1972)).^2 While it is fairly straightforward to measure the transitory price impact of a block trade, it is a much greater challenge to measure the transitory price im- pact when a large institutional parent order is executed in perhaps hundreds of smaller child order executions. The key innovation of our approach is that we estimate the temporary component at each point in time, and in particular whenever a child order executes. By summing over all child orders, we can thus measure the effect of the temporary component on overall trading costs. To be more precise, we extend the classic Perold (1988) “implementation (^1) See for example the US Securities and Exchange Commission’s 2010 concept release on equity market structure (Release No. 34-61358). 2 See Duffie (2010) for an extensive discussion of temporary price impacts from large infor- mationless demands for liquidity.
shortfall” approach to decompose ex-post transaction costs into various com- ponents, one of which accounts for the trading costs associated with transitory pricing errors. Because trading cost analysis is often performed on an insti- tution’s daily trading, we first illustrate our transaction cost measurement ap- proach at a daily frequency. However, our methods are much more precise when more disaggregated trading data are available. Using detailed information on the intraday child order executions from a larger institutional parent order, we show how the transitory price component evolves with trading on a minute-by- minute basis, and we show how this transitory price component contributes to overall implementation shortfall. In some ways, our work is most closely related to Almgren et al. (2005), who assume a particular functional form for both permanent and transitory price impacts, with limited persistence in the latter. They then apply their model to a large set of institutional orders to characterize permanent and transitory components of transaction costs as a function of various stock and order char- acteristics.^3 In contrast, we allow the data to determine the persistence of the temporary component.
1 Implementation Shortfall
Even for those who are intimately familiar with trading cost analysis, Perold (1988) is worth a re-read. For example, he frames the discussion on p.4:
After selecting which stocks to buy and which to sell, “all” you have to do is implement your decisions. If you had the luxury of transacting on paper, your job would already be done. On paper, transactions occur by mere stroke of the pen. You can transact at all times in unlimited quantities with no price impact and free of all commissions. There are no doubts as to whether and at what price your order will be filled. If you could transact on paper, you would always be invested in your ideal portfolio. There are crucial differences between transacting on paper and trans- acting in real markets. You do not know the prices at which you will be able to execute, when you will be able to execute, or even whether you will ever be able to execute. You do not know whether you will (^3) Engle and Ferstenberg (2007) also estimate implementation shortfall costs on a sample of institutional orders, focusing on the variance of the execution costs as well as their mean.
Denote the prices at which transactions take place by pij ; pij are net of incremental costs such as commissions and transfer taxes. Let the price of security i be pbi at the beginning of the period and pei at the end. While the pij must be transaction prices, the two benchmark prices can be either trade prices or quote midpoints. Assuming there are no net cash flows into or out of the real portfolio, all transactions are financed with proceeds of other transactions. That is, at each time j, ∑^ tij pij is zero when summed over i = 1 to N. Let the value of the paper and real portfolios at the beginning of the period be Vb: Vb =
nipbi. (2)
Let the end-of-period values of the real and paper portfolios be Vp and Vr , respectively: Vp =
nipei and Vr =
ωei pei. (3) The performance of the paper portfolio is Vp −Vb, and the performance of the real portfolio is Vr − Vb. The implementation shortfall is the difference between the two. The performance of the real portfolio can be expanded as ∑ (ωie pei − ωbi pbi ) =
ωei (pei − pbi ) −
pbi (ωei − ωbi )
ωei (pei − pbi ) −
(pij − pbi )tij. (4)
The performance of the paper portfolio can be expanded as ∑ ni(pei − pbi ). (5)
Subtracting the real portfolio performance from paper portfolio performance completes the calculation:
Impl. Shortfall =
(pij − pbi )tij ︸ ︷︷ ︸ Execution Cost
(pei − pbi )(ni − ωei ) ︸ ︷︷ ︸ Opportunity Cost
The term (pij − pbi ) is the per-share cost of transacting at pij instead of at pbi , and this cost is applied to tij traded shares. The weighted sum is the total execution cost relative to the pre-trade benchmark. The term (pei − pbi ) is the paper return on security i over the period. The opportunity cost is the sum of
these returns weighted by the size of the unexecuted orders. While opportunity costs are a real concern for institutional investors, our methodology does not offer much insight into them, and in the rest of the chapter we focus only on the execution cost component.
2 Observed prices, efficient prices, and pricing
errors
The implementation shortfall incorporates the total price impact of a large order. However, to better understand the sources of the shortfall, it may be useful to decompose the price impact into its permanent and transitory components. To do this one must define and measure the efficient price and any deviations from it at each moment in time. We take the standard approach of assuming the efficient price is unpredictable, i.e., it follows a random walk. Absent trading frictions, the efficient price at the daily or intraday frequency can be characterized as a martingale process. Let mj be this latent price:
mj = mj− 1 + wt. (7)
Sometimes the quote midpoint is assumed to represent this latent price. However, quote midpoints are not generally martingales with respect to all available order flow, in which case Hasbrouck (1995, p.1179) proposes to view the random-walk component of a Stock and Watson (1988) decomposition as the “implicit efficient price.” Hasbrouck (2007, Ch.4 and Ch.8) constructs an effi- cient price more generally as the projection of mt onto all available conditioning variables, i.e., the so-called filtered state estimate:
m˜ij = E∗^ [mj |pij , pi,j− 1 ,.. .] , (8)
where E∗^ [.] is the linear projection of mij on a set of lagged prices.^4 A standard approach to implementing such a projection is through ARIMA time series econometrics (Hasbrouck (2007, Ch.4)). The filtered estimate can be enriched by expanding the set of conditioning variables with trade-based variables (e.g., signed order flow), news-based variables (e.g., the Reuters sentiment score of (^4) The observed pij in this section can be either trade prices or quote midpoints. In this chapter we always use midquotes.
blessing or a curse - a blessing if you have just extracted a pre- mium for supplying liquidity, a curse if you have just been bagged by someone who knows more than you do.
The efficient price estimate enables one to further refine the standard imple- mentation shortfall calculation of equation (6) by recognizing the size of these two components. The execution cost component of the implementation shortfall can be rewritten as:
Execution Cost =
(pij − mˆij )tij ︸ ︷︷ ︸ Liquidity Cost
( ˆmij − mˆbi )tij ︸ ︷︷ ︸ Informational Cost
( ˆmbi − pbi )tij ︸ ︷︷ ︸ Timing Cost
The first component captures liquidity cost relative to the efficient price. If one buys at a price above the efficient price, one effectively pays a liquidity premium, and if one buys at a lower price one earns the premium. The liquidity costs incorporate both the bid-ask spread and any transitory price effects. For example, if a sequence of trades causes the current quoted price to differ from the efficient price, this temporary price impact is captured in the liquidity cost component. This differs from the standard approach to measuring temporary price im- pact, which compares the price immediately after execution to a price some time later. In the standard approach, the temporary impact reflects the cor- relation between the direction of the order and subsequent price movements. For example, there is temporary impact if prices fall after the completion of a large buy order. The state-space approach captures this general idea, as it incorporates future price movements to estimate the permanent and temporary price decomposition. However, the main advantage of the state-space approach is that it calculates efficient prices throughout the execution period. The tem- porary component can be measured and incorporated into the liquidity cost component for each of the N executions. In contrast, the standard approach can only measure the temporary price impact at the end of the execution period based on its dissipation thereafter. The second component of the implementation shortfall captures the infor- mational cost, as it measures the covariation between executed signed order flow and the efficient price change. This is sometimes referred to as the permanent
price impact of the trades. If for some reason signed flow does not correlate with efficient price changes, then the informational cost is zero. In most finan- cial markets, however, the order flow is potentially informationally motivated, so this component is positive on average. For example, in a classic market- making model a liquidity supplier cannot distinguish informed from uninformed flow and therefore charges all incoming flow the same price impact (see, e.g., Glosten and Milgrom (1985)). In reality, a small informational cost component could reflect the skill of a trader or algorithm in camouflaging the order and having it perceived as uninformed. This component can also reflect variation in the information environment over time. For example, informational costs may be greater just before scheduled earnings announcements. The third component measures whether the timing of the trade is correlated with the temporary component. If the parent order is a buy, for example, then starting it when the quote midpoint is above the efficient price increases the overall cost of the trade, all else equal. Conversely, starting a buy order when the price is below the efficient price should improve its overall execution. We capture this by assigning a negative timing cost when a trade begins in these favorable conditions.
3 Illustration of Approach
Decomposing the price process into its permanent and transitory components is fundamental to our approach. Hasbrouck (2007, Ch.8) provides a detailed discussion of the challenges in identifying the two components. Here we follow an approach developed for analyzing cyclical macroeconomic time series. This approach puts enough structure on the persistence of the transitory price com- ponent to identify the two components. Morley, Nelson, and Zivot (2003, p.240) show that the most parsimonious allowable specification for the temporary com- ponent is an AR(2):
“... the order condition for identification of the unrestricted UC- ARMA(p,q) model, in the sense of having at least as many moment equations as parameters, is p > 0, p > q + 2, and it is just satisfied with p = 2, q = 0.”
3.1 Implementation shortfall calculations
To illustrate our approach, we use two different examples with trading data observed at different frequencies: one example with daily trading data, and one example of a parent order where we observe the size, time, and price of the individual child order executions during the trading day. In the daily example, we have two months’ worth of trades by the same fund in the same stock, ag- gregated at the daily level, and we estimate the efficient and transitory price components at a daily frequency. This approach is most relevant to investors that make each day’s trading decisions overnight while the market is closed, because in that trading environment implementation shortfall is naturally cal- culated relative to the previous closing price. It is worth noting that the decomposition in (11) requires the efficient price estimate at the time of each transaction, ˆmij. In the daily example, however, we only calculate end-of-day efficient price estimates because we do not know when the trades actually take place during the day. This timing mismatch re- duces the precision of the implementation shortfall decomposition and may also introduce bias. The main issue is the allocation of the shortfall between the first two terms of (11), the liquidity and information costs. These two components can be thought of as corresponding to temporary and permanent price impacts, respectively. If there is positive correlation between the direction of trading and the movement in the efficient price, then using an estimate of the efficient price prior to transaction j will overestimate the liquidity cost and underestimate the information cost. Conversely, using an estimate of the efficient price after trans- action j will underestimate the liquidity cost and overestimate the information cost. If only coarse execution data are available and temporary components are sufficiently persistent, however, the decomposition may still prove useful. For the intraday example, we obtain an efficient price estimate for each minute of the trading day. We use these efficient price estimates to evaluate the execution of a single parent order that is gradually executed over the course of about 30 minutes. The intraday horizon allows for an evaluation of the high- frequency price dynamics during order execution. To calculate our implementation shortfall decomposition we use equation (11) with the prices at time j modified as follows:
- the subscript i is dropped as there is only one security;
- pj is the average price at which the institution’s trades execute at time j;
- pb^ is the quote midpoint prior to beginning execution;
- mˆj is the estimate of the efficient price at time j;
- mˆb^ is the estimate of the efficient price prior to beginning execution.
Using these prices the per-share execution costs can be represented as: ( ∑ ( ︸p j − mˆ︷︷j )sign( tj (^) ︸) Liquidity Cost j
|tj | +
( ˆ ︸m j − mˆ︷︷b)sign( tj )︸ Informational Cost j
|tj |+
∑ ( ˆ ︸m b^ − p︷︷b)sign( tj )︸ Timing Cost j
|tj |
|tj |
3.2 Daily estimation
For our first example, the execution data are from a long-short equity hedge fund with approximately $150 million in assets under management and an average holding period of about one month. For each stock traded by this fund, we know the total number of shares bought and sold each day along with the weighted average execution price. In this case, we do not have information on individual intraday trade executions. This is the standard granularity for institutional trading cost analysis, because this information along with a pre- trade benchmark price (such as the previous closing price, the opening price on the day of execution, or the price at the time the order is released) is sufficient to measure implementation shortfall. The chosen example is for AEC, which is the ticker symbol for Associated Estates Realty Corporation, a real estate investment trust (REIT) listed on the New York Stock Exchange with a market cap of around $650 million during the sample period. We examine the fund’s trading in AEC during November and December of 2010. The fund traded a total of 559,356 shares of AEC during this time period on 20 separate trading days. The stock has an average daily volume of roughly 460,000 shares over these two months, so the analyzed trades constitute about 2.8 percent of the total trading volume in AEC during this interval. The implementation shortfall decomposition is illustrated based on daily data and one investor’s trades in a single security. The index j runs over days and the price snapshot is taken at the end-of-day (closing) price, i.e., the bid-ask midpoint at the end of the trading day.
that when the model is implemented at the daily horizon, the liquidity cost component measures the average difference between execution prices and the post-trade efficient price at the close. The gap between trade time and mea- surement of the efficient price argues against making direct use of the numbers as estimates of the cost of temporary price moves when the price decomposition is performed at the daily horizon. Instead, we advocate using this component on a relative basis to compare executions across brokers, across stocks, and over time. To illustrate the breakdown of execution costs across days, Figure 2 plots the size of the total costs and each of its components for each day’s trading. As in Figure 1, the size of the dot is proportional to the amount traded (|tj |).^7
Figure 2 here.
As is often the case with execution cost measurement, there is substantial variation in the costs. Daily implementation shortfalls in this case are between -2.5 and 3.3%. The total costs are highest in the beginning of the sample, especially for the first few large orders, suggesting that the fund quickly became aware of its price impact and subsequently traded in smaller sizes. For these first few large orders, the timing costs are negative, indicating that these orders began when prices were relatively attractive, but the large informational costs quickly swamped the timing benefit. Because we are using an end-of-day post- trade efficient price estimate to split the price impact into liquidity (temporary) and informational (temporary) components, we do not want to overinterpret this part of the decomposition. However, because it is a post-trade price, our liquidity component bears a strong resemblance to the traditional measure of the temporary component discussed earlier. In fact, some traders regularly measure trading costs against a post-trade price. Our innovation is to gain additional insight by using a post-trade efficient price from the state space model rather than use a closing quote or closing auction price.
3.2.1 Re-calculation based on filtered estimates
It is also possible to decompose the implementation shortfall using filtered es- timates of the efficient price instead of smoothed estimates by substituting ˜mj (^7) On most days the fund traded in only one direction. However, on three days the fund bought and sold shares. On those days, only the net trade enters the analysis along with the average price across all trades that day. For example, if the fund bought 35,000 shares at $ and sold 5,000 shares at $16, then the net trade that day was a buy of 30,000 shares at a price of (35,000$15-5,000$16)/30,000 = $14.83.
for ˆmj in equation (15). The filtered estimates yield:
Avg Liquidity Cost 10.4 basis points Avg Information Cost 69.4 basis points Avg Timing Cost -28.2 basis points Avg Total Cost 51.6 basis points Of course, the total implementation shortfall is calculated using observed prices, so it remains unchanged. The timing cost component using filtered estimates is of particular interest, because it has a natural interpretation as the gross short-term alpha conditional on the subset of information included in the model available at the designated pre-trade time (the previous close in this case). Using filtered estimates, the timing cost component for this example is more negative at -28.2 basis points, indicating that an important source of overall return for this investor (or equivalently, an important source of trading cost minimization) is trading against temporary pricing errors.
3.3 Intraday estimation
Our second example uses data from a well-known firm that provides equity transactions cost analysis to institutional clients. We know the size and release time of the parent order, and the size, price, and time of each child order execution. To illustrate the method, we choose one such parent order arbitrarily from a set of recent large orders in less active mid-cap stocks. We also require the order to be executed in one day. The chosen example is a December 13, 2012 sell order in HMST, which is the symbol for Homestreet, Inc., a Nasdaq-listed community bank on the west coast of the U.S. with a market cap of around $360 million. The sell order is for 6,365 shares, and the stock has an average daily volume of 119,000 shares during December 2012. The order is released right around 11:00am, and it is fully completed in 50 child order executions over the space of about 30 minutes. During the half hour from 11:00am to 11:30am, total trading volume in this symbol was 34, shares, so this client ended up trading 18.6% of the total volume during this interval.^8 We estimate the state space model using NBBO midpoints at each minute during regular trading hours for 15 trading days from December 1, 2012 through (^8) There was no news released on HMST that day, and during the 11am-11:30am period, the S&P500 fell by 0.2%, compared to a share price drop of about 2% over the same interval in HMST. Thus, it appears that most of the price moves documented here are idiosyncratic.
Avg Liquidity Cost 48 basis points Avg Information Cost 219 basis points Avg Timing Cost -5 basis points Avg Total Cost 262 basis points The overall implementation shortfall is 262 basis points, and the large infor- mation cost component reflects the fact that this order is selling as the estimated efficient price is falling. The negative timing cost component of -5 basis points simply reflects the fact that the sell parent order was released at a time when the observed midpoint was slightly above the estimated efficient price. Perhaps the most interesting component of our decomposition is the liquidity cost, and it is particularly useful to compare our implementation shortfall de- composition to a more traditional one. Recall that the liquidity cost component measures the average difference between execution prices and the estimated ef- ficient price in effect at the time. While the child orders here execute an average of 48 basis points below the estimated efficient price, the liquidity cost would only be 9 bps if we compare trades to quote midpoints in effect at the time of the child order execution. This is a substantial difference and highlights that the temporary component in prices clearly contributes to the overall trading costs for this order. Figure 3 illustrates the estimates by plotting the observed end-of-minute NBBO midquote, the efficient price estimate, and the investor’s trades each minute. An initial burst of selling coincides with a sharp price decline. We can- not make causal statements, but it is certainly possible that the selling pressure from this parent order caused the price decline. Much of the decline appears to be temporary. The share price bounces back by noon once this order is com- pleted and the selling pressure abates. This armchair empiricism is confirmed by the efficient price estimate, which never moves down as far as the observed quote midpoint and is as much as 14 cents above the midquote during this order execution. The deviation between the observed midquote and efficient price be- gins to appear as child orders begin to execute. After selling 4,365 shares in the space of five minutes from 11:05 to 11:09 (or 23% of the 19,096 HMST shares that trade in this interval), the transitory component reaches its maximum de- viation. Thereafter, execution slows and the transitory component gradually shrinks.
Figure 3 here.
To illustrate the minute-by-minute breakdown of execution costs Figure 4 plots the size of the total costs and each of its components for trades in each minute. As in Figure 3, the size of the dot is proportional to the number of shares filled in each minute. As noted earlier, the efficient price moves down sharply as the first few minutes of selling unfold. This is reflected in the initial upward trend in the informational cost component. The liquidity component increases rapidly from 39 basis points for executions at 11:05am to 64 basis points for the 11:11am fills. Thereafter, the liquidity component generally de- clines, although the scaling of the graph makes this difficult to see. The timing component is constant at -5 basis points, as this illustration is for a single par- ent order. Because the informational costs are by far the largest component of the implementation shortfall, the pattern for total costs closely tracks the informational cost component.
Figure 4 here.
4 Conclusion
In this chapter, we decompose a sequence of observed asset prices into a per- manent and temporary component. We use this price process decomposition to provide a novel and useful decomposition of the standard implementation shortfall transaction cost measure. Investors often think in terms of earning the spread, evaluating individual executions vs. the prevailing quote midpoint. Our methodology provides an alternative benchmark. Individual executions should be evaluated against the estimated efficient price, which can be far from the current quote midpoint (a root-mean-squared average of 51 basis points in the case of HMST, our intraday example). Our methodology also captures the fact that a sequence of trades in the same direction can generate or contribute to a temporary component, and it allows an institutional investor to measure how much its own trading has done so. This seems particularly important in the current automated equity market structure, where transitory price impact may be due to some traders following order anticipation strategies. An institutional investor or algorithm provider can use these empirical techniques to discern whether its algorithms or trad- ing practices minimize these temporary price moves. The empirical examples indicate that the temporary component could be an important contributor to
and uses the pykalman package^10. The EM algorithm is combined with a param- eter grid search for the AR parameters of the pricing error process: ϕ 1 and ϕ 2. The choice for a grid on this subset of model parameters is informed by studying convergence based on random sets of starting values. It turns out that the pa- rameters at the optimum are particularly sensitive to starting values of ϕ 1 and ϕ 2. Grid search involved a grid over [− 0. 8 , − 0. 6 , − 0. 4 , − 0. 2 , 0 , 0. 2 , 0. 4 , 0. 6 , 0 .8]^2 and finer grids with step sizes down to 0.05 around the optimum.
References
Almgren, Robert, Chee Thum, Emmanuel Hauptmann, and Hong Li. 2005. “Equity market impact.” Journal of Risk July:57–62.
Brogaard, Jonathan, Terrence Hendershott, and Ryan Riordan. 2012. “High Frequency Trading and Price Discovery.” Manuscript, University of Califor- nia, Berkeley.
Dempster, Arthur P., Nan M. Laird, and Donald B. Rubin. 1977. “Maximum Likelihood from Incomplete Data via the EM Algorithm.” Journal of the Royal Statistical Society 39:1–38.
Duffie, Darrell. 2010. “Presidential Address: Asset Price Dynamics with Slow- Moving Capital.” Journal of Finance 65:1237–1267.
Durbin, Jim and Siem Jan Koopman. 2001. Time Series Analysis by State Space Models. Oxford: Oxford University Press.
Engle, Robert F. and Robert Ferstenberg. 2007. “Execution Risk.” Journal of Portfolio Management :34–44.
Gˆarleanu, Nicolae and Lasse H. Pedersen. 2012. “Dynamic Trading with Pre- dictable Returns and Transaction Costs.” Journal of Finance (forthcoming) .
Glosten, Lawrence and Paul Milgrom. 1985. “Bid, Ask, and Transaction Prices in a Specialist Market with Heterogeneously Informed Agents.” Journal of Financial Economics 14:71–100.
Hasbrouck, Joel. 1995. “One Security, Many Markets: Determining the Contri- butions to Price Discovery.” Journal of Finance 50:1175–1199. (^10) Package documentation is at http://pykalman.github.com/.
———. 2007. Empirical Market Microstructure. New York: Oxford University Press.
Hendershott, Terrence and Albert J. Menkveld. 2011. “Price Pressures.” Manuscript, VU University Amsterdam.
Kraus, Alan and Hans Stoll. 1972. “Price Impacts of Block Trading on the New York Stock Exchange.” Journal of Finance 27:569–588.
Menkveld, Albert J. 2011. “High Frequency Trading and the New-Market Mak- ers.” Manuscript, VU University Amsterdam.
Menkveld, Albert J., Siem Jan Koopman, and Andr´e Lucas. 2007. “Modelling Round-the-Clock Price Discovery for Cross-Listed Stocks using State Space Methods.” Journal of Business & Economic Statistics 25:213–225.
Morley, James C., Charles R. Nelson, and Eric Zivot. 2003. “Why are the Beveridge-Nelson and Unobserved-Components Decompositions of GDP so Different?” The Review of Economics and Statistics 2:235–243.
Perold, Andr´e F. 1988. “The Implementation Shortfall: Paper versus Reality.” Journal of Portfolio Management 14:4–9.
Shumway, Robert H. and David S. Stoffer. 1982. “An Approach to time series smoothing and forecasting using the EM algorithm.” Journal of Time Series Analysis 3:253–264.
Stock, James H. and Mark W. Watson. 1988. “Testing for Common Trends.” Journal of the American Statistical Association 83:1097–1107.
