Documente Academic
Documente Profesional
Documente Cultură
Phil Mackintosh
+ 1 212 325 5263
Trade Strategy
Laurent Boldrini
+44 20 7888 3128 Evolution of Impact Cost Modeling
Key points An Evolution Rather than Revolution
Trading costs can be a significant contributor to portfolio performance, and by
Electronic execution and high frequency
inference, stock selection. Not surprisingly, better understanding and
trading has revolutionised markets, but
estimating these costs is critical to the success of investors around the world.
pre-trade systems evolved far less.
In this report we look at how impact cost Impact Cost estimation methods have changed significantly in the space of a
models have evolved over time: generation. The advent of computers, with ever increasing capacity, followed
by electronic exchanges and tick databases makes more complex calculations
Inventory risk & cost models:
far easier to complete. However from a mathematical standpoint, this has
Based on option pricing concepts,
been more of an evolution than a revolution as weaknesses caused by key
assume a liquidity provider has a one-
assumptions were gradually overcome. In this report, we discuss the models
sided payoff based mostly on a stocks
that became industry standards throughout the last two decades.
volatility risk.
Semi-empirical models: moved Cost Estimates Should Affect Stock Selection
away from option based methodology, We know that smaller trades, and more liquid names, cost less to trade. For
and fit more flexible curves using more that reason, it is critical that portfolio managers dont naively construct
factors of trade cost. portfolios based purely on alpha scores. In some cases, round trip costs in
Optimal Schedule models: Split the expensive stocks to trade can offset the additional alpha earned from growing
concept of risk and return, a big the trade size. Optimising alpha against post-trade-cost returns adds an
improvement on inventory models. additional constraint (and complexity) to the portfolio construction process
They also elegantly solve the trading but it should resultant in more of the alpha being captured.
trade-off (cost vs risk).
But even with the improvements over time, Trade Strategy Affect Execution Results
we think cost models are still not that Impact cost models are also important for traders as they help identify
helpful to traders. Specifically, these trades that have high expected costs & liquidity as well as providing a
models rely far too much on volatility: benchmark for normal shortfall. We see proof of trader reliance on pre-
Market-wide shifts in volatility cause a trades in practice in our execution systems. We find an unusually large
relatively small change in average proportion of executions are submitted at 20-30% participation, consistent
execution costs with the typical conclusion from the latest evolution of cost models that we
Real execution risk evolves in a very discuss in section 4 below.
non-normal way (unlike stock volatility) Exhibit 1: In-line executions mostly use around 20% aggression
Event risk is more important than 20,000
18,000
annualized stock volatility Count
16,000
Time decay of alpha is more important 14,000
than the time decay of volatility 12,000
All Day
7.5
6 hours
10
3 hours
20
25
1 hour
Participation
30
30 mins
40
5 mins
50
(%)
30 Sec
100
Duration
Source: Credit Suisse Portfolio Strategy
Portfolio Strategy
0
2. Trading Faster = A little More Costs
5
Trade
1
7.5 10 15 20 25 30 40
0.1
Size 0 1 2.5 5
Aggression (%)
(ADV%)
Also in chart 2, moving from left to right shows the changes in
impact cost caused by more aggressive execution. This shows
Exhibit 3: Volatile stocks costs more to trade that for each sized [colored] orders, aggression increases cost.
70 70
Low Volatility Stocks
60
High Volatility
Stocks
As the slope of (2) is not as steep as for (1)
60
Average Average above, we conclude aggression has a less
Shortfall Shortfall
(bps)
50
(bps)
50
significant cost than trade size.
40 40
30
3. More Volatile = More Costs
30
Separating the results into volatile and less-
20
15
20
20
20
volatile baskets, we see that more volatile
15
10
7.5
5
10 10
7.5
5
10 stocks cost more to trade (for the same
Trade Size 2
relative trade-size), especially as size
2
1 Trade Size 1
(ADV%) 0.5 0 0
(ADV%) 0.5
25 30 40
1 2.5 5 7.5 10 15
Aggression (%)
20
1 2.5 5 7.5 10 15 20
Aggression (%)
25 30 40
increases (exhibit 3).
Exhibit 4: Longer trades have a wider range of results 4. Trading Slower = More Risk
Deviation
60
We also found that as the duration of a trade increased, the
(bps) 50 dispersion of shortfall results increased (exhibit 4). Importantly:
The magnitude of this variation was non-trivial (large)
4
40
0
2.5
5
7.5
10
15
6hours
25
3hours
30
1hour
40
30 mins
50
5mins
75
Participation
30Sec
100
bad executions
30 from being in the trade before other market participants. Even for
Small cost
increase information-less trades, market trends can also affect shortfall
10
and the slower trades are executed, the more exaggerated this
-10
effect becomes. Put simply, trading becomes a trade-off.
-30
Not surprisingly impact models also treat uncertainty, as a cost in
-50
0 10 20 30 40 50 some way. However many generalize to use a stocks volatility as
Agression (%) a symbol of trend risk. As we show in the following pages this
Source: Credit Suisse: Portfolio & Derivatives Strategy generalization can cause issues for model users.
2
Portfolio Strategy
12
More volatile stocks do tend to cost more than execute than less volatile
10
8
stocks (Exhibit 3)
6 This can be seen in Exhibit 6, where we fit an inventory risk model to our
4 latest real shortfall data (note: this line looks straight because the bottom axis
2
0
is a quasi-log scale it is actually a root function). This example uses
1 2.5 5 7.5 10 15 20 25 30 40 average volatility of 20% - as our data is from 2006/7, this is reasonable.
Trade Size (% ADV) Weaknesses
Assumes constant aggression
The most obvious, and significant, weakness of this model was its
assumption of a constant aggression level. There is no flexibility to adjust a
trade for execution strategy. This fails to address a key component of
execution analysis the tradeoff between aggression and impact. And it
leaves the model too inflexible for real-world trading decisions.
Some also say these models have a fixed 1-day duration, and in reality many
were calibrated using OTD data (just like we have in Exhibit 6). This is
contrary to the concept of inventory at-risk mathematics as if we assume
1-day trades, it means all executions are at different levels of aggression.
Inventory Risk isnt the same all day
Exhibit 7: Exposure to (Trade & Positions) A key assumption of this model is also that cost evolves in the same way that
140 100 exposure to volatility evolves as a root function of time. This is true for an
Evolution of exposure 90 option, where the whole position expires simultaneously. But there are two
120
(for a Position) 80 important mathematical implications for this model:
100 70
In the calculation above, the term for time (t) is actually just % of daily
Volatility (bps)
Trade Residual
60
80
50
volume. For an order to actually be at risk for this time, would actually
60 Evolution of exposure require the trade to executed at 100%.
40
(for a Trade)
40 30 Trades are usually executed in slices, over time. This incrementally
Residual
20 reduces the actual exposure to volatility throughout the day. Over the life
Trade Exposure
20
Order Exposure 10 of an order, this means that during the day, inventory risk is actually far
0 0 from a root function (see Exhibit 7, which shows the actual root function
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 exposure to (red line) versus the actual exposure to from a trade
Time (Days) (modeled using a 20-vol stock with 10 slices traded TWAP over the day).
However, this latter point is somewhat academic as the root function is a
good fit to observed execution data, and the integral of the outstanding
position can easily be included in the calibration of the constant k.
3
Portfolio Strategy
Model Cost
Days to complete
70
8.0 Ultimately, increasing aggression actually reduces the expected cost.
60 Model Time
50 6.0 This is the opposite of what we observe in the market! We see in Exhibit 9,
40 for very patient executions, the modeled costs (blue line) also increase
30
4.0 exponentially, while observed costs (blue bars in exhibit 9) are falling.
20
2.0 2. Confusing risk, with cost?
10
Our main complaint with these two models however is the use of volatility,
0 -
1% 5% 10% 15% 20% 25% 30% 35% 40%
most commonly associated with risk, to calculate cost and the implied
Aggression assumption that variation will go against the market maker. We discuss this
Source: Credit Suisse Portfolio Strategy in more detail on the next page.
4
Portfolio Strategy
But options payoffs really are one sided and put-call parity proves
that stocks are (almost) as likely to go up, as to go down.
We also find that execution results are (almost) as likely to be
positive, as negative (Exhibit 10). Over most execution timeframes
-200 -175 -150 -125 -100 -75 -50 -25 0 25 50 75 100 125 150 175 we find that a trader has around a 44% chance of actually beating
Implementation shortfall (bps)
Source: Credit Suisse Portfolio arrival price. Consequently, the correct way to interpret stock
Strategy volatility and its effects on execution results is to look at Exhibit Y3
in two dimensions:
The average () represents the expected cost.
The the stocks volatility () describes the distribution of results
Non-normality reduces confidence anyway!
Although the returns of a stock are said to be stochastic (random and assumed to be normally distributed), the execution results
are clearly non-normal. Compare the normal curve [exhibit 10, red line] with the actual results distribution [blue bars].
The fact that intraday returns are also far from normal contributes to this. In an earlier report we highlighted the pointy
mountains of single stock daily returns (Looking for Alpha in All the Wrong Places - Part 2, see exhibit 11). This shows that
most stock returns are far more boring (stable) than predicted by a normal distribution but some of the time, stocks move far
more than we expect. This is typically called a fat-tailed distribution.
This feature has implications for any model using a standard volatility calculation. As these imply that stock returns will evolve in
a normal way, real execution outliers will be fewer than predicted, but will have far more impact on the P&L.
Exhibit 11: Actual distribution of daily stock returns
8000
Actual Distribution
7000
Normal Distribution
600
6000
Actual Distribution
Normal Distribution
Observations
500
5000
400
Observations
4000
300
3000
200
2000
100
1000 0
2 3 4 5 6
0 Standard Deviations
-5 -4 -3 -2 -1 0 1 2 3 4 5
Standard Deviations
Source: Credit Suisse Portfolio Strategy
5
Portfolio Strategy
Survey says: Spreads rise with Volatility Exhibit 14: Average Bid-Ask spreads for S&P500 stocks,
Our recent analysis in As The Dust Settles: Analyzing compared to the VIX
Microstructure Changes, showed that spreads were clearly
affected by, and very correlated to, volatility (exhibit 14).
However, spread costs are reasonably certain and easy to
calculate so they are usually a separate component of an
impact cost model.
Additionally, although spreads increased, their contribution to
total impact cost is usually a small fraction and the
magnitude would not explain the change in Exhibit 13.
6
Portfolio Strategy
VIX (Pts)
However, our findings in exhibit 15 seem to show that a 25 40
higher volatility regime, on its own, didnt affect average 20 30
shortfall much at all. 15
20
In fact, during the crisis period we also saw that (especially for 10
5 10
VWAP orders):
0 0
Shortfall increased most across more aggressive May-08 Aug-08 Sep-08 Nov-08 Nov-08 Feb-09
strategies (as they crossed the spread more often) and Source: Credit Suisse Portfolio Strategy
Exhibit 16: Changes to shortfall during the crisis period
The amount of degradation in performance was close to
-
the actual increase in spreads (see exhibit 16)
Shortfall Deterioration in
(2)
And was far from the 4x increase that the market
(4)
Crisis (bps)
experienced (and expected in Exhibit 12)
(6)
Very passive orders (which cross the spread less) were Inline
(8)
roughly in line with historic shortfall. VWAP
(10)
This seems to indicate that market-wide shifts in volatility
(12)
increase trading costs mostly because they increase spreads. Agressive Moderate Passive
This means that inventory cost models would have Execution Style
systematically overstated expected cost during the crisis. Source: Credit Suisse Portfolio Strategy
A Beta way to Calibrate? Exhibit 17: Example Changes to 6 month Volatility & Beta
It seems that although volatility is a useful tool to 80% 7
rank the relative trading impact between stocks it
is actually not as important to determine absolute 70% Volatility 6
impact especially during a market-wide volatility 60% Beta
regime shift. 5
50%
Volatility
Beta
normalize for market volatility (m). Doing this 40%
effectively would remove the impact of market-wide 3
30%
volatility increases, but retain the relationships of 2
higher and lower volatility stocks 20%
Potentially, replacing with Beta, which could 10% 1
achieve this. Beta includes stock volatility (s) as 0% 0
well as a market volatility divisor of (m). Our results Jul-07 Dec-07 May-08 Oct-08 Mar-09
show that this was a far more stable metric over the
recent market sell off (see Exhibit 17) . Source: Credit Suisse Portfolio Strategy
7
Portfolio Strategy
3. Semi-Empirical Models
Description & History
Mathematics: In a paper published in the Journal of Risk in 2005; Almgren, Thum,
Almgren et al tested a variety of descriptive Hauptmann and Li proposed that impact is a 3/5 power law, with specific
factors and fit a curve to the average dependence on trade duration, daily volume, volatility and shares outstanding.
shortfall observed from actual executions. (See: Almgren Impact Cost 2005)
They discovered the importance of liquidity, This model does away with the assumptions of inventory cost, where
aggression and duration to trade cost exposure to a stocks volatility over time is responsible for the accurate pricing
of an execution. Instead, they prove by regression that what factors had the
They considered execution to have 2 most descriptive power:
components:
Relative Turnover is important: measured as ADV/shares outstanding,
Permanent cost: created by the
information from a trades direction and Aggression is important: measured as trade/total shares traded
size. This introduced the concept of Duration is important: Espcially as combined with agression it reflects
signalling risk trade size
Temporary cost: reflecting the Stock volatility is important.
market move required to attract Market cap is not important.
liquidity in a short timeframe.
Bid Ask Spread was not important
A generic formula for these models is:
The best curve was a 3/5th power: compared to the root curves used
Cost = (X/ V)*( / V) + sgn(X) |
X.
|
3/ 5 in inventory models this has slightly lower costs for small orders, but
VT with higher costs for large trades.
Where = 0.314 Consequently, these are considered empirical (cost) models.
= 0.142
X = Trade (Shares) Strengths
V = Avg Daily Volume
Critically, because this model sought to fit a curve to the data, the cost of very
T = Time (Duration)
= Shares Outstanding passive executions was shown to be smaller than for larger orders.
This was also instrumental in highlighting other important drivers of cost. In
addition to volatility and normalized trade size (trade/ADV). It also led to their
separation of the permanent and temporary costs which highlighted the
trade-off between duration (more signaling) and aggression (less time for
traders to provide offsetting liquidity).
Weaknesses
On its own, this model typically suggests that all executions should be
executed very slowly as that reduces the temporary impact. However this
solution ignores losses created by trend and opportunity costs.
However, we highlight that a solution to this problem (summarized in part 4,
over) was actually published (also by Almgren) well before this 2005 paper.
8
Portfolio Strategy
T re n d C o s t
Exhibit 19: Almgren Chriss in Theory Exhibit 20: Optimum paths for different risk aversion
Cost (bps)
80 80
60 60
40 40
20 20 Optimum
- -
0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70
Agression (%) Agression (%)
9
Portfolio Strategy
Weaknesses
The major weakness of this approach is that a naive application of the
process can result in suboptimal trading and as we show on the next page,
more aggressive trading than needed. For reasons detailed below we think
an intuitive understanding and use of this model is difficult.
This model also use a linear impact assumption however as we discussed in
section 3, Almgren has since created a better cost curve.
1. Optimal solution is not the lowest cost
The fact this model has an optimal solution is very neat, and quite elegant
mathematically. However we think many traders confuse the risk term with
trend cost and assume the optimal cost is actually the sum of both terms,
rather than the trade-off between risk and return. In fact, Almgren & Chriss
specifically note that the expected value of the risk term = 0 (exhibit 21).
2. How do you calibrate ?
Perhaps more importantly, Almgren & Chris talk about optimizing for all levels
of risk aversion. Used correctly, the term would change be lower for
cashflow or delta neutral trades, and higher for event driven or single sided
trades (as they exhibit different amounts of execution risk).
In practice this is rarely done. Partly because it is difficult to interpret the
term especially because it is expressed in $2 which means the value of
tends be un-intuitive (its just a very small number).
We also think this explains one of the more unusual observations from page 1
why such a high proportion of in-line executions target 20% participation
(see exhibit 1). This is likely because all traders use a similar generalized
term leading to the same optimal aggression level. In the real world, trade
dynamics should create a lot more dispersion (or granularity) of execution
strategies, as we discuss more in later pages.
3. Still Uses Stock Volatility as Risk Proxy
Finally, these models still calibrate their risk term using each stocks historic
volatility. As we discussed on page 5, execution results are both non-normal
& actually far lower than the exposure to volatility x t, even after
accounting for incremental executions over time. Basically, the inventory
cost has made a reappearance as the execution risk. But we can see from
Exhibit 22 how different actual execution dispersion (green line) is to the
estimate using x t (blue line) and full day exposure (red line)
Exhibit 21: Slippage over execution time Exhibit 22: Slippage over execution time
1.4%
Volatility Exposure
1.2% Working Order Slippage
VWAP - Past 2
Years Observed Slippage
1.0%
Std Dev of Prices
0.6%
E[f()] = 0
0.4%
0.2%
0.0%
-200 -175 -150 -125 -100 -75 -50 -25 0 25 50 75 100 125 150 175 0.0 0.2 0.4 0.6 0.8 1.0
Implementation shortfall (bps) Time (Days)
Source: Credit Suisse Portfolio Strategy
Source: Credit Suisse Portfolio Strategy
10
Portfolio Strategy
25
20 Execution Risk
Using real results (for cost and risk) and Almgren & Chriss
15 Impact Cost approach, we can estimate the frontier for our real
Efficient Fronteir executions:
10
5 Note that the Almgren and Criss optimization first
- converts risk (Blue area) to $2 shrinking its value
0 10 20 30 40 versus impact cost.
Agression (%)
Our hypothetical real fronteir (green line) = Cost + x
Source: Credit Suisse Portfolio Strategy Risk.
For our selected level of risk aversion, this shows that the
Exhibit 25: Frontier with theoretical () exposure frontier is optimal around 5% aggression.
10% Trade Size - Model Fronteir
70 3. But Almgren & Chriss use Theoretical Risk
60 Risk However in their model, Almgren & Chriss use theoretical
Efficient Fronteir
50 Impact Cost
risk based on the stocks exposure to its volatility () over
time. This is similar conceptually to what the inventory risk
Cost (bps)
40
models called cost, as the risk component reduces as in
30 proportion to time.
20
Not surprisingly, this curve actually starts higher, and
10 decays slower, than our real execution results. As a result,
- the models optimal result is much more aggressive (see
0 10 20 30 40 50 60 70 exhibit 25).
Agression (%)
Source: Credit Suisse Portfolio Strategy
11
Portfolio Strategy
1.00%
0.80% A cashflow trade may have no new information, especially if it is into an
0.60% index fund.
0.40% A value fund may find that they have a lot of valuable information but
0.20% that the market takes months to adjust for it
0.00% A momentum trader may find that their information decays faster, over a
Source: Credit Suisse Quant Research, Portfolio Strategy matter of weeks.
An event driven trader may have a huge expected return, but must trade
quickly before the event they anticipate becomes public.
Total Alpha differs by signal
Exhibit 27: Days before Alpha maximizes We can quantify this effect by looking at the performance of the alpha factors
100%
from our Quant Research team (see exhibit 26 & 27). These show that total
90% alpha varies across factor. Reversion strategies have the highest alpha Value
80% the lowest.
70%
But pre-trade alpha of just 100bps can easily be reduced significantly by
Alpha Decay
60%
trading costs especially for very large trades. The selection of aggression
50%
may be critical to retain more alpha.
40%
30% Price Reversal And Alpha Decay Varies by signal
20% Earnings Momentum In addition, different alpha signals decay at different rates. For example, 60%
10% Relative Value of the factor performance occurs in:
0%
3 days for their price reversal factor
0 10 20 30 40 50 60 70 80 90 100 110 120
Days 3 weeks for their momentum factor
Source: Credit Suisse Quant Research, Portfolio Strategy
Around 3 months for their value factor.
12
Portfolio Strategy
In this instance, optimal trading strategies are pretty clear from the different
dynamics of quant factors and are clearly more related to the alpha factors
Exhibit 28: Correlation increases as traders than dispersion of execution results:
focus more on broad economic news, and less
The factor with the most alpha also has the fastest alpha decay. This
on stock specific news
clearly lends itself to faster trading which is required to minimize
opportunity cost, but can be borne thanks to the stronger expected
performance.
Conversely value trades have less alpha, but also need to be owned for
months to capture the alpha. Slowing these executions may save more
costs than the alpha that would accumulate.
X-Treme Traders? Event Driven get Aggressive
Most investment strategies rely on new information in some form-or-other to
generate their outperformance. But event driven traders have some of the
most significant alpha (and fastest rates of alpha decay).
The addition of news can lead to an instantaneous revaluation of a company,
Source: Credit Suisse Quant Research, Portfolio Strategy industry, or even the whole market. Sometimes this causes the market to
gap; even when markets are closed (no signaling is required). [This also
explains why stock markets are not entirely random (or normal), why
Exhibit 29 Earnings Surprises vs Price Moves in correlations, beta & volatility change over time; and also why stock returns are
Q4-08 (Market Adjusted) fat tailed (see exhibit 28)].
less than -50%
-50% to -40%
Some typical event driven trades include earnings releases & takeover
-40% to -30%
announcements (see exhibits 29 & 30), and they affect all traders in a stock:
-30% to -20%
For event driven traders the eventual arrival of news should create
-20% to -10% Avg negative
-10% to -1% surprise px increased urgency to complete and adverse or unexpected news, a
0%
reaction (-3.7%)
reason to aggressively exit a position.
Avg positive
1% to 10%
surprise px Even for other (nave) traders, it significantly increases execution risk
10% to 20% reaction (1.7%)
20% to 30%
which could be reduced by trading the specific stock with more urgency
30% to 40%
than the alpha signal may suggest.
40% to 50%
14.0
(TCHC May 6, 2008)
80,000
However this means portfolio managers need better understand their trade-
70,000
lists and traders will need more information from them. It also means
13.5
60,000
traders need work harder, as execution strategies might ultimately need to be
13.0
different for each stock in a trade-list.
Volume (shares)
50,000
12.5
Price
40,000
12.0
30,000
11.5
20,000
11.0 10,000
10.5 -
9:30 10:30 11:30 12:30 13:30 14:30 15:30
13
Source: Credit Suisse Portfolio Strategy
Portfolio Strategy
Raymond Hing +44 20 7888 7247 raymond.hing@credit-suisse.com This material has been prepared by individual traders or sales personnel
of Credit Suisse and its affiliates ('CS') and not by the CS research
Laurent Boldrini +44 20 7888 2041 laurent.boldrini@credit-suisse.com department. It is not investment research or a research recommendation,
Marwan Abboud +44 20 7888 0082 marwan.abboud@credit-suisse.com as it does not constitute substantive research or analysis. It is provided for
informational purposes, is intended for your use only and does not
Sikandar Samar +44 20 7888 0604 sikandar.samar@credit-suisse.com constitute an invitation or offer to subscribe for or purchase any of the
products or services mentioned. The information provided is not intended
Asia to provide a sufficient basis on which to make an investment decision. It is
Murat Atamer +852 2101 7133 murat.atamer@credit-suisse.com intended only to provide observations and views of individual traders or
sales personnel, which may be different from, or inconsistent with, the
observations and views of CS research department analysts, other CS traders or sales personnel, or the proprietary positions of CS. Observations and views
expressed herein may be changed by the trader or sales personnel at any time without notice. Trade report information is preliminary and subject to our formal
written confirmation.
CS may, from time to time, participate or invest in transactions with issuers of securities that participate in the markets referred to herein, perform services for or
solicit business from such issuers, and/or have a position or effect transactions in the securities or derivatives thereof. The most recent CS research on any
company mentioned is at http://www.csfb.com/researchandanalytics.
Backtested, hypothetical or simulated performance results have inherent limitations. Simulated results are achieved by the retroactive application of a
backtested model itself designed with the benefit of hindsight. The backtesting of performance differs from the actual account performance because the
investment strategy may be adjusted at any time, for any reason and can continue to be changed until desired or better performance results are achieved.
Alternative modeling techniques or assumptions might produce significantly different results and prove to be more appropriate. Past hypothetical backtest
results are neither an indicator nor a guarantee of future returns. Actual results will vary from the analysis.
Past performance should not be taken as an indication or guarantee of future performance, and no representation or warranty, expressed or implied is made
regarding future performance. The information set forth above has been obtained from or based upon sources believed by the trader or sales personnel to be
reliable, but each of the trader or sales personnel and CS does not represent or warrant its accuracy or completeness and is not responsible for losses or
damages arising out of errors, omissions or changes in market factors. This material does not purport to contain all of the information that an interested party
may desire and, in fact, provides only a limited view of a particular market.
14