ITG Financial Engineering, August 2016
A Side-by-Side Comparison Between
ITG’s Size-Adjusted Spread Cost
Estimates and the “True” Realized
Costs of Institutional Investors
AUTHORS
Onur Albayrak, PhD, Researcher, Financial Engineering
Milan Borkovec, Managing Director, Head of Financial Engineering
INTRODUCTION
ITG’s size-adjusted spread (SaS) cost estimates provide guidance on the
anticipated costs associated with instantaneous spot trade executions, measured
relative to the prevailing mid-quote rate at the trade time1. The underlying data for
our model contains dealer quotes from 6 global banks and 5 major ECNs. By
varying the manner in which we consolidate the limit order book across trading
venues / liquidity providers, we are able to reflect different trading styles and
credit tiers as well as varying degrees of sophistication of market participants.
The use of the empirical limit order book enables us to construct cost estimates
for instantaneous trading at various consolidation levels, deal sizes, as well as at
various times of the day2.
1
For details of the methodology and applications, see for instance Borkovec, Cochrane, Domowitz,
and Escobar, “The Cost of Liquidity in the FX Market”, Journal of Trading, Vol. 20, No. 1 (Winter
2015): 6-19, and Brandes Y., “Spread Revisited: How Size Adjusted Spread Metrics Can Deliver
FX Trading Insights” http://www.itg.com/thought-leadership-article/spread-revisited-how-sizeadjusted-spread-metrics-can-deliver-fx-trading-insights/.
2 ITG’s SaS cost estimates are currently available for four different consolidation types reflecting
the cost of instantaneously sweeping the limit order book for four different aggregation schemes
(“ALL”, “CECN”, “AVG”, “MIN”) depending on the end users’ objectives and capabilities. “ALL” and
“CECN” correspond to consolidating liquidity across all available venues and all ECN venues,
respectively. “AVG” and “MIN” do not involve limit order book consolidation across different trading
venues. These latter schemes rely on the liquidity available from a single bank dealer. “AVG”
computes the cost of climbing the limit order book of each dealer bank separately, and then
calculates the average cost (across banks) for a given trade quantity. “MIN” starts in a similar
manner, but instead of calculating the average cost across all participating banks, it determines the
lowest cost. The cost computed in this manner is typical for a trader who has good business
relationships and sufficient credit standing with all the bank dealers and thus can act on the best
dealer quote provided by the banks for any deal size at any moment in time.
1
In this study, we provide answers to the following important questions:
• Are ITG’s SaS cost estimates reflective of observed costs of spot trades
for large institutional investors?
• Which of the supported consolidation levels best represents the realized
costs of ITG’s FX Peer clients?
• Is the execution performance of ITG’s TCA clients uniform and
consistent across various trade scenarios?
For our empirical analysis, we utilize twenty four most liquid currency pairs in our
ITG FX Peer database in 2015. ITG FX Peer data includes contributions from a
majority of the ITG FX TCA clients. As such, it is not limited to a specific
execution style or trading pattern. The data retrieved from ITG FX Peer database
is cleansed by ensuring the validity of the information in the benchmark and
execution prices. Details are given in the following sections.
Our findings can be summarized in the following bullet points:
• ITG’s SaS cost estimates serve as meaningful benchmarks to assess
the realized FX costs of large buy-side investors.
• ITG’s Peer group database clients pay attention to their executions, as
their average costs mimic remarkably well the cost estimates for the
consolidation type “MIN”, especially for larger deal sizes.
• Our analysis shows that the execution performance can be further
improved for very small deal sizes, as well as during off-London trading
hours.
SETUP OF ANALYSIS
ITG FX Peer database holds the execution information of most of ITG’s FX TCA
buy-side clients in the foreign exchange market. For the year 2015, the ITG FX
Peer database encompases more than 755 thousand trades, for a total allocation
of 2.4T US dollars. At the initial stage of our analysis, we study the reliability of
the trade timestamps and benchmark rates for all of the spot transactions. As
such, we retrieve the benchmark rates from ITG’s internal Market Data (MD)
group, check their validity by comparing it with a second data source and
subsequently perform an actual cost analysis to check whether the trade
timestamps of spot trades are reliable and can be included in our analysis.
Results will be presented in terms of the empirical costs defined by the below
formulas.
10000 for “Buy” trades and
10000 for the “Sell” trades,
where !"#$%&'( is the mid-quote rate at execution time, and )*!#+&! denotes
the execution rate of the trade. Then, we bucket the trades according to the deal
size in base currency.
TESTING BENCHMARK PRICES
An individual spot trade in the ITG FX Peer database contains its trade
timestamp and spot rate that the deal was executed with. In order to compute the
associated realized SaS cost, the prevailing mid-quote rates at trade execution
were collected. To verify the validity of the benchmark rates, we leverage the
existing market database used in ITG’s FX TCA service which aggregates data
from 6 global banks and 5 major ECNs and provides the best bid and offer for
tradeable quotes for any given day/time and currency pair. For robustness of the
test, rates were also used from a secondary data set that uses the same data
2
sources. The secondary data set aggregates these data sources independently
and adopts 15 second buckets to report the prevailing mid-quote rate at the end
time of each bucket. In order to validate the benchmark rate for every trade in
the FX Peer database, we determine which 15 seconds trading interval the trade
corresponds to and subsequently compare the two data points at the end of that
interval. If the two mid-quote rates are more than 1 bps3 away from each other,
we mark this trade as a “bad benchmark trade” and exclude it from the
subsequent analysis.
TESTING TIMESTAMPS
One reason to suspect that trades have incorrect timestamps is exemplified
when execution rates are very different compared to their corresponding
benchmark rates. In order to evaluate the seriousness of this issue, we create an
“acceptable” window around the executed rate, depending on whether the trade
is a buy or sell. We measure how far away from this “acceptable” window the
execution rate is using ITG’s SaS cost estimates for consolidation type “ALL”, as
well as the minimum, and the maximum rate surrounding the trade. More
specifically, we require that the executed rate meets the following two criteria in
order to pass our timestamp test:
,-"+&! . )*!#/-"+&! . ,&*+&! 0 " ∙ 2&2 for Buy trades, and
,-"+&! 3 " ∙ 2&2 . )*!#/-"+&! . ,&*+&! for Sell trades,
where
2&2is the size-adjusted spread cost for the consolidation type “ALL”,
" is a factor showing how far away (in terms of the size-adjusted
spreads) the execution is from the mid-quote rate, and
,-"/,&*+&! is the minimum/maximum rate within the three 15 second
bins surrounding the execution. A chart demonstrating this requirement is shown
in Exibit 1.
Exhibit 1: Illustration of the filtration algorithm for trades with bad
timestamps.
Source: ITG
3
The 1 bps threshold is a reasonable threshold for the 24 currency pairs that are used in this
analysis. Robustness checks in changing the threshold have not resulted in material different
results.
3
In our subsequent analysis we use the value" 4, which is chosen to avoid
discarding too many trades and biasing our results. The asymmetric requirement
on the costs of trading can be justified by the argument that our institutional
clients are not in the business of making the market in FX. In other words, they
are the initiators of their trades, most of the time. Our heuristic makes sure that
we only accept trades where execution timestamps are reasonable, in the sense
that transactions are not associated with huge profits and losses. We tested our
heuristic on a subset of clients’ trades that are known to have valid timestamps in
the ITG’s FX Peer universe and performing very well in terms of transaction
costs, encompassing approximately 20% of the US dollar amount of the
EURUSD trades.
Exhibit 2 shows the resulting cost curves for EURUSD during London hours (i.e.
7:00 to 18:00 GMT) for two filtration methods: the red curve corresponds to our
heuristics discussed above (asymmetric filtration, i.e. size-adjusted spread costs
are added to the trade-initiated side only), while the black curve corresponds to
symmetric filtration. In both cases, we set " 4 and 10 (shown by yellow and
orange curves for asymmetric and symmetric requirements, respectively). The
proximity of the curves in Exhibit 2 and the fact that the choice of " did not have a
major impact on the cost curves suggests that our filtration heuristic is
meaningful4. These findings are consistent across other currency pairs as well.
Exhibit 2: Cost curve comparison as a result of the symmetric and
asymmetric SaS requirement on trade filtration, for EURUSD during London
hours.
Source: ITG
4
The subset of trades shown in Exibit 2, have very low execution costs, pointing to excellent
trading performance compared to their peers. (Compare red curve in Exhibit 2 with orange line in
Exhibit 3)
4
FIXING BAD TIMESTAMPS
Timestamps in FX execution data is notoriously bad, therefore it is crucial to
design a good algorithm to assign correct timestamps to the executions that are
outside of our “acceptable window” definition. Instead of just discarding these
trades, we search for a time period where the execution would be more
reasonable such that it is not associated with huge profits, or huge losses.
Starting from the reported execution time, we search for a 15 second interval
before the trade to find the time period where the execution is inside the
“acceptable window”. The closest time period to the reported execution time is
chosen to be the new timestamp for the execution and the average mid-quote
within that period is used as the benchmark to calculate costs5.
We tested this algorithm on the executions reported in ITG’s FX Peer database
and observed that including the fixed executions have minimal impact on the
empirical costs. Results are shown for EURUSD on Exhibit 3, for the trades
during the London hours in 2015, other currency pairs provide similar behaviour
as well.
Exhibit 3: Cost curve comparison for the corrections applied to the
timestamps, for EURUSD during London hours.
Source: ITG
METHODOLOGY AND DESCRIPTIVE STATISTICS
After the removal of the trades that failed our filters, we move on to calculate the
empirical costs. All clients are aggregated together, except for the clients with
irregular trading patterns (e.g. clients with the majority of their executions at and
around the London close) and trades that are very close to London Close (as
5
If the algorithm could not find an acceptable window within the last 3 hours the trade is excluded
from further analysis. The 3 hours horizon has proven to be a conservative horizon in the sense
that most of the trades can be kept.
5
these trades have different SaS cost characteristics from regular trading). It is
also worthwhile to mention that trades that share the same set of information (i.e.
same client, side, currency, date, and execution time) are combined to form a
single trade, as we consider them as allocation trades that technically belong to
the same deal.
In order to make a meaningful aggregation across all currencies, times of the day
and clients, actual costs are normalized using ITG’s SaS cost estimates of the
consolidation type “ALL” (i.e. available liquidity is consolidated across all venues
and liquidity providers and SaS cost estimates are computed based on that
aggregated liquidity pool). Other ITG SaS cost curves (“ALL”, “AVG”, “MIN”, and
“CECN”), which are also normalized, are also included in the analysis to present
insights about the best fit to the actual trading costs. Trades are analyzed for two
distinct trading periods separately: London hours (between 7:00-18:00 GMT) and
off-London hours.
Descriptive statistics of our spot transaction sample before and after the data
clean-up are shown in Tables 1 and 2 below, together with the US dollar (in
millions) amount that was discarded with our filtration. Table 1 displays the
number of observations and US dollar volume traded for the average client
(aggregated across all currencies for each particular client) in 2015. Similarly,
Table 2 shows the number of trades and the nominal US dollar amount in 2015
for an average currency pair (aggregated across all clients). The results in the
tables are based on all the trades across the day. Both tables indicate that we
can keep most of the trades in ITG’s FX Peer database. The “Removed Trades”
column also includes trades that were discarded due to “activity spikes”
(explained above).
Table 1: Average number of trades and US dollar volume (in mln) for spot
transactions in 2015 for a representative client in ITG FX Peer Database
(before and after the data cleaning).
Client
Packaged
Trades
Removed Trades
Remaining
Trades
Remaining %
Average
17353
Client
Packaged
Trades USD
2327
15026
87%
Removed Trades
USD
Remaining
Trades USD
Remaining %
Average
$83,636.82
$9,734.49
$73,902.33
88%
Source: ITG
Table 2: Average number of trades and US dollar volume (in mln) for spot
transactions in 2015 for a representing currency pair in ITG FX Peer
Database (before and after the data cleaning).
Currency Packaged
Trades
Average
12164
Removed Trades
Remaining
Trades
Remaining %
1883
10281
79%
Currency Packaged
Trades USD
Removed Trades
USD
Remaining
Trades USD
Remaining %
Average
$8,242.65
$50,564.75
77%
$58,807.40
Source: ITG
6
SUMMARY OF ACTUAL SaS COSTS IN ITG FX PEER UNIVERSE
In what follows, we report the results for trades during London hours and offLondon hours separately. Not surprisingly, using only the trades remaining after
filtration (see Tables 1 and 2), we observe that most of the clients are more
active during the London hours, i.e. there are significantly less observations for
off-London hours. Also, the deal sizes are usually smaller for spot transactions
during off hours. As a consequence, we show and discuss cost curves (i.e. costs
as a function of deal size) for London hours only. Costs for small deal sizes are
subsequently compared for the two trading periods separately.
Normalized aggregated cost curves for the most liquid currency pairs are given
in Exhibit 4, overlaid with ITG’s normalized SaS cost curves for all consolidation
types. The currency pairs EURUSD, USDJPY, GBPUSD, USDCHF, AUDUSD,
USDCAD, and EURGBP are used for the aggregation. The average cost curve is
shown by the solid red line and the min/max values (displaying the lowest/highest
normalized average costs for the set of currency pairs) are presented by the by
the same red color with dashed curves. Unlike the “CECN” cost curve, the
relative difference between the actual cost curve and the “ALL” consolidation type
decrease for large trade sizes. We observe a similar behaviour of actual costs to
the “MIN”, where execution is always done with the best (cheapest) dealer in our
dealer universe. Results clearly show that clients appear to execute large deals
mostly when they receive good quotes and thus can outperform our “CECN” and
“AVG” cost estimates. Particularly for the smallest deal sizes, there is still room
for improvement in best execution, as the average aggregated cost in this size
segment is 1.7 times the SaS cost estimate for consolidation type “ALL” and,
most importantly, it is much higher than the “CECN” cost estimate. We should
also note that the “AVG” cost curve decreases relative to the “ALL” for large order
sizes, which could be the result of the banks getting more competitive for the
larger deal sizes.
7
Exhibit 4: Average normalized aggregated cost curves for the eight most
liquid currency pairs in ITG FX Peer Database in 2015 during London hours.
Source: ITG
Next, we study the performance of the smallest deal sizes (corresponding to the
first observation points in Exhibit 3) in more detail. Normalized costs for deals with
notional value of 1 million in the base currency and smaller, together with the
corresponding normalized cost estimates of ITG’s FX cost curves for active
currency pairs, are shown in Exhibit 5A for London hours and in Exhibit 5B for offLondon hours. Similar to Exhibit 4, we see that the performance of ITG’s Peer
clients is in the vicinity of the SaS cost estimates for the consolidation type “MIN”.
These results imply that our SaS cost estimates are meaningful cost benchmarks
and that the institutional clients in ITG’s FX Peer database have the potential to
lower costs further by actively monitoring and executing their spot transactions
through ECNs during London hours. As one can easily see, costs for small deal
sizes are higher during the off hours relative to our SaS “MIN” and “CECN” cost
estimates, suggesting that ITG’s institutional clients should monitor transactions
that are being handled off-hours with at least as much attention as the London
hours.
8
Exhibit 5A: Average normalized costs for small deal sizes in 2015 during
London hours.
Source: ITG
Exhibit 5B: Average normalized costs for small deal sizes in 2015 during
off-London hours.
Source: ITG
CONCLUSIONS
The demand for FX measurement and best execution services is growing. Our
analysis shows that ITG’s SaS cost estimates for FX spot transactions provide a
meaningful benchmark for actual costs of our institutional investor clients in the
ITG FX Peer universe. The FX trading environment is getting “equitized”, with the
recent market shifts to electronic trading through ECNs, the introduction of nolast-look trading venues, and the development of the foreign exchange market
global code by the Bank of International Settlements6. It is expected that we will
observe further declines in transaction costs in the future. ITG’s SaS cost
estimates can guide institutional investors toward that goal.
6 Bank for International Settlements, “FX Global Code: May 2016 Update”,
https://www.bis.org/mktc/fxwg/gc_may16.pdf
9
The results presented in this paper demonstrate that the ITG Peer universe
clients pay particular attention to executions of larger deal sizes during regular
London hours. Their average empirical costs in those size segments are close to
ITG’s cost estimates for consolidation type “MIN”, but slightly higher. We believe
that the performance of those clients can be improved even more if they carry
through the same process for trades of smaller notional values, on and off the
London trading hours.
© 2016 Investment Technology Group, Inc. All rights reserved. Not to be reproduced or retransmitted without permission. XXXX-XXXX
These materials are for informational purposes only and are not intended to be used for investment purposes. The information contained
herein has been taken from sources we deem reliable but we do not represent that such information is accurate or complete and it
should not be relied upon as such. No guarantee or warranty is made as to the reasonableness of the assumptions or the accuracy of
the models or data used by ITG or the actual results that may be achieved.
Broker-dealer products and services are offered by: in the U.S., ITG Inc., member FINRA, SIPC; in Canada, ITG Canada Corp.,
member Canadian Investor Protection Fund (“CIPF”) and Investment Industry Regulatory Organization of Canada (“IIROC”); in Europe,
Investment Technology Group Limited, registered in Ireland No. 283940 (“ITGL”) (the registered office of ITGL is Block A, Georges Quay,
Dublin 2, Ireland). ITGL is authorised and regulated by the Central Bank of Ireland; in Asia, ITG Hong Kong Limited (SFC License No.
AHD810), ITG Singapore Pte Limited (CMS Licence No. 100138-1), and ITG Australia Limited (AFS License No. 219582). All of the
above entities are subsidiaries of Investment Technology Group, Inc. MATCH NowSM is a product offering of TriAct Canada
Marketplace LP (“TriAct”), member CIPF and IIROC. TriAct is a wholly owned subsidiary of ITG Canada Corp.
10