FX Execution Algorithms: TWAP, VWAP, the Fix, Risk Transfer and Transaction Cost Analysis

An FX execution algorithm is software, usually supplied by a bank or a specialist provider, that splits a large currency order into many small “child” orders and works them over time and across trading venues to reduce the cost of execution. Institutions choose between an algorithm, a risk transfer (one price from a dealer for the whole amount, immediately) and a fixing order (a trade at a published benchmark rate such as the 4pm WM/Reuters fix), and they measure the result with transaction cost analysis (TCA).

In short: every large order faces a trade-off between paying a dealer to take the risk now, paying the market’s spread and impact gradually while bearing the price risk yourself, or accepting a benchmark rate set by everyone else’s trading. A BIS Markets Committee study group estimated in 2020 that execution algorithms handled 10-20% of global spot FX, or roughly $200-400 billion a day. They lower expected costs, but they shift market risk from the dealer to the client, and for large orders that risk is usually much bigger than the cost being saved. The FX Global Code asks algorithm providers to disclose how their algorithms work, what they charge and where conflicts arise, and encourages standardised TCA data.

This sits alongside how the FX market is structured and the FX fixing scandal, which explains why the 4pm fix was redesigned in 2015.

The primary sources

BIS Markets Committee, “FX execution algorithms and market functioning” (Markets Committee Papers No 13, October 2020), produced by a study group chaired by Andréa Maechler of the Swiss National Bank, drawing on a survey of algorithm providers and users.
The FX Global Code (December 2024 version), Principle 10 on order handling and fixing orders and Principle 18 on algorithmic trading and aggregation services, with the Global Foreign Exchange Committee’s TCA Data Template (July 2021).
Martin Evans, Peter O’Neill, Dagfinn Rime and Jo Saakvitne, “Fixing the Fix? Assessing the Effectiveness of the 4pm Fix Benchmark”, FCA Occasional Paper 46 (October 2018).
Martin Evans and Dagfinn Rime, “Microstructure of foreign exchange markets”, Norges Bank Working Paper 6/2019, which measures execution costs on EBS.

Three ways to execute a large order

Risk transfer. The client asks one or more dealers for a price for the full amount and trades it at once. The BIS describes the dealer’s spread as “the compensation paid to the liquidity provider for taking on the market risk,” an “almost instantaneous ‘all-in’ price that typically depends on the transaction size, the prevailing liquidity conditions and market volatility.” The client’s cost is known before it trades.

Execution algorithm. The client hands the order to an algorithm, which trades in pieces. In the BIS’s words, the result is “a trade-weighted average price known only at the end of execution plus an associated fee.” The client keeps the market risk until the last piece is filled.

Fixing order. The client asks a dealer to trade at a benchmark rate, most commonly the WM/Reuters 4pm London closing rate. The client gets exactly the benchmark, plus or minus an agreed charge, and the dealer takes on the job (and risk) of trading around the fixing window. Passive managers whose portfolios are valued at the fix may prefer this, because otherwise, as the FCA paper notes, they incur tracking error against their benchmark.

Why size is expensive: depth on the order book

Quoted spreads in EUR/USD are a fraction of a basis point, but quoted size is limited. Evans and Rime’s Norges Bank paper computed what it would have cost to execute market orders of different sizes against the EUR/USD limit orders actually resting on EBS, sampled every minute of the London day over a month:

Order size	Share of minutes with enough depth to fill	Median cost vs mid (basis points)
$10 million	100%	0.91
$20 million	99.96%	1.27
$40 million	98.50%	1.92
$50 million	91.31%	2.18
$60 million	66.10%	2.33
$100 million	9.26%	2.39

The cost per unit rises with size, and beyond about $50 million the book usually cannot absorb the order at all. The authors note that splitting a large order “can produce an adverse shift in limit orders before all the market orders are executed,” so “the cost of executing a USD 100 million purchase will be greater than 10 times the cost of a USD 10 million purchase.” They also warn that spreads measured in the 4pm fix window “are poor approximations to the spreads applicable for the overwhelming majority of spot trades.” Every institutional execution method is an answer to this depth problem.

The algorithm types

The BIS report classifies FX algorithms into six families:

Time-sliced (TWAP): equal slices at regular intervals. The BIS calls the TWAP “the most simplistic EA” and “the most widely used type of algorithm.”
Historical volume-sliced (VWAP): slices sized to the usual intraday volume pattern.
Percent-of-volume (POV): trades a set share of observed market volume, so it speeds up when the market is busy.
Pegged or tracker: rests passive orders at or near the best bid or offer, trying to earn the spread rather than pay it.
Implementation shortfall: uses impact and volatility models to balance the cost of trading fast against the risk of trading slowly.
Limit-based or sweeping: takes available liquidity aggressively up to a limit price.

Many providers also offer hybrids that switch between types as prices move.

The execution trilemma

The BIS frames algorithm choice as a trilemma built on André Perold’s 1988 definition of implementation shortfall: the difference between the price when the decision was made and the price actually achieved. It has three parts:

Market impact and spread costs: the price moves against you because you are trading, plus the spread you cross.
Market risk: the price moves for reasons unrelated to you while the order is being worked.
Opportunity cost: the order may not complete in the time allowed.

“No single EA can optimise all three components of implementation shortfall at the same time.” A sweeping algorithm minimises market risk by trading immediately but has the largest impact. A pegged algorithm minimises impact but is “exposed to substantial market risk as execution tends to be slow.” A TWAP gives “a high degree of execution certainty” but cannot respond to conditions. A POV adapts to liquidity but “cannot guarantee completion within a fixed period.”

Worked example 1: reading a TCA report

TCA compares the price achieved with a benchmark. The GFXC’s TCA Data Template, introduced in July 2021 after the BIS report, sets out the fields an algorithm provider should give the client for each parent order, including the mid at arrival, the average rate excluding and including the algorithm fee, and an estimated risk transfer price for the whole order at the start time.

The template’s own illustrative sample is a sale of EUR 6 million against the yen. Computed in Python from its sample fields:

Measure	Rate	Cost vs arrival mid
Mid at arrival	120.711000	(benchmark)
Average rate, excluding fee	120.706158	0.40 bp
Average rate, including fee	120.703319	0.64 bp
Estimated risk transfer price	120.695000	1.33 bp

The algorithm fee works out to 0.24 basis points. Including it, the algorithm beat the risk transfer price by 0.69 basis points, about JPY 50,000 on EUR 6 million. That is how TCA is read: against the arrival price for total cost, against the risk transfer price for whether working the order was worth it, and with and without fees so that charges cannot hide inside the average rate.

One order proves nothing, because the market may simply have moved favourably while it was being worked. TCA is meaningful only across many orders, which is why the template standardises the fields.

Worked example 2: risk transfer vs TWAP vs the fix

The next example is stylised. The costs are assumptions chosen to be in a plausible range for a liquid major pair, not market quotes; the risk figures follow from an assumed volatility. All of it was computed in Python.

A fund must buy EUR 500 million against dollars, about $575 million at an assumed rate of 1.15, and decides to do so at 2pm London. One basis point is worth $57,500. Assume EUR/USD volatility of 7% a year, spread evenly over 252 trading days of 24 hours (the convention used in our last look example), which is about 9.0 basis points per hour.

For a TWAP over T hours, the standard deviation of the average price relative to the arrival price is the hourly volatility times the square root of T/3. For a fixing order, the relevant risk is the move between 2pm and the 4pm fix, two hours.

Method	Assumed expected cost	Expected cost ($)	Standard deviation of outcome	95% range ($)
Risk transfer now	1.50 bp	86,250	0	known in advance
TWAP over 1 hour (incl. 0.25 bp fee)	0.85 bp	48,875	5.2 bp	+/- 586,000
TWAP over 4 hours (incl. 0.25 bp fee)	0.60 bp	34,500	10.4 bp	+/- 1,171,000
Fixing order at 4pm	0.50 bp	28,750	12.7 bp	+/- 1,435,000

Three lessons come out of this table.

Risk dwarfs cost. The algorithm saves about $37,000 in expected cost over the risk transfer, but exposes the fund to a one-standard-deviation outcome of about $299,000 either way. The savings are real on average; on any single order they are noise.
Slower is cheaper and riskier. Stretching the TWAP from one to four hours lowers expected impact but doubles the risk.
The choice depends on risk appetite. Scoring each method as expected cost plus a penalty times its standard deviation, the one-hour TWAP beats risk transfer only if the penalty is below about 0.125 basis points per basis point of risk (for the four-hour TWAP, about 0.09). A fund that cares mainly about average cost over thousands of orders will use algorithms; a corporate treasurer who must report a known rate for one large payment may rationally pay the risk transfer price.

The fixing order looks cheapest and riskiest, but its risk is measured against the 2pm decision price. For a fund whose benchmark is the fix, that risk largely disappears, because its performance is measured at the same rate. Choosing the benchmark before choosing the method is the first step in professional execution.

The 4pm fix: how it is calculated and what changed in 2015

The WM/Reuters closing spot rate is the most widely used FX benchmark. The FCA paper describes the method: one trade is sampled each second from the reference interdealer venues during the fixing window, medians are taken separately of trades at the bid and at the offer, and the fix is the mid of those two medians.

After the 2013-2014 rigging scandal, the Financial Stability Board recommended on September 30, 2014 that the window be widened. On February 15, 2015 it moved from one minute (3:59:30 to 4:00:30 London time) to five minutes (3:57:30 to 4:02:30). Evans, O’Neill, Rime and Saakvitne studied the effects with trader-identified data from Thomson Reuters Matching:

Price efficiency improved. Short-term price reversals around the fix “decrease steadily throughout our sample period, and disappear from 2015 onwards.”
Manipulation became harder. Simulated outlier trades had “half the impact with a 5-minute fix window in comparison to a 1-minute window,” although the impact was under 1 basis point in both cases.
Tracking error rose. Lengthening the window made the fix harder to replicate by trading, worsening attainability “by a magnitude of between 2 and 5 times for the largest currencies,” which “significantly increases the tracking error of market participants, and thus trading costs, for those participants that use the benchmark for rebalancing purposes.”

The BIS report adds a warning from market participants that users of “simplistic” TWAP algorithms executing against the 4pm fix “can be systemically disadvantaged” because their trading is predictable. The FX Global Code’s Principle 10 governs dealers handling fixing orders: they “should not intentionally influence the benchmark fixing rate to benefit from the fixing,” should not share client fixing information, and should price fixing transactions “in a manner that is transparent and is consistent with the risk borne.” The history behind those rules is in the FX fixing scandal.

Stress test: March 2020

The BIS report includes a case study of the Covid-19 turmoil. Volatility “more than quadrupled in March 2020,” and “risk transfer spreads widened significantly, especially for larger-sized orders, as market-makers were less inclined to take risk onto their balance sheets.” Most providers reported algorithm volumes more than double their normal level. Data from the TCA firm BestX showed passive pegged and time-sliced algorithms “generally outperformed risk transfer during this period,” while simple TWAPs did worse than more adaptive algorithms, and some dynamic algorithms “needed manual retuning in March.” When dealers charge more to warehouse risk, working the order yourself becomes relatively more attractive, provided you can bear the risk.

Agency, principal, and what the Code requires providers to disclose

An algorithm provider can act as agent, sending child orders to the market for a fee; as principal, facing the client on every fill; or as riskless principal, principal on the trade but without market risk. The distinction matters because a principal provider may fill child orders from its own inventory (internalisation), and the client needs to know whose liquidity it is interacting with.

Principle 18 of the FX Global Code says providers of algorithmic trading or aggregation services “should provide adequate disclosure regarding how they operate.” The disclosures include whether the provider “could execute as Principal,” “the fees applicable to the provision of the services,” “general information regarding how routing preferences may be determined,” and “any conflicts of interest that could impact the handling of any Client order (for example, arising from their interaction with their own principal liquidity, or particular commercial interests in trading venues or other relevant service providers).” Providers are “encouraged to disclose pertinent information to be used for the purpose of Transaction Cost Analysis (TCA) in a market-wide standardised format,” such as the GFXC template.

Venues matter too. The TCA template records whether each child order went to a firm or last look venue; last look lets the liquidity provider reject a trade after seeing it, and the BIS notes that a “liquidity mirage is likely to be exacerbated if the practice of last look is permitted.” The practice is explained in last look explained.

What can go wrong

Information leakage. A predictable schedule, such as a TWAP that trades every 30 seconds into the fix, can be detected and traded ahead of.
Phantom liquidity. Fragmented venues can show the same liquidity several times; it disappears when an algorithm tries to take it all.
Non-completion. Passive and volume-linked algorithms may not finish, leaving the client with an open position at the worst time.
Hidden charges. Fees can be explicit or embedded in fills. TCA that excludes fees, or compares against a convenient benchmark, flatters the provider.
Benchmark mismatch. Measuring a fix-benchmarked fund against arrival price, or a cash trade against the fix, produces misleading conclusions.
Model failure in stress. Impact models calibrated in calm markets misjudge thin ones, as the March 2020 retuning showed.

Why this matters to retail traders

Retail accounts do not use execution algorithms, but the same arithmetic applies on a smaller scale. A signal follower who enters after the provider is effectively executing late against the provider’s arrival price; the difference is implementation shortfall, discussed in slippage and execution on copied trades. And the spreads a retail broker charges are many times the fractions of a basis point quoted in this article, as transaction costs in retail FX shows.

How professionals execute

Decide the benchmark first: arrival price, risk transfer price, TWAP over the window, or the fix, based on how performance is actually measured.
Estimate before trading. Pre-trade models estimate expected cost and risk for each method, as in the table above.
Match urgency to information. If the order is driven by news the market will soon price, speed matters more than impact.
Use several providers and compare them with standardised TCA over many orders, not one.
Read the disclosures: principal or agent, internalisation, venue selection, last look, and fees.
Watch the fix. If you must trade at the fix, understand that the five-minute window reduces manipulation risk but adds tracking error.

The short version

Large currency orders can be executed by risk transfer, by an algorithm, or at a benchmark fix. A risk transfer costs more on average but gives a known price; an algorithm such as a TWAP costs less on average but leaves the client exposed to price moves while it works, and in our stylised $575 million example that risk was several times larger than the cost saved. The BIS estimated in 2020 that algorithms handle 10-20% of spot FX, and found they performed well in March 2020 when dealers’ risk transfer spreads widened. The 4pm WM/Reuters fix moved to a five-minute window in February 2015, which made it harder to manipulate but harder to replicate. The FX Global Code asks algorithm providers to disclose their role, fees, routing and conflicts, and the GFXC’s template standardises the data needed for transaction cost analysis. This is an explanation of institutional practice, not execution advice.