Execution algorithms: VWAP, TWAP, POV and smart order routing

When people say most trading is algorithmic, this is mostly what they mean. Execution algorithms decide how an order reaches a market after the trading decision has already been made — how to slice it, when to release each piece, where to send it. They do not predict anything. That is precisely why they work: the objective is measurable, the counterfactual is observable, and success means a few basis points of cost rather than a forecast that came true.

Published 27 August 2026 · Updated 28 August 2026 · AI Trading Book Editorial · Reading time about 12 minutes

Two different questions

Every trade involves two decisions that are often collapsed into one. What should I hold, and how do I get there.

The first is the prediction problem, and it is the one that gets attention. It is also the one where the evidence is weakest: careful directional accuracy sits a few points above a well-chosen baseline, and AI hedge fund indices have underperformed passive benchmarks, as covered on our evidence page.

The second is the execution problem. Given that you have decided to buy a hundred thousand shares, how do you acquire them without pushing the price up against yourself before you finish? This has a measurable objective, a benchmark you can compare against, and a clear success criterion. It is unglamorous and it is where automation demonstrably earns its keep.

The confusion between the two matters for reading any statistic about algorithmic trading. ASIC estimates algorithmic trading at approximately 85% of Australian listed equities turnover and about 94% of SPI 200 futures; Select USA reports 60 to 75% of turnover in developed equity markets as algorithmic. Those figures are frequently quoted as evidence that machines are predicting markets. They are mostly measuring machines executing decisions that humans and portfolio models made.

The trade-off that defines the field

Execution has one central tension, and every algorithm is a position on it.

Trade the whole order immediately and you consume the available liquidity at progressively worse prices, moving the market against yourself — that is market impact, and it is a certain cost you inflict on yourself. Trade slowly to minimise impact and you are exposed for longer to the price moving for reasons unrelated to your order — that is timing risk, and it is an uncertain cost the market may or may not impose.

No setting avoids both. Faster execution converts uncertain timing risk into certain impact cost; slower execution does the reverse. The right point on the curve depends on order size relative to liquidity, on how urgently the position is wanted, and on how volatile the instrument is.

Scheduled algorithms: VWAP and TWAP

The simplest approach is to decide the schedule in advance and follow it.

TWAP

Time-weighted average price slices an order into equal parts released at regular intervals across a chosen window. Buying 120,000 shares over six hours means roughly 20,000 an hour regardless of what the market is doing.

Its virtues are simplicity and predictability of completion. Its weakness is the same predictability: a regular pattern is detectable by other participants, and it ignores liquidity entirely, so it may push into a quiet period and pull back during a busy one.

VWAP

Volume-weighted average price paces the order by expected volume instead of by the clock, trading more when the instrument is historically busy. Since most equity markets have a pronounced intraday pattern — heavier volume at the open and close, lighter in the middle — this typically causes less impact than TWAP for the same total size.

VWAP is also a benchmark, which is where much of its popularity comes from. An institution can instruct a broker to achieve VWAP and afterwards measure objectively whether they did. That measurability is a genuine advantage and a subtle problem: an algorithm optimised to match a benchmark is optimising for the benchmark, not necessarily for the client's total cost.

Its dependence on a volume forecast is the main failure mode. On a day when volume arrives in an unusual pattern — an unexpected announcement, an index event — the schedule is wrong, and the algorithm trades heavily at the wrong times.

Participation: percentage of volume

A POV algorithm targets a proportion of actual traded volume rather than a forecast. At 10% participation it aims to be roughly one share in every ten that trades, so it speeds up when the market is active and slows when it is quiet.

The advantage over VWAP is that it reacts to what is happening rather than what was predicted. The cost is that completion time is uncertain: on a quiet day the order may not finish, which converts an execution problem into a position problem. Participation rate is the control — higher finishes sooner and costs more impact.

There is also a subtle self-reference. Your own trading is part of the volume you are measuring, so a high participation rate partly chases its own tail. Implementations handle this, but it is a reason participation rates above roughly a fifth of volume are unusual for large orders.

Implementation shortfall

Implementation shortfall is the honest measure of what execution cost you. It compares the price at the moment the decision was made — the arrival price — with what was actually achieved, and it includes the cost of any quantity that never got filled.

That last component is what distinguishes it from VWAP as a benchmark. An algorithm can beat VWAP comfortably while performing badly on implementation shortfall, by waiting for good prices and failing to complete. The unfilled portion is a real cost — you did not get the position you decided you wanted — and VWAP does not see it.

As an algorithm family, implementation shortfall strategies solve the trade-off explicitly rather than following a fixed schedule, taking a view on urgency, volatility and expected impact and adjusting as conditions change. They are the most demanding to build and the most defensible conceptually, because they optimise the thing you actually care about.

Smart order routing

Scheduling decides when a slice is sent. Routing decides where. In a fragmented market the same instrument trades on several venues with different fees, liquidity and queue dynamics, and a smart order router chooses among them.

The decision involves more than the best displayed price. A venue showing a good price with little size behind it may be worse than one a tick away with depth. Fee structures differ, sometimes inverting the ranking after costs. And sending an order to one venue reveals information that affects the price at others, so sequencing matters.

Routing and scheduling are usually bundled in a single product, which is why the distinction gets lost. They are separate decisions with separate failure modes, and they are also separately regulated: providers of direct electronic access fall within the UK algorithmic trading regime independently of whether they run scheduling logic.

Order types that do execution work

Some of the same job is done by order types rather than by algorithms, and the boundary is blurry.

• Iceberg orders display only part of their size and replenish as the visible portion fills, reducing information leakage. The cost is queue position, since the refreshed portion usually goes to the back.
• Pegged orders track a reference such as the bid, offer or midpoint automatically instead of sitting at a fixed price.
• Hidden orders do not display at all, trading impact reduction for a much lower fill probability.
• Minimum quantity conditions prevent a large order being picked apart by many tiny fills, each of which reveals its presence.
• Immediate-or-cancel takes whatever is available at a price and cancels the rest, leaving no resting order to be detected.

The theme running through all of these is information. A large order is valuable information about future price pressure, and most execution technique is about revealing as little of it as possible while still getting done.

Where AI actually enters

Execution is one of the areas where machine learning has a plausible and specific role, precisely because the objective is well defined and the feedback is fast.

• Volume forecasting. VWAP needs a prediction of the intraday volume curve, and that is a forecasting problem with abundant data and a clear error measure.
• Impact modelling. Estimating how much a given order size will move a given instrument is empirical and improves with data.
• Adaptive scheduling. Reinforcement learning is an active research area here, since execution is naturally framed as sequential decisions with a measurable reward.
• Venue selection. Learning which venues fill reliably under which conditions is a pattern-recognition task.

Note the difference from price prediction. Each of these has a defined target, a large sample, and an outcome observable within hours. A price forecast has a noisy target, a small effective sample and feedback that takes months to distinguish from luck. That asymmetry is why the evidence for AI in execution is better than the evidence for AI in prediction, and it is consistent with what regulator surveys find: in the Bank of England and FCA survey of 118 firms, the leading use cases were process optimisation, cybersecurity and fraud detection rather than return generation.

How the rules apply

An execution algorithm determining order parameters with limited or no human intervention meets the definition of algorithmic trading. For an authorised firm in the UK that brings MAR 7A and onshored RTS 6 into play: conformance testing under article 6, annual self-assessment under article 9, stress testing under article 10 and kill functionality under article 12. ASIC's CP 386 proposes an equivalent framework for Australia, including a requirement for controls enabling immediate suspension.

This is the context for the FCA's multi-firm review of algorithmic trading controls published on 21 August 2025. Across the ten firms examined, pre-trade controls were adequate, but some firms using third-party algorithms did not have a good technical understanding of how those algorithms were built. Execution algorithms are overwhelmingly bought rather than built, which makes that finding specifically about this category — a firm routing flow through a vendor's algorithm carries the same responsibility as one that wrote its own.

Does any of this matter to a retail trader?

Mostly not, and understanding why is more useful than the algorithms themselves.

Execution technique exists to solve market impact, and market impact is a function of order size relative to available liquidity. A retail order in a liquid instrument is small enough that it consumes a negligible fraction of the book, so there is nothing to minimise. A limit order does the job, and slicing it would add complexity without reducing cost.

Three situations change that. Illiquid instruments — small caps, wide-spread pairs, thin crypto markets — where even a modest order is large relative to the book. Concentrated moments, such as trading into an announcement when the book thins out. And accounts large enough that position size becomes meaningful in the instruments traded.

The more transferable lesson is about costs generally. Spread and slippage scale with turnover and are frequently larger than the edge a strategy claims, which is the arithmetic set out on our accuracy page. Execution algorithms exist because institutions concluded that managing those costs was worth building infrastructure for. A retail strategy that ignores them is ignoring the thing institutions spend most on.

What we could not establish

• The split between execution and signal-generating algorithms within published algorithmic-trading shares. ASIC and Select USA publish aggregate figures without breaking them down by purpose. Industry understanding is that execution dominates, but no source quantifies it. Data not found.
• Comparative cost performance between algorithm types. Transaction cost analysis is largely proprietary. Data not found.
• Any regulator assessment of execution algorithm effectiveness. Supervisory work addresses controls and governance, not performance. Data not found.

Key takeaways

• Most algorithmic trading is execution, and statistics about algorithmic share are usually measuring this rather than prediction.
• Speed against impact is the only trade-off, and every algorithm is a position on it.
• VWAP is a benchmark as well as an algorithm, which is both why it is popular and why beating it is not the same as executing well.
• Implementation shortfall counts what you did not get, which is the cost other benchmarks miss.
• AI has a real role here — defined objective, fast feedback, large samples — unlike in price prediction.
• Retail rarely needs execution algorithms but should take the cost lesson seriously.

Frequently asked questions

What is an execution algorithm?

Software that decides how to work an order into the market once the trading decision has already been made — how to slice it, when to send each piece and where to send it. It does not decide what to buy or sell, only how to get the order done at the best achievable cost.

How is execution different from prediction?

Prediction asks what will happen to the price. Execution asks how to acquire or dispose of a position without moving the price against yourself. The second has a measurable objective and a clear counterfactual, which is why it is the part of algorithmic trading with the most solid evidence behind it.

What is VWAP?

Volume-weighted average price, used both as a benchmark and as an algorithm. The algorithm splits an order across a period in proportion to expected volume, trading more when the market is historically busy, aiming to achieve a price close to the period's VWAP.

What is TWAP?

Time-weighted average price — an algorithm slicing an order into equal parts released at regular intervals, ignoring volume entirely. Simpler and more predictable than VWAP, which is both its advantage and its weakness, since predictability can be detected.

What is the difference between VWAP and TWAP?

VWAP paces the order by expected volume, so it trades more during busy periods. TWAP paces by the clock, trading the same amount each interval. VWAP usually causes less impact in instruments with a pronounced intraday volume pattern; TWAP is more suitable when the volume pattern is unreliable.

What is a percentage of volume algorithm?

An algorithm that participates at a set proportion of actual traded volume — at 10% participation it aims to be roughly one in every ten shares traded. Unlike VWAP it reacts to real volume rather than a forecast, so completion time is uncertain.

What is implementation shortfall?

Both a measure and a family of algorithms. As a measure it is the gap between the price when the decision was made and the outcome actually achieved, including unfilled quantity. As an algorithm it explicitly balances market impact against the risk of the price moving while the order waits.

What is smart order routing?

Deciding which venue or venues an order goes to, and in what sequence, across a fragmented market where the same instrument trades in several places. It is a separate decision from how the order is sliced over time, though the two are usually bundled.

What is an iceberg order?

An order that displays only a portion of its size, replenishing the visible part as it fills. It reduces the information a large order reveals, at the cost of losing queue priority each time the hidden portion refreshes.

Do execution algorithms use AI?

Some do, particularly for volume forecasting and adaptive scheduling, and reinforcement learning is an active research area here. But the majority of execution volume runs on well-understood scheduling and routing logic that predates machine learning.

Are execution algorithms regulated?

Yes, for firms. An execution algorithm that determines order parameters with limited or no human intervention falls within the definition of algorithmic trading, bringing MAR 7A and onshored RTS 6 into play for authorised firms in the UK.

Do retail traders use execution algorithms?

Rarely and usually without needing to. Retail order sizes are generally small relative to available liquidity, so market impact is negligible and a simple limit order does the job. Execution algorithms matter when order size is large relative to what the market can absorb.

How much trading is execution rather than prediction?

No published split exists, but the share of algorithmic trading is high — ASIC estimates approximately 85% of Australian listed equities turnover, and Select USA reports 60 to 75% across developed markets. Industry understanding is that execution accounts for the large majority of that, though we found no source quantifying it.

What is the main trade-off in execution?

Speed against impact. Trading quickly moves the price against you; trading slowly exposes you to the price moving for unrelated reasons while you wait. Every execution algorithm is a position on that trade-off, and there is no setting that avoids both.

Sources

• ASIC — "ASIC moves to modernise trading system rules", 27 August 2025 — algorithmic share of approximately 85% of Australian listed equities and about 94% of SPI 200 futures.
• Select USA — 60–75% of turnover in US equities, EU and major Asian markets as algorithmic; plateau around 70–80%.
• FCA Handbook, MAR 7A; UK-onshored MiFID II RTS 6 — articles 6, 9, 10 and 12; direct electronic access within scope.
• FCA — multi-firm review of algorithmic trading controls, 21 August 2025 — adequate pre-trade controls; limited technical understanding of third-party algorithms among some users.
• ASIC — CP 386, 27 August 2025 — proposed controls enabling immediate suspension of a Trading Algorithm.
• Bank of England and FCA — "Artificial intelligence in UK financial services – 2024", 21 November 2024 — leading use cases process optimisation 41%, cybersecurity 37%, fraud detection 33%.

Informational research only. Nothing on this page is financial, legal, tax or investment advice, or a recommendation to use any execution method. Diagrams are schematic illustrations of logic, not plots of market data.