Choosing wisely
With so many algo strategies available across providers, what should buy-side firms be thinking about when matching a strategy to a particular order?
The proliferation of algo strategies is a positive development — but more choice doesn’t automatically mean better outcomes. The question is whether firms have a structured way to navigate that choice.
It starts with the objective. Is this about minimising cost, managing risk, hitting a benchmark, or reducing footprint? Different objectives point to different strategies — TWAP for time risk distribution, POV for market camouflage, aggressive strategies for urgency. The best selection frameworks are simple decision trees, not complex models.
Size relative to daily volume is a critical input. A $200 million order in EUR/USD, where daily turnover runs into the hundreds of billions, is a very different execution challenge than the same notional in USD/TRY or USD/ZAR, where it represents a much larger share of available liquidity. Cross pairs add another layer of complexity — a EUR/JPY order is typically constructed from two legs, each with its own liquidity profile and spread dynamics, making the algo’s ability to manage both components efficiently a key differentiator. The liquidity profile of the pair matters just as much as its size.
Every order has a fingerprint
How should the characteristics of a specific trade — currency pair, size, urgency, time of day — shape the algo selection process in practice?
Every order has a fingerprint – and the more precisely the algo selection reflects that fingerprint, the better the outcome.
Currency pair determines liquidity depth, venue availability, spread behaviour, and time-of-day patterns. Size matters not just in absolute terms but relative to average daily volume — this determines how “loud” the order will be in the market. Urgency is worth interrogating: is it a genuine constraint or a preference? The distinction matters enormously for strategy choice.
Time of day shapes everything. London morning and Asia afternoon are effectively different markets with different liquidity profiles, spread dynamics, and participant behaviour. Expected holding period such as intraday versus multi-day affects which benchmark is appropriate and therefore which strategy makes sense. “One size fits all” selection remains the most common missed opportunity we see.
How important is a provider’s ability to internalise flow, and what should buy-side firms understand about how internalisation affects their execution quality?
Internalisation or matching a client’s order against the provider’s own internal flow rather than routing to external venues can be one of the most significant factors in execution quality, yet it’s often underappreciated in the selection process. When an algo can match flow internally, the benefits are tangible: minimal market footprint, reduced information leakage, and typically tighter pricing than what’s available in the lit market.
End-to-end internalisation matters because it directly addresses adverse selection — the risk that external counterparties are only willing to trade when conditions favour them rather than the client. When flow is matched internally within a provider’s own ecosystem, both sides of the trade are natural, which removes the information asymmetry that drives adverse selection in the broader market. The more of an order that can be filled this way, the less exposed the client is to the signalling costs and predatory dynamics that come with external venue interaction.
How internalisation is accessed relative to external liquidity is equally important. The right approach is to access internal liquidity only when it offers a genuinely better outcome for the client – tighter pricing, lower market impact, or reduced signalling — rather than defaulting to it for the provider’s convenience. At J.P. Morgan, internal liquidity is accessed when it is most advantageous to the client, competing on equal terms with external sources rather than being preferred automatically.
Feeding the machine
What pre-trade information has become most valuable for informing algo selection and parameterisation, and how well is the buy-side community using it today?
The quality of an algo’s execution often depends less on the algo itself and more on how well it’s parameterised. The inputs matter as much as the engine.
Historical volume profiles are among the most valuable — they reveal when liquidity typically appears for a given pair, and these patterns are reasonably consistent outside of event days. Spread behaviour follows similar rhythms: time-of-day, day-of-week, and event-driven widening all create predictable windows that should inform when and how aggressively to execute.
Volatility regime is another critical input. Whether the market is trending, ranging, or stressed calls for fundamentally different parameterisation — a set of parameters that works well in a calm market may produce poor results when volatility spikes.
Expected cost modelling brings these inputs together: given this order’s size, pair, and current conditions, what should execution cost? The answer provides a baseline against which outcomes can be measured.
The gap we observe is that many firms select the right algo but accept default parameters. Pre-trade analytics exists to turn raw information into a specific execution plan — but it’s underutilised.
Once an algo is running, what real-time signals help distinguish genuinely adaptive execution from a strategy that’s simply following a predetermined schedule?
“Adaptive” has become a popular label, but there are meaningful differences in how algos respond to changing conditions during execution. The signals worth watching are the ones that reveal whether the algo is genuinely reacting to the market or simply ticking through a schedule.
Spread movement is one of the clearest indicators. Does the algo recognise when spreads are favourable and lean in, or does it maintain rigid pacing regardless? Similarly, volume patterns matter. An algo that adjusts participation when market volume is above or below average is responding to real conditions rather than following a clock.
Momentum signals add another layer. A genuinely adaptive algo modulates its aggression based on whether the market is moving for or against the order, rather than treating every interval the same. Fill rates offer a real-time feedback loop. If the algo is learning from its own recent success or failure and adjusting, that’s a sign of genuine intelligence.
The buy-side can test this directly: run similar order profiles through different strategies and compare the behaviour, not just the outcomes. The patterns in how the algo trades often reveal more than the final number.
Peer comparison adds further context: how does this execution compare to similar orders under similar conditions?
How can firms build a feedback loop where historical execution data meaningfully improves future decisions, rather than just sitting in a report?
The most valuable data any firm has is its own execution history. The question is whether it flows back into the decision process or simply accumulates in reports that nobody reopens.
Clients that can leverage the depth of data available from their liquidity providers have a real opportunity here. The raw material exists – detailed fill data, venue-level performance, timing analysis, cost attribution – but it needs to be systematically mined rather than passively received.
At J.P. Morgan, our FX Alternatives framework takes this a step further. It allows us to backtest client orders at different speed configurations, leveraging historical data and adjusting for market impact to produce what-if scenarios. Rather than debating in the abstract whether a faster or slower execution would have been better, the framework provides a data-driven answer: here’s what the cost profile would have looked like under different approaches, given the conditions that actually prevailed.
Provider assessment follows naturally. Comparing performance across providers for similar orders, controlling for conditions, gives a more objective basis for allocation than reputation or relationship alone. And automation helps — systematic tagging, classification, and comparison can surface insights without manual effort. The firms that build this discipline create a compounding advantage over time.
Knowing what worked
Arrival price remains the most common benchmark. What are its limitations, and what additional dimensions give a fuller picture of execution quality?
Arrival price has earned its place as the default benchmark, but for a richer understanding of execution quality, it’s worth considering what it captures — and what it misses.
Arrival price measures from a single point in time, which may or may not represent the true decision point. For orders that take hours to execute, it conflates market drift with execution quality — the price may have moved significantly for reasons entirely unrelated to the execution itself, and disentangling the two is difficult.
Additional dimensions worth examining include implementation shortfall decomposition, which separates delay cost, market impact, and timing luck into distinct components. Spread capture — whether the algo earned or paid the spread — reveals how effectively it used passive strategies. Volume participation shows how visible the order was in the market. And signalling cost measures whether the market moved adversely during execution in a way that suggests information leakage.
Peer comparison adds further context: how does this execution compare to similar orders under similar conditions? No single metric tells the whole story. A dashboard view that brings these dimensions together gives more useful insight than any one number in isolation.
How should TCA evolve to become a forward-looking decision tool rather than a backward-looking compliance exercise?
TCA started as a compliance and reporting requirement. The more interesting question is how it becomes a decision-making tool that actively improves future execution.
Backward-looking TCA answers “how did we do?” — necessary but insufficient. Forward-looking TCA answers “what should we do differently?” — and that’s where the value compounds.
Segmentation is key. When you slice performance by pair, size bucket, time of day, strategy, and provider, patterns emerge that aggregate numbers hide. A firm might discover that its preferred strategy consistently underperforms for a specific pair during certain sessions, or that one provider delivers better outcomes for larger orders while another excels at smaller, more urgent flow.
This enables genuine hypothesis testing: “we believe a participation strategy works better than TWAP for EUR/USD orders over $100M during London morning” — does the data support it? If it does, that insight feeds back into selection frameworks and parameter defaults. If it doesn’t, the assumption gets revised.
The maturity curve runs from reporting to analysis to insight to action. Most firms are somewhere in the first two stages. The competitive advantage lies in reaching the last two.
Looking ahead, how do you see the relationship between data, algo intelligence, and performance measurement evolving over the next few years?
The intersection of data, intelligence, and measurement is where the most interesting developments in FX algo execution are likely to emerge.
More data will become available — deeper venue analytics, broader market microstructure signals, and richer internalisation metrics. Algo intelligence will continue to improve, with better real-time adaptation and more sophisticated cost models that learn from an ever-growing dataset of execution outcomes.
While fully AI-driven execution remains ahead of us, AI is already being deployed in the areas surrounding algo execution — integration with trading workflows, smarter order staging, and streaming pre-trade analytics that surface relevant context directly to execution desks. These tools help traders make more informed decisions about strategy selection and parameterisation without removing the human from the process. Over time, this kind of AI-assisted workflow is likely to become the norm rather than the exception.
Perhaps most significantly, measurement will become more integrated. Rather than TCA being a purely post-trade exercise, we’re moving toward real-time performance monitoring during execution -giving both the algo and the human overseeing it the ability to assess and adjust while the order is still live.
The convergence is the key theme: pre-trade, execution, and post-trade are becoming a continuous loop rather than separate phases. For buy-side firms, the ones that invest in understanding and using data across this entire loop will extract more value from their algo programmes than those who treat each stage in isolation.
In a market where everyone has access to similar tools, the edge increasingly lies in how well you use them.
