From measurement to prediction: How FX execution analytics is becoming a strategic trading tool

An important qualification to start this discussion is that adoption of FX execution analytics remains uneven. The most sophisticated users increasingly apply analytics to execution decisions in real time, while others still use it primarily to assess performance after a trade.

The strategic shift is from asking ‘How much did this trade cost us?’ to asking ‘How should we trade, with whom, through which venue, and at what time to achieve the best outcome?’ and the broader market structure reinforces the case for analytics.

For example, the BIS’s analysis of the 2025 triennial survey highlights the FX market’s fragmented execution landscape and the importance of dealer internalisation. In such a market, a trader needs to understand not just the price received but the quality of execution relative to realistic alternatives.

For many years, FX transaction cost analysis was primarily a post-trade performance measurement and best execution tool. The industry has progressed from analysing what happened towards influencing what traders do next further than most firms’ reports would suggest but not as far as the technology allows.

So says Paul Lambert, CEO NCFX, who observes that TCA was built to demonstrate best execution after the fact and for many firms that remains its main function. “Leading institutions, however, now use analytics to assess market conditions, plan trades, monitor them live and adjust LP panels and algo choices based on the results,” he says. “The shift is in the question being asked. Instead of only “How did we execute?”, firms increasingly ask, “Given the market we face now, what should we do?”

Since execution quality covers impact, information leakage, opportunity cost and risk – not just a slippage figure – TCA is best seen as one component of a broader framework that includes pre-trade to assess conditions and choose a strategy, in-flight to track the order against the live market and post-trade to measure the outcome and feed the lessons into the next decision.

Consistency is key

What makes that framework work is consistency, suggests Lambert. “A firm that plans a trade using one data source, monitors it with a second and measures it with a third cannot learn anything reliable, because the numbers never reconcile. A lifecycle approach needs a single independent reference running through every stage, from pricing and execution to valuation, risk and reporting.”

The quality, granularity and independence of the benchmarks and market data against which execution is measured are fundamental, because a measurement is only as good as its benchmark.

Forwards show the problem clearly, explains Lambert. “Most curves are quoted only at standard tenors and broken dates are interpolated between them. But forward points jump around month-ends, central bank meetings and holidays and interpolation errors at those points can exceed the spread being measured. Emerging market currencies and NDFs add thin liquidity and fixing complexities on top.”

Independence completes the picture. Measuring execution against a liquidity provider’s own prices lets that provider mark its own homework. Granularity, independence and traceability are prerequisites, not refinements.

According to Lambert, analytics can estimate execution costs, available liquidity and likely market impact before an order is sent to market accurately enough to change decisions, provided the estimates are built on reliable, consistent reference data and read as probabilities rather than promises. 

“Live monitoring is already practical,” he adds. “With a streaming independent benchmark, each fill can be judged as it happens and deviations flagged while there is still time to act.”

A benchmark is only as useful as it is relevant. If it is built with data from firms with a different credit profile, market access or trading style, it is a map of the wrong city: it can tell a client they were badly treated when they executed in the best way available to them.

“You also need to know how much data sits behind a benchmark and how it is built,” explains John Stead, VP strategy and marketing at smartTrade Technologies. “Granularity matters as much as volume since per-LP, per-tenor and properly timestamped data tells you far more than a daily average. Independence matters too. If a benchmark is built by the LP or venue you re measuring, you will always wonder whose interests it serves.”

He suggests that how much weight you give a benchmark depends on the volume of data behind it. A EUR/USD spot benchmark draws on vast amounts of data whereas EUR/CZK spot has a fraction of it. Move into swaps, NDFs and emerging markets and the data gets thinner still. In those markets, he recommends treating a benchmark as a range – not a single number – and judging execution against firms that trade like you.

“Analytics can estimate execution costs, available liquidity and likely market impact before an order is sent to market well where the data is deep,” says Stead. “We have had good results estimating market impact before an order goes out but everything rests on sample size – and that can’t be said often enough.”

Cut the data by tenor, time of day and LP and the variables multiply quickly. “Once the data gets thin, there is a trade-off. You either accept a smaller sample, which is less reliable for drawing conclusions, or you group: clients by cluster, time of day and tenors into buckets, currencies that behave alike. Working out which approach gives the most useful answer takes skill and domain expertise.Available liquidity and execution cost are the same story: estimates are good where the data is deep, and rougher where it is thin.”

The FX market is impressively competitive. If a trader or firm executes one trade a month, then execution costs may not dominate their P&L, whereas high frequency traders live and die by execution quality and liquidity.

Human judgement required 

Many firms are medium frequency and use ongoing stats to route orders to the right LP considering market conditions. That feedback loop works well for liquid pairs but executing a large trade in USD/ZAR still requires excellent human judgement.

That is the view of Walter Bell, chief technology officer at Ideal Prediction, who observes that the term TCA carries different meanings to various players. “In our view, execution analysis and managing liquidity are wonderfully complex,” he says. “The analysis doesn’t stop after calculating the market impact of a trade – you can understand the behaviour of your LPs and significantly improve your P&L. The leap is to go from just analysing filled trades to investigating the streaming quotes from LPs and ECNs, which is a much larger and richer data set.”

Bell reckons that terms such as analytics, decision intelligence, data and AI are often too generic to differentiate companies’ value propositions and that what is more specific is understanding how a firm will help your team.

“From our perspective, there is little differentiation pre/in-flight/post trade – it is about the end goal,” he says. “Analytics firms need to deliver the insight that will evolve execution to make more money, lower risk, adhere to the FX Global Code or help a client.”

Data quality is obviously table stakes for providing any analysis: times must be accurate and prices should be unskewed. Independence is always helpful as a neutral benchmark. 

“That said, some pricing may be client-specific, so we often use both independent and private data for our work,” says Bell. “Sterling spot is straightforward and bespoke swaps can be hard. Some data (for example, in illiquid swaps and EM pairs) isn’t conclusive by its nature but it is worth measuring regardless and then deeply researching the anomalies.”

Like all models, execution analytics are great at predicting based on data they have already seen. For example, if you are trading G10 pairs in calm markets, the estimates will be excellent. As you go down the liquidity curve or encounter disruptive markets, those estimates will naturally have more variance.

“We have seen many interesting phenomena pre-trade across asset classes, especially when optimising high frequency execution,” explains Bell. “For example, on the maker side, the pre-trade move can indicate if your trade is part of a large client order that swept the market. On the taker side, an LP may be low latency and first to react to price changes.”

“Over my career, I have seen many teams get very excited to use analytics to quickly change knobs. Those parameter changes might improve execution or they might add noise. The firms we have seen sustainably succeed are methodical and make data-driven decisions.”

Improving execution quality

For algo execution, banks can use real-time, short-term predictions and venue/LP selection to improve outcomes. The feedback loop measuring the quality of child trades can likewise improve the parent algo order execution quality.

In forwards, swaps and NDFs, much pricing data is indicative, stale or imprinted with trading activity, explains Will Holt, head of FairXchange, United Fintech.

“We have found that creating a quality reference price is not an insurmountable problem,” he says. “What really matters is sourcing genuinely independent pricing sources. Each one may be imperfect but using the same techniques deployed in liquid markets does allow robust references to be constructed.”

As for how accurately analytics can estimate execution costs, available liquidity and likely market impact before an order is sent to market, Holt observes that users of pre-trade analytics need to be certain of their purpose.

“They need to be clear on whether they are trying to exactly estimate the execution cost or if the analytics is being used to decide on a particular course of action,” he says. “When the aims are clear, pre-trade analytics can turn ad hoc decisions into a repeatable and explainable process.”

There is considerable interest across the industry in the extent to which execution analytics could operate while an order is still live and ultimately recommend the most appropriate execution route for each trade based on predicted rather than historical performance.

Holt notes that monitoring in-flight orders in a systematic manner can help remove the impulse to override algorithm parameters. “When the market is moving away from arrival the temptation is to override,” he adds. “In-flight statistics and regime metrics provide indications that the market is no longer in the same regime as order creation and the parameters can be adjusted accordingly. We have found that in-flight analytics more often than not discourages a trader from overriding their initial order.”

On the issue of the capacity for enhanced analytics to change the relationship between buy-side firms and liquidity providers, for example by replacing simplistic LP scorecards with a more constructive understanding of execution quality, Holt says the technology has already made considerable progress.

“Execution analytics have progressed a long way from basic scorecards that measure simple metrics such as spreads, fill ratios and reject rates,” he adds. “Better shared analytics turn the conversation to the specifics of timing, sizing duration and microstructural effects in liquidity pools.”

Bell refers to AI tools as useful applications that are used in many aspects of his firm’s software. When investigating execution quality, the tools can quickly scan many independent variables to identify the root drivers, such as LP behaviour changes, market volatility or scheduled economic releases.

“Another major consideration is the privacy of these tools,” he says. “Are you comfortable uploading your trades to tech firms? In our view, best practice is to use privately hosted LLMs and thoughtfully control your sensitive data.”

On the question of the importance of integrating analytics directly into the EMS, algo wheel or trading workflow, Bell observes that models can make mistakes. “Having run automated trading groups for 20 years, I have seen amazingly profitable models break overnight as market conditions change,” he adds. “My general advice is to use models on data you have previously seen and be mindful of the results when markets evolve. Integrating real-time, short-term price predictions and evaluating child order execution quality can improve algo performance. It is important to incorporate this analysis, with ongoing evaluation of the underlying models’ performance.”

Liquidity provider value

Bell says liquidity providers deliver value on a number of fronts, including consistency, size and pricing during disruptions that this complexity is not easily summarised by simple stats. For example, a multi-billion dollar asset manager has different goals than a systematic high frequency trading firm, which creates a challenge for analytics to optimise liquidity to attain those different goals.

Looking ahead, he notes that execution can work in a closed loop, especially when market conditions are consistent with the data used to train the models. “In my experience, market disruptions throw all the rules out the window.  As much as I am a technology enthusiast, I recommend human oversight. Governance, along with supervision and compliance, mitigates model risk.”

When asked what role machine learning and AI can play in identifying relationships between order characteristics, market conditions and historical execution outcomes that conventional TCA might miss, Stead notes that his firm has been using machine learning for more than a decade to cluster clients and LPs that behave alike.

“What generative AI adds is explanation,” he says. “In the smartTrade Agentic Copilot (STAC), machine learning does the computation and generative AI does the explanation. Each morning a desk sees which clients’ flow is costing it (the loss, the reasoning and the charts) and can question it in plain language rather than wading through pages of statistics. The AI finds the signal and the trader decides what to do with it, including setting triggers for when a client changes cluster or an LP changes behaviour.”

He agrees that it is vital to integrate analytics directly into the EMS, algo wheel or trading workflow, noting that every time a person has to read a report and act on it, you add latency and operational risk.”

“Our rule of thumb is to automate everything by default. The exception has to earn its place: a process that is rare and very expensive to automate. The same goes for an algo wheel in that analytics should set the weights as results come in, not wait for a quarterly review.”

Stead adds that enhanced analytics could change the relationship between buy-side firms and liquidity providers if both sides agree on one set of data and one way of reading it. 

“Most LP scorecards today are simple rules (a hit ratio below a set level means fewer orders), reviewed monthly or quarterly. Agree what matters to each side and both can see where they stand in real time.”

As for whether FX execution could evolve towards a closed feedback loop in which every trade improves the models used to plan and execute the next one, Stead suggests this is where the feedback loop was always heading.

“Our models learn with every trade,” he says. “Each one confirms or adjusts a hypothesis and the more data we have, the more confident the predictions. STAC is what closes the loop between one trade and the next. Every trade brings new data and STAC’s data science agent checks it against its models. Either it confirms the analysis or it’s a sign that something needs investigating to keep execution at its best.”

He adds that although AI will do more and more of the heavy lifting, firms will still need experienced people with domain, technical and analytical skills to review it, interpret it and check its progress. In this scenario, sensitive changes still go to an authorised approver and every decision is recorded for audit.

Predictive routing progress

Predictive routing is the natural next step in the progression from historical analysis to pre-trade planning, real-time decision support and increasingly automated execution. Today’s algo wheels mostly allocate orders based on past averages – a predictive approach asks which route suits this particular order in current conditions. Making it work requires consistent measurement across every route and continued experimentation, because a system that only uses its favourite routes stops learning.

Lambert believes machine learning and AI can play a significant role in identifying relationships between order characteristics, market conditions and historical execution outcomes that conventional TCA might miss and describes the importance of integrating analytics directly into the EMS, algo wheel or trading workflow as critical.

“A standalone report separates analysis from action: the trader has to leave the trading environment, interpret the findings and translate them into a decision and much of the value is lost along the way,” he says. “Insight that arrives after the decision has been made is a report, not decision support. Analytics have to reach the trader (or the algorithm) at the moment of choice, inside the EMS, OMS or algo wheel where trading actually takes place.”

As execution becomes more automated, analytics also need to be readable by machines as well as people. Technology should adapt to the client’s workflow, not the reverse, but integration must never compromise independence: the benchmark has to remain separate from the venue or provider it is being used to assess.

Lambert notes that single-metric scorecards reduce a complex relationship to a handful of historical averages and can reward the wrong behaviour. Contextual analytics show where each LP performs well or poorly under different conditions and change the conversation.

“The result is a relationship built on evidence, in which liquidity can be tailored rather than simply ranked. Banks gain as well, because an independent view of how their pricing compares with the market is useful to them too.”

He agrees that human judgement does not disappear but rather its role shifts from execution to design and oversight – setting objectives, handling the exceptional trade, challenging models when they drift and remaining accountable for results.

FX execution analytics is becoming a source of trading intelligence and competitive differentiation, rather than simply a way to demonstrate best execution or challenge bank pricing. But its strategic value depends on whether firms use the information to change trading behaviour, not merely produce more sophisticated reports.