Tradefeedr Brings AI-Powered Analytics to the Next Generation of TCA

Artificial intelligence is opening up new possibilities for transaction cost analysis (TCA), allowing traders and portfolio managers to interrogate execution data, explore trading performance and generate customised insights without relying on specialist quantitative or programming skills. But as AI becomes more deeply embedded in execution workflows, questions around analytical consistency, transparency and accountability are becoming increasingly important.

In this Q&A, Balraj Bassi, CEO and Co-Founder of Tradefeedr, discusses with TradingTech Insight the practical implications of AI for TCA, the importance of common analytical standards, and the potential for execution analytics to evolve beyond retrospective measurement towards continuous decision support.

TTI: Tradefeedr describes AI as a delivery mechanism for TCA rather than a replacement for the underlying analytical methodology. Where do you draw the boundary between what AI should be allowed to interpret or generate and what needs to remain deterministic, particularly when the output may feed into best-execution decisions?

BB: For us, the boundary is very clear. AI can help users interrogate data, interpret results, identify patterns and decide which analysis to run, but it should not invent the underlying numbers or redefine the methodology.

The calculation of metrics such as slippage, arrival cost, market impact or benchmark performance needs to remain deterministic. The same data, methodology and parameters should produce the same result every time, irrespective of which AI model or interface is being used.

AI then sits on top of that analytical layer. It can translate a natural-language question into a request for the appropriate data and analytics, combine different outputs and explain what those results may mean.

That distinction becomes particularly important when analysis feeds into best-execution processes. A user should always be able to move from an AI-generated observation back to the underlying calculation, data and methodology that produced it.

We see AI as making institutional analytics dramatically more accessible, not making the analytical process less rigorous.

TTI: Historically, getting genuinely customised insight from execution data has often required quant skills and the ability to work directly with APIs or tools such as Python. What changes when traders and portfolio managers can interrogate the same data conversationally, and what kinds of questions are users now asking that they previously struggled to answer?

BB: The biggest change is that the barrier between having the data and actually being able to interrogate it starts to disappear.

Historically, many trading desks have had enormous amounts of execution data available to them, but answering a new or highly specific question might still require a quant, an API query, Python code or a new report to be built. That naturally limits how frequently people experiment with the data.

Natural-language access changes that dynamic. A trader or portfolio manager can move from looking at a standard report to asking very specific follow-up questions immediately.

For example: How did our performance in EURUSD compare with peers when trading more than $100 million? Which execution strategies performed best during periods of higher volatility? Are we paying more for liquidity than we were six months ago? Has our use of faster or slower algorithms changed, and has that affected performance?

Those questions were always theoretically answerable. The difference is that they can now be asked by the person making the trading decision, without that person needing to know how the underlying database or API is structured.

That creates a much more iterative relationship with execution data. Instead of consuming a fixed set of reports, users can investigate why something happened and immediately follow the next question.

TTI: Natural-language access makes sophisticated analytics easier to use, but it also introduces the possibility that users ask ambiguous questions or that an AI assistant misinterprets their intent. How does Tradefeedr ensure that an apparently simple question produces an analytically valid, reproducible and auditable answer?

BB: This is exactly why we believe the AI layer has to sit on top of a controlled analytical framework.

A natural-language question may be ambiguous, but the analytical operations available underneath it should not be. The AI is effectively translating the user’s question into requests against defined datasets, parameters, benchmarks and calculations.

Where a question requires parameters that have not been specified, the system should either apply clearly defined defaults or ask the user for clarification rather than silently making analytical assumptions.

Equally importantly, the answer needs lineage. It should be possible to understand which data was queried, which analytical function was used, which filters were applied and which benchmark or methodology produced the result.

That means the conversational interface can be flexible while the underlying analysis remains reproducible.

We think that separation is essential for institutional use. The goal is not simply to produce an answer that sounds plausible. It is to produce an answer that can be verified.

TTI: You argue that customisation needs to sit on top of common, deterministic definitions for metrics such as slippage, market impact and arrival cost. Why is that standardisation becoming more important as AI enters execution analytics, and what problems emerge if different firms or AI systems are working from different definitions of the same metric?

BB: AI actually makes standardisation more important, not less.

As analytics become easier to generate, users will inevitably create more customised analysis. That is extremely powerful, but only if everyone is starting from a common analytical foundation.

If two institutions use the term “market impact” but calculate it differently, or two AI systems interpret “arrival cost” using different timestamps, benchmark definitions or treatment of fees, the outputs may look directly comparable when they are not.

AI can amplify that problem because it can generate analysis very quickly and present the result extremely convincingly.

The solution is to separate the definition of the metric from the way a user chooses to interrogate it. Clients should have enormous flexibility in asking questions, segmenting data and combining analytics, while the core definitions and methodologies remain consistent and transparent.

That is also one of the benefits of an industry-wide dataset and common analytical framework. It gives the market a shared language.

TTI: With the Tradefeedr MCP server, clients can use their own AI assistants rather than being tied to a Tradefeedr model or interface. Why did you choose that architecture, and what does it mean in practice for data governance, security, model choice and accountability when an AI-generated answer is challenged?

BB: We deliberately do not believe that clients should have to adopt a particular AI model simply because they use Tradefeedr analytics.

Different institutions already have different policies, technology stacks and approved models. Some are building their own internal assistants, others use enterprise versions of external models, and some firms currently place significant restrictions on the use of generative AI.

The MCP architecture allows Tradefeedr to provide the trusted data and analytical capability while the client retains control over the AI environment through which it is accessed.

That separation is important from both a technology and governance perspective. Tradefeedr remains responsible for the integrity of the data, permissions and deterministic analytical functions we provide. The client remains in control of its chosen model, internal AI policies and the way that model is deployed within its organisation.

It also reduces model dependency. The value should reside in the quality and structure of the underlying data and analytics, not in which large language model happens to be leading the market at a particular point in time.

And if an AI-generated answer is challenged, there should be a clear route back to the underlying Tradefeedr query and analytical result. The model should not become an opaque layer between the user and the evidence.

TTI: TCA has traditionally been retrospective, explaining what happened after an order was executed. To what extent can AI move execution analytics towards a more continuous or predictive process – identifying patterns, generating hypotheses or informing execution decisions before and during trading rather than simply analysing them afterwards?

BB: This is potentially one of the most interesting developments.

Traditional TCA has largely focused on explaining what happened after execution. That will remain important, particularly for best execution and governance, but the same analytical infrastructure can increasingly support decisions before and during trading.

Predictive execution models already use historical data to estimate the likely cost, duration and risk of an order under different execution strategies, given its size and prevailing market conditions. AI can make these capabilities more accessible to traders, helping them identify relevant historical patterns, explore alternatives, understand forecasts and assess the trade-offs for a particular order.

We would distinguish that from allowing AI to autonomously make execution decisions. There is a significant difference between providing better contextual intelligence to a trader and delegating responsibility for the trade itself.

The nearer-term opportunity is therefore to move from retrospective TCA towards continuous execution intelligence: helping users understand what has happened, what is changing and what that may imply for their next execution decision. Predictive models provide the forecasts; AI makes them easier to explore, interpret and use.

TTI: Tradefeedr’s Industry Advisory Board brings together buy-side trading desks, banks and technology providers. What have practitioners told you they actually want from AI in execution analytics, and where has that feedback caused you to change either the product, the security model or your assumptions about how much autonomy users are comfortable giving AI?

BB: One of the clearest messages from the market is that there is a significant gap between enthusiasm for AI and what institutions are currently prepared to allow AI to do.

Practitioners are very interested in being able to interrogate their data more easily, automate repetitive analysis, identify unusual patterns and generate customised insight. What they are generally much less interested in is handing an opaque AI system control over their analytics or execution process.

We have therefore focused heavily on using AI as an access layer rather than replacing the core product with an AI interface.

Our standard platform, reports, APIs and deterministic analytics remain fundamental. That matters because not every user within an institution will necessarily be permitted to use generative AI, and different institutions will progress at very different speeds.

The same feedback influenced our approach to MCP. Rather than telling institutions which model they need to use, we want them to be able to connect Tradefeedr’s analytical capabilities into the AI environments their own security and technology teams have approved.

The market is effectively telling us: give us the flexibility and productivity benefits of AI, but do not remove control, transparency or accountability.

We think that is a very sensible position.

TTI: You are extending the AI-ready analytics layer beyond FX into fixed income, equities and futures. How difficult is it to create meaningful cross-asset TCA when execution methods, liquidity structures and relevant benchmarks vary so widely, and what becomes possible for institutions once those different markets can be interrogated through a common analytical layer?

BB: The challenge is that cross-asset analytics cannot simply mean taking an FX TCA model and applying it to every other asset class.

The market structure, execution workflow and relevant measures of performance can be very different across FX, equities, futures and fixed income. Even concepts that sound similar may require different methodologies depending on the asset class.

Our approach is therefore to standardise the infrastructure and analytical principles without pretending that every market is identical.

Each asset class needs analytics that genuinely reflect how that market trades. At the same time, the underlying data architecture, terminology, governance and access layer can be made much more consistent.

That becomes particularly powerful for large institutions. Today, execution data often sits in separate systems across asset classes, with different reporting processes, different datasets and sometimes entirely different teams maintaining them.

Once those datasets can be accessed through a common analytical layer, users can begin asking questions across the organisation rather than within individual technology silos.

A global head of trading could potentially move from asking how algorithmic execution is changing in FX to how transaction costs are evolving in equities, futures or fixed income through the same interface.

Ultimately, we think the opportunity is bigger than AI-enabled TCA. It is about creating a common institutional data and analytics layer across markets, with AI providing a much simpler way for people to access and interrogate it.

Looking ahead, Tradefeedr sees opportunities to extend these capabilities across asset classes, bringing FX, equities, fixed income and futures analytics together through a more consistent data and analytical framework. While each market requires its own execution metrics and methodologies, a common access layer could allow trading desks to investigate performance across markets without navigating multiple disconnected systems.

For solution providers such as Tradefeedr, the longer-term opportunity extends beyond making existing TCA processes faster or easier to use. By combining standardised analytics with more flexible access to execution data, AI could help institutions move towards a more continuous approach to evaluating trading performance, informing decisions before and during execution as well as afterwards. The challenge will be to deliver those benefits while preserving the transparency, methodological rigour and human oversight that institutional trading requires.