Trading

What is transaction cost analysis (TCA), and how does AI improve it?

Transaction cost analysis (TCA) measures the full cost of executing trades, including spread, slippage, market impact and fees, before and after the trade. AI can improve it by analyzing more granular execution and market data, finding cost patterns sooner and linking analysis to the next execution decision.

Published October 27, 2026 | For heads of trading, technology and compliance

The TCA feedback loopStep 1: Pre-trade estimate (Likely cost and strategy). Step 2: Execution (Venues and algorithms). Step 3: Post-trade measurement (Against benchmarks). Step 4: Attribution (Spread, impact, delay). Step 5: Updated assumptions (Back-tested and approved). The last step feeds back into the first.Pre-tradeestimateLikely cost andstrategy1ExecutionVenues andalgorithms2Post-trademeasurementAgainst benchmarks3AttributionSpread, impact,delay4UpdatedassumptionsBack-tested andapproved5The TCA feedback loopStep 1: Pre-trade estimate (Likely cost and strategy). Step 2: Execution (Venues and algorithms). Step 3: Post-trade measurement (Against benchmarks). Step 4: Attribution (Spread, impact, delay). Step 5: Updated assumptions (Back-tested and approved). The last step feeds back into the first.Pre-trade estimateLikely cost and strategy1ExecutionVenues and algorithms2Post-trademeasurementAgainst benchmarks3AttributionSpread, impact, delay4Updated assumptionsBack-tested and approved5
Fig. 1 Post-trade findings update the pre-trade assumptions, so measurement feeds the next execution decision.

What does TCA actually measure?

Commissions are only the visible part of trading cost. TCA compares what a trade actually achieved against a benchmark, then attributes the difference to its causes. Measured well, it turns execution from a black box into a managed cost.

ComponentWhat it captures
SpreadThe cost of crossing from the bid to the offer
SlippageDifference between a benchmark, such as arrival price, and the average fill price
Market impactPrice movement caused by the order itself
Delay and opportunity costCost of waiting, or of portions that never executed
Fees and commissionsExplicit charges from brokers, venues and clearing

How do pre-trade and post-trade TCA differ?

Pre-trade analysis estimates likely cost for an order of a given size, liquidity and urgency, and helps choose an algorithm, venue or schedule. Post-trade analysis measures realized cost against benchmarks such as arrival price or volume-weighted average price, and explains the result.

The value comes from connecting the two. Post-trade findings should update the assumptions used pre-trade, closing the loop between measurement and execution.

A hypothetical example

Hypothetical, with made-up round numbers. A desk buys 100,000 shares when the arrival price is $50.00 and the average fill is $50.05. Slippage is $0.05 per share, or $5,000 on a $5,000,000 order, which is 10 basis points. TCA then asks why. Was the spread wide, did the order consume too much of the available volume, or was one venue consistently worse? Each answer points to a different action.

Where can AI add value?

  • Granular analysis. Fill-level data can be analyzed across many orders, venues and conditions, which periodic reports often cannot.
  • Pattern detection. Models can surface costly venues, time-of-day effects or algorithm behavior that averages hide.
  • Better cost estimates. Machine learning can improve pre-trade impact estimates, which are usually calibrated on historical data.
  • Unstructured context. News and market events can be tagged so unusual results are explained rather than ignored.
  • Explanations. Natural-language summaries can help reviewers understand outliers faster.

AI does not fix poor inputs. If order, fill and quote timestamps are misaligned, the analysis will be misleading however advanced the model.

What data and technology foundation does it need?

Reliable TCA depends on a data foundation as much as on models. Orders, fills, quotes and venue data usually arrive from order and execution management systems, brokers and market-data vendors. Streaming and event data can be captured as it arrives, while batch loads fill gaps and corrections. A lakehouse built on open table formats such as Delta Lake or Apache Iceberg lets analysts reproduce a past analysis from a point in time, and a semantic layer defines each benchmark and metric once so desks do not calculate arrival price in different ways. A feature store and a model registry keep the inputs and versions behind machine learning (ML) estimates traceable, and MLOps monitoring watches for drift. All of this can run on AWS, Microsoft Azure or Google Cloud, whichever matches your existing estate.

What should you check before investing?

  1. Confirm timestamp quality and clock synchronization across order, execution and market data.
  2. Agree a clear benchmark policy with traders so results are accepted.
  3. Check coverage across asset classes, venues and strategies, with consistent tagging.
  4. Ask how estimates are back-tested and when models are retrained.
  5. Make sure reviewers can see why an order was flagged.
  6. Confirm retention and access controls for trade data.

How does TCA relate to best execution?

Dealers and asset managers generally have obligations to seek good outcomes for clients, and TCA is a common way to evidence that oversight is happening. The specific requirements depend on your regulator, role and the instruments traded. Confirm current expectations with the relevant regulator, such as the Canadian Investment Regulatory Organization (CIRO) in Canada, the U.S. Securities and Exchange Commission (SEC) and the Financial Industry Regulatory Authority (FINRA) in the United States or the Financial Conduct Authority (FCA) in the United Kingdom, and with legal counsel. TCA supports a best execution process; it does not by itself satisfy one.

What are the common pitfalls?

  • Teams sometimes choose a benchmark that flatters results because it is easy to beat.
  • Orders are sometimes compared without adjusting for size, volatility, liquidity or urgency, which makes strategies look better or worse than they are.
  • The unexecuted part of an order can be ignored, which hides opportunity cost.
  • Conclusions can be drawn from samples that are too small to support them.
  • Reports can be produced that nobody acts on.

How should results be used?

Review them regularly with traders, not only with compliance. Group orders into comparable buckets, such as size relative to average daily volume, and compare brokers, algorithms and venues within each bucket. Feed findings into routing rules, broker reviews and pre-trade estimates, and record decisions so you can show how results led to action. Share a short summary with portfolio managers so trading cost is visible in investment decisions, since high-turnover strategies are especially sensitive to it.

What governance does AI-based TCA need?

Models that estimate impact or flag orders are models like any other. Keep an inventory, document the data and method, back-test against realized costs, monitor for drift when market conditions change, and keep a person responsible for acting on outputs. If estimates guide live execution decisions, manage changes to them under formal change control.

Where ARCHR fits

ARCHR can help trading and asset management teams assess their execution data, build the pipelines and analytics that make TCA reliable, and design AI components with validation, monitoring and human review built in. The approach is shaped around the client's existing systems and market-data sources.

This article is for general information only and is not legal, financial or investment advice.

FAQ

Frequently asked questions

Who uses transaction cost analysis?

Asset managers, trading desks, brokers and other institutions use TCA to evaluate execution quality, control costs and support best execution oversight.

Is TCA only a post-trade report?

No. Pre-trade TCA estimates likely cost to guide strategy, while post-trade TCA measures what happened and feeds back into the next decision.

What data does good TCA need?

Accurate order and fill timestamps, benchmark prices, venue and strategy tags, and fee data. Poor timestamps are the most common reason TCA results are unreliable.

Does AI replace traders or TCA vendors?

No. AI is one analytical technique. Traders keep responsibility for execution decisions, and many firms combine it with existing TCA tools.

Let's talk about your data and AI priorities

Tell us what you are trying to solve. We aim to reply within one business day.