In financial markets, the past arrives badly labelled. One exchange calls an auction one thing; another invents 30 variations. Protocols change. Clocks disagree. A cancellation appears without the context that made it meaningful. Then a highly paid quantitative researcher, hired to test an idea about liquidity, spends the week persuading yesterday’s files to open.
BMLL has built a business around that indignity. The London company collects the most granular historical feeds it can obtain - raw packet captures where possible, direct exchange data otherwise - and turns them into a consistent account of what happened inside order books. Not merely the best bid and offer, but orders arriving, changing, trading and disappearing. The result is sold as data, a managed research environment and precomputed analytics.
This is plumbing, but unusually consequential plumbing. A strategy tested on missing or misread events can look clever until it meets the market. A best-execution report built on inconsistent venue data can be impossible to defend. An AI model trained on noisy history becomes a very fast student of bad habits. BMLL’s proposition is that clean history should be shared infrastructure, not a private rite of passage.
The price of a dirty past
Ask what BMLL costs and the honest answer is: the company does not publish a rate card. Enterprise pricing depends on markets, asset classes, years of history, delivery and compute. Some samples and trials are available through BMLL and cloud marketplaces. The larger cost in its sales argument is the alternative - exchange licences, storage, parsers, quality checks, changing specifications and the specialists required to keep all of it alive.
BMLL says users can avoid spending as much as 80 percent of their time sourcing, cleaning and “munging” data. Treat that as a vendor’s claim, not a universal stopwatch. Yet the underlying complaint is familiar: historical market data is vast, fragmented and fussy. The first thing to fail is often not the trading hypothesis. It is the pipeline beneath it - a sequence gap, a timestamp mismatch, an undocumented venue state, a corporate action that does not line up.
“Firms now differentiate themselves by what they do with the data.”Paul Humphrey, chief executive
The firm itself required plenty of capital to industrialize the fix. Its public company history reports $36 million across seed and Series A financing, then a $26 million Series B in 2022 and a $21 million Optiver-led round in 2024. That is $83 million before Nordic Capital acquired BMLL in October 2025 and injected additional primary capital. The acquisition price was not disclosed. A neat figure would be satisfying; an invented one would rather defeat the spirit of the company.
Three depths, three different stories
Level 1 is the shop window: best bid, best offer and the size displayed there. Level 2 adds the shelves behind it, aggregating interest at each price. Level 3 is the stockroom ledger. Every individual order has an identifier and a lifecycle, with inserts, updates, cancellations and executions preserved at the precision supplied by the venue.
How much of the book can you see?
More depth does not guarantee a better decision. It does let researchers ask questions the shallower views cannot answer.
That last layer lets an analyst study queue position, resting time, fill probability, modifications and post-cancel behaviour. A broker can replay a smart-order router against actual market conditions. An exchange can compare its liquidity with rival venues. A regulator can inspect patterns that vanish in aggregated views. A researcher can join a firm’s own child orders to matching-engine events and ask whether an algorithm behaved better after a change.
BMLL reconstructs its Level 1 and Level 2 products from Level 3 source data rather than accepting a shallower historical feed as gospel. It offers both a lossless harmonized form, which retains venue-specific fields, and a tighter normalized schema for comparisons. The distinction matters. Standardization is useful only until it sands off the feature someone needs to investigate.
A refinery, not a terminal
The company’s production line has four broad jobs: collect, curate, engineer and analyze. BMLL says it keeps original source data indefinitely, an unglamorous choice that makes history reproducible. If its method improves, or an error appears, the company can regenerate the record rather than patching an already processed copy.
Clients enter through three main doors. Data Feed sends petabytes of history and derived analytics into a client’s own environment by API, files, object storage or cloud-data platforms. Data Lab puts the information beside JupyterLab, VS Code, Python libraries, scheduled jobs and scalable CPU or GPU compute. Vantage gives people who do not want to code a visual route into more than 500 Level 3-derived metrics.
For firms that already have the analysts and systems but do not want another ingestion factory.
For researchers who want the data, notebooks, scheduling and compute in one managed workplace.
For market-structure and execution questions that should not require a fresh Python project.
The delivery network is part of the product. Snowflake users can query native tables beside proprietary trade records. Databricks users gained another route in April 2026. AWS hosts core infrastructure and lists Data Lab in its marketplace. FactSet, Nasdaq Data Link, Deutsche Börse and other distributors take BMLL content into workflows clients already use. Several of these names are investor, partner and potential competitor at once. Financial data has always enjoyed complicated seating plans.
Born when the available answer was “build it”
Hugh Christensen started BMLL in 2014 after work in Cambridge University’s engineering and machine-learning laboratories with exchange groups. His recollection is admirably plain: he became an entrepreneur after realizing no existing solution solved the problem those businesses faced. The original name, Bayesian Machine Learning Limited, carried the academic ancestry in full.
The early answer was Data Lab, launched in 2017. Data Feed followed in 2019. Paul Humphrey, a market veteran whose career included Euronext, Tullett Prebon and Thomson Reuters, became chief executive that year. Vantage arrived in 2022. The sequence shows what changed BMLL’s mind about packaging: customers did not all want the same doorway. Some needed a laboratory, some wanted data pushed into existing systems, and others needed charts without code.
Geography expanded in the same practical fashion. Early venue coverage included LSEG, Borsa Italiana, CME, Euronext and Tradeweb. Japan and major US venues followed. By 2024 BMLL said it covered more than 100 venues and 98 percent of the MSCI All Country World Index. The current product site says more than 140 venues, spanning equities, ETFs, futures, options and regulated prediction markets.
The buyers are asking expensive questions
Banks and brokers use the data for transaction-cost analysis, algorithm tests, smart routing and best-execution evidence. Market makers examine queue dynamics and pricing. Asset managers and systematic funds back-test strategies and study market impact. Exchanges benchmark liquidity and member behaviour. Regulators and academics investigate market structure. BMLL’s named users and collaborators include Optiver, Berenberg, Kepler Cheuvreux, Aquis, SIX and École Polytechnique; its site also displays large institutions such as Bloomberg, UBS and Morgan Stanley.
Berenberg offers the most tangible public example. Its data-science and algorithmic-engineering teams layer client trade data onto BMLL market events by linking order IDs and matching-engine timestamps. The bank says that supports an A/B testing environment across five years of trade and market history. This is the sale in miniature: not a signal in a box, but the evidence needed to decide whether an execution change helped.
“BMLL has mastered the heavy lifting around the data normalisation process.”Marc Berthoud, SIX
The competitive difference is continuous maintenance. Anybody can buy files from one exchange. The harder job is doing it across regions, retaining sequence and condition codes, absorbing protocol revisions, managing vendor licences and applying the same quality method tomorrow. BMLL acts as Vendor of Record, so clients do not separately negotiate each covered exchange licence. It also places data and compute together, reducing the absurdity of dragging petabytes across a network before asking a question.
What Nordic Capital bought
Optiver’s 2024 investment was a useful endorsement because a market maker is unusually qualified to find bad order-book data. One year later, Nordic Capital acquired the company in partnership with management and minority shareholder Optiver. Management stayed in charge and retained shares. Nordic said the new capital would extend venue coverage, historical depth, asset classes, analytics and distribution.
The post-deal moves have followed that plan. In 2026, BMLL added Databricks delivery and SpiderRock US options print data. It launched an investor-flow analytics partnership with Exponential and a collaboration in which SIGMA AI uses BMLL history to benchmark real-time execution alerts. In July, former Rimes chief Brad Hunt became chairman while Lee Hodgkinson remained an independent director.
BMLL is moving from supplying research history toward becoming the base layer for other products - flow analytics, options studies and real-time decision support measured against the past.
That is also where the AI label becomes useful rather than decorative. BMLL does not make a model wise. It can make the training and benchmark record more consistent, more complete and easier to reproduce. The company’s strongest advantage is not a mysterious algorithm. It is accumulated engineering, licences, raw history and the staff who know how exchange minutiae can ruin a confident conclusion.
The copyable lesson is boring on purpose
Readers outside trading can borrow the operating idea. Keep raw inputs. Separate lossless harmonization from convenient normalization. Put compute beside large data. Version the analysis. Automate recurring jobs. Preserve identifiers that let customers join your record to theirs. Make one underlying asset usable through code, feeds and visual tools instead of building three unrelated products.
It will not fit every situation. A desk making decisions in microseconds still needs live feeds; BMLL’s core history is generally available T+1. A small team studying one venue and a short period may prefer direct files or a narrow specialist. A firm whose edge depends on proprietary capture may refuse to share the same foundation as rivals. And clean data cannot rescue a weak hypothesis, remove model risk or guarantee profitable execution.
What BMLL can do is remove a bad excuse. When a back-test fails, the researcher should be able to blame the idea before blaming a broken parser. In a market fond of speed, the company has made its case by tending patiently to the past.