The questions have become perennial: build versus buy, cloud versus on-premise, in-house versus managed platform. But the discussion that emerged from a recent A-Team Group webinar on quantitative research infrastructure pointed to something more interesting than a binary choice. The build-versus-buy question, panellists agreed, has matured into a more sophisticated conversation about where firms locate their genuine proprietary edge – and what they can safely externalise without giving it away.
The webinar, Navigating the Build vs Buy Dilemma: Cloud Strategies for Accelerating Quantitative Research, sponsored by OneMarketData/KX and moderated by TradingTech Insight editor Mike O’Hara, brought together two buy-side practitioners – Renato Guerrieri, Head of Quantitative Strategy for Liquid Alternatives at Downing, and Chris Kelliher, Portfolio Manager in the multi-asset systematic strategies team at Fidelity Investments – alongside Peter Simpson and Mick Hittesdorf from OneMarketData/KX. A live audience poll midway through the session put the result starkly: most attending firms now operate a hybrid mix of self-built and managed components. Pure-play strategies at either end of the spectrum are increasingly the exception.
Where the edge sits
“I don’t see build versus buy as a religious, dogmatic debate,” said Guerrieri. “Build what is proprietary and buy what is a commodity. The mistake is building everything for control, or buying everything and accidentally outsourcing your brain.”
Kelliher drew a sharper line around data specifically. “If it’s data that’s central to the alpha generation process, then there absolutely is the potential for it to be something fundamental,” he said. “If you’re a portfolio manager or researcher trying to articulate an edge, getting the data externally creates a challenge to understand the edge and how that edge will persist.”
The corollary is that traditional, standardised data sets are increasingly recognised as commodity infrastructure rather than competitive advantage. As Guerrieri put it: “For most firms, maintaining another version of standard historical market data is not where the edge actually sits. Nobody is going to find much alpha just from historical prices.”Firms are no longer asking whether to outsource their data infrastructure, but which layer of the research stack to outsource, and what discipline to maintain around the boundary.
The commodity layer is harder than it looks
If historical market data is increasingly treated as commodity, that does not mean it is easy to provide well. The audience poll asking what firms find hardest about building and maintaining a market data environment in-house produced an unambiguous answer: ongoing maintenance and data operations burden, well ahead of upfront build cost.
“That’s the one that people notoriously overlook,” said Hittesdorf, who spent nearly three years building a cloud-based market data platform at an options market-making firm in Chicago before joining OneTick. “The thinking is, once I get this platform built and get my data, I’m done. But that’s really when the hard work starts.”
Simpson pointed to a constant stream of venue-driven change: “The National Stock Exchange of India at the end of last week pulled its delivery of level three data, and updated a completely new spec on Monday. That continuously happens. You have to always understand each of the venues you’re subscribing to, and understand that they will change.”
For Guerrieri, the deeper cost is the effect on research quality. “Firms often underestimate all the boring problems,” he said. “Identifiers, corporate actions, missing data, correction, entitlements, point-in-time logic, versioning, monitoring – all the small breaks that quietly damage your research quality. The biggest risk is false confidence. You can get a clean-looking backtest from a fragile data foundation, and that is very dangerous, especially when allocating capital. It does not look obviously wrong; it just makes you believe the wrong thing.”
This is the case for externalising the data layer to specialists – but it also exposes the limits of what managed platforms can realistically deliver. Asked how well they handle deep historical data quality across corporate actions, symbology changes and survivorship, Guerrieri was direct. “You are not going to find anyone who is good at everything,” he said, citing his own multi-asset experience. “If you have derivatives that are OTC or direct contracts with the banks, you cannot expect a vendor to cover all of the 10 major banks. But maybe they are great with managing corporate actions in small caps.”
The point cuts both ways. Building everything in-house produces operational burden and false confidence; outsourcing everything produces coverage gaps that need to be papered over with internal workarounds. The hybrid model is not a compromise but a recognition that no single approach handles the full surface area.
What this means for the researcher
The infrastructure choice shapes what researchers can practically investigate – and, more subtly, what they choose to investigate at all.
“The API is faster and definitely helps,” said Guerrieri. “But it also changes what questions the researchers are willing to ask. If data access is slow, messy, or painful, researchers will naturally avoid certain questions. They stay closer to what is easy to test. That is going to limit the research agenda.”
Kelliher emphasised consistency as a related benefit. “If two people ask the same research question and use the same data, they’re going to get the same answer. Having this streamlined data set can help you test things that wouldn’t be possible if things are coming in disjoint places.”
Both panellists were careful to caveat the productivity argument. “If the research process is weaker internally, faster infrastructure just helps you reach the wrong answer faster,” said Guerrieri. “The real objective is not just speed by itself – it is time to validate.”
That distinction is where the strongest case for managed infrastructure lives. Multi-year backtests, scenario grids and regime-change simulations all place uneven demands on compute, and the case for elastic, cloud-native infrastructure is strongest where workloads are genuinely bursty and parallelisable. It is weakest where firms move to the cloud without the cost controls to manage what they are running. “Cloud just converts technical inefficiency into a larger invoice,” as Guerrieri put it. “Without control, it’s just another expensive flexibility.”
Vendor dependency and architectural discipline
If hybrid is the dominant model, the question becomes how firms maintain optionality across the boundary. The panel’s response was notable for emphasising architecture as much as contractual protection.
“Convenience is valuable, definitely, but it should not become captivity,” said Guerrieri. “Firms need clarity on access, exit routes, continuity, service levels, and data usage. But they also need modular research code, open interfaces, lineage, and the ability to move or reproduce a critical workflow.”
From the vendor side, Hittesdorf advocated for interoperability as a deliberate architectural choice. “I’ll lean into interoperability, and only take advantage of proprietary features when they provide distinct and differentiating value. Things like open data formats – Parquet and Iceberg – are designed to allow you greater ownership of your data and the use of common tools and APIs.”
Simpson pointed to the importance of contractual foresight, and of treating proprietary symbology with caution. “If a vendor is providing you proprietary symbology, why would you go down that path when you’ve got standardised MIFID classifications and standard symbologies like FIGI, ISIN, SEDOL or CUSIP? Try to avoid proprietary lock-in. The price might be very attractive in year one, but by year three it’s going to change.”
The underlying point is that hybrid architecture is not a default state to be drifted into. It requires deliberate decisions about what is encapsulated, what is exposed, what is standardised and what is allowed to be proprietary.
AI is shifting the line
The subject of AI came up relatively late in the conversation, but it sharpened the central question. AI is not changing the build-versus-buy decision so much as moving the boundary between what counts as commodity and what counts as proprietary.
“AI is a force multiplier across the industry,” said Hittesdorf. “There’s a temptation to think that as an end user, now you can build it all yourself – just point Claude Code at it and let Claude build it. There’s a lot of domain knowledge, however, that is specialised with respect to market data that AIs just don’t have. At the same time, your vendors are taking advantage of that same capability and delivering more software and better data faster.”
Kelliher highlighted how AI is changing vendor interfaces themselves. “Instead of accessing the data in a raw way, maybe a vendor is able to use an LLM to have a prompt-based framework. That can be a powerful concept in terms of making the whole process more seamless.”
Guerrieri was sharper on what AI has and has not changed. “It is definitely helping on the workflows, but not on the model. We have not invented new models.” On agent-driven research, his framing was characteristically pointed: “Agents need rails. Without the rails, they are basically expensive interns with superpowers.”
That observation found an immediate echo on the vendor side. Simpson noted that managed platforms are already seeing the consequences at scale. “We’ve seen the amount of queries ramping up significantly, and that is ever-growing as more analysis is agentically driven. We don’t want a customer’s AI agents to run thousands of queries to build up tick-by-tick a set of metrics. We need to provide more feature sets that are relevant to our customers’ needs.”
The implication is significant. As agents become a larger share of how research is conducted, the value of managed platforms shifts from raw data delivery to pre-engineered feature sets and MCP tooling that agents can call efficiently. The commoditisation line is moving up the stack.
Decision-ready infrastructure
For all the talk of hybrid models and architectural discipline, the panel converged on a forward-looking thesis that is more demanding than the build-versus-buy framing usually implies.
“The biggest shift will be toward decision-ready research infrastructure,” said Guerrieri. “Firms don’t just need more data, more compute, or more tools. They need a cleaner workflow from data to research to portfolio decisions, with lineage, reproducibility, monitoring, and governance built in. The winning setup will be hybrid: managed infrastructure where scale, coverage, and maintenance are a commodity, and internal ownership where research logic, portfolio construction, and investment judgement are proprietary and paramount.”
Hittesdorf described a similar dynamic as a continuum. “A new desk or a new pod can start in the cloud and very quickly research and eliminate strategies. Once something has proven to have value, it can be brought more in-house. That continuum – from experimentation and failing fast to in-house infrastructure where there’s real edge – is a very healthy strategy.”
Simpson closed on the market structure changes that will reshape both sides of the equation over the next 24 months: volume growth, the move towards 24/7 trading, increased tokenisation, further liquidity fragmentation, and a growing convergence between real-time and historical data at the level three level. The infrastructure that handles all of that will need to do so under conditions that no current architecture has yet been fully tested against.
For firms currently reassessing their infrastructure, Guerrieri proposed a single test: “Does this setup reduce the time to validate insight? If it does, it is helping the investment process. If it is only adding a complexity layer, it is probably just another layer of infrastructure theatre.”
That is the discipline behind the hybrid consensus. Build what is proprietary. Buy what is commodity. And be honest about which is which.
The recorded webinar can be found here.
Subscribe to our newsletter
