Research that survives contact with a real market.

Research and production are different problems. Research rewards iteration speed and expressive libraries; execution rewards determinism, predictable memory and low tail latency. Most systems end up bilingual, and the boundary between the two is a design decision rather than an accident.

Market data is large, strictly ordered and almost always queried by time and symbol. General-purpose relational stores are the wrong shape for it, and the cost of that mismatch shows up as research that is too slow to iterate on.

The layer that most often decides whether a backtest means anything. What matters is not the vendor name but the resolution, the timestamp discipline, whether the history is point-in-time, and whether the feed the research used resembles the feed production will see.

Useful where it is genuinely orthogonal to price and survives an honest evaluation. Most alternative datasets are shorter, sparser and more revised than they appear, which is a research design constraint before it is a sourcing question.

Where most integration work actually lives. Saying “the API” hides the distinction between a REST endpoint and a binary session protocol with sequence numbers, gap fill and a resend request, and the second is what an order path is built on.

A delivered system usually has to sit alongside an OMS or EMS a client already runs, not replace it. Execution logic is quantitative work in its own right: the cost of crossing the spread is one of the few edges that can be improved without predicting anything.

Market structure differs enough between venues that it changes strategy design, not just configuration: tick size, auction mechanics, order types, self-match prevention and fee schedules all show up in realised performance. Who provides liquidity on a given book is part of that structure, and knowing it is microstructure research rather than a commercial relationship.

The part that decides whether the position on the screen is the position that exists. Reconciliation against the clearing record is not administrative overhead; it is how a silent execution bug gets found before it compounds.

Second-order risk is where this firm's method has most to say. The distinction that matters is between risk that is reported and risk that is understood, which is usually a question of whether the cross terms were computed at all.

A strategy that cannot be operated has not been finished. Most of what separates a research prototype from a running system is in this layer, and almost none of it is visible in a backtest.

Relevant when a mandate is genuinely latency-sensitive, which is less often than it is claimed. Where it applies, these are the constraints a design has to respect and the terms in which execution quality gets measured. This is execution engineering: the firm builds and measures against this infrastructure rather than operating a proprietary low-latency trading business.