Build or buy an automated trading system

By the TradeAgentic engineering team·Published 30 August 2026·Updated 30 August 2026

The question behind the question

Firms that ask whether to build or buy usually already have someone capable of building it. The real question is not capability, it is whether the two to three years and the permanent maintenance obligation are the best use of that person. Below is what building actually costs, and where buying stops making sense.

What building actually involves

The strategy is the small part

Most people estimating this project price the strategy work, because that is the interesting part. The strategy is perhaps a fifth of it. The rest is data handling, order lifecycle, reconciliation, protection, restart behaviour, entitlement handling and the research harness, none of which is interesting and all of which is required.

Reconciliation is where projects stall

Rebuilding realised profit and fees from the broker's own record, and reconciling positions and protective orders before placing anything new, is unglamorous and much harder than it sounds. Systems that skip it appear to work and quietly diverge from reality.

Restart behaviour decides whether you can sleep

What happens to an open position when the process dies, the machine reboots or the connection drops for ten minutes. Getting this right is most of the difference between a prototype and something you would leave running.

The research harness is a second product

A backtester with a real cost model, honest out-of-sample validation and reproducible results is a substantial piece of software in its own right. Without it you cannot tell whether a change helped, which means you cannot safely change anything.

The costs that outlast the build

Maintenance never ends

Brokers change APIs, exchanges change entitlements, operating systems change security requirements. A system nobody maintains stops working, usually at the worst time. Budget a permanent fraction of an engineer, not a project.

Key person risk becomes structural

If one person built it, the firm's trading depends on that person staying. This is the risk that most often converts a successful in-house build into a licensing conversation two years later.

Opportunity cost is the real number

Two to three years of a capable engineer is the actual price, and it is usually far larger than any license. The question is whether what you would build is meaningfully different from what you could buy, and often it is not.

When building is genuinely right

When the edge is the system itself

If your advantage is a structural or latency edge that no licensed product could express, you have to build it, and no license is a substitute. This is a real case and it is rarer than people think.

When you cannot accept an outside dependency

Some mandates make any external software dependency unacceptable regardless of cost. That is a policy answer, not an economic one, and it settles the question.

When you already have the team and the harness

If you already run a research harness and an execution stack, adding to it is incremental rather than a new build, and the economics look completely different.

The middle option most firms miss

A source license is buying and building

A license that includes source is the arrangement firms usually want when they say they need to build: a working, reconciled, protected system today, with the ability to change it. You inherit part of the maintenance obligation, which is the honest trade, but you skip the two years.

Evaluate before you estimate

Run a licensed system in paper for a quarter before committing to a build. It is the cheapest way to discover what you actually need, and firms frequently find that the thing they were about to build already exists.

Estimating your own build honestly

Count the parts nobody demos

Write down every component before estimating: data ingestion and gap handling, order lifecycle and partial fills, position and protection reconciliation, restart and recovery, entitlement handling, the research harness, and the interface. Estimate each separately. The total will surprise whoever proposed the project.

Price the second year, not the first

First-year enthusiasm carries a build a long way. Ask what happens in year two when the person who built it is on another project and a broker changes an API. That is the year the true cost appears.

Decide what proves it works before you start

Agree the acceptance standard in advance: reconciled to the broker, surviving a mid-session restart, reproducing a backtest from one command. Builds without a defined finish line do not have one.