Build vs. buy: the trading platform decision.
A framework for the build, buy, or white-label decision, drawn from engagements on each side of the question.
The wrong question
The question 'build or buy?' is the wrong starting point. Modern trading platforms comprise multiple layers, and the appropriate choice can differ for each.
The remainder of this note works through each layer.
Layer one: order management
Off-the-shelf OMS products work, but enforce a specific workflow. For institutions with standard workflows, this is acceptable. For institutions with workflow specialisation — bespoke compliance, unusual order types, exotic asset classes — the platform constrains the business.
Heuristic: if the trading workflow can be described in a single page, buy. If it requires extended documentation, build.
Layer two: execution routing
Execution routing is more frequently appropriate to buy than to build. The relevant market-microstructure work is provided by venues and FIX engines. Buying applies unless the institution is a market maker or high-frequency firm.
Layer three: market data
Market data is the layer at which most build decisions become untenable. Vendor brochures present streaming quotes, snapshots, and historical bars. The realities — conflation, throttling, exchange permissioning, corporate actions, normalisation across venues, replay — are the work.
Licensing the feed and the normalisation layer is generally appropriate. Distribution is built only where specific latency or fan-out requirements demand it.
Layer four: position keeping and P&L
For asset managers, this layer is appropriate to build — typically with an experienced specialist. Off-the-shelf portfolio platforms are designed for the average case. Incorrect P&L causes lasting client damage.
Layer five: client-facing applications
Web, mobile, and desktop trading clients are where customer experience is determined. Purchased clients carry the vendor's UX. White-label clients permit reskinning within constraints.
Building this layer is appropriate where retention depends materially on the application. White-labelling and subsequent replacement is a valid alternative.
Layer six: AI, analytics, and insights
Production-grade AI requires a model-serving layer, a monitoring stack, a feedback loop, and an experienced machine-learning engineer. AI features marketed without these foundations tend to underperform when examined.
Build where staffing supports it. Defer the feature otherwise.
Total cost
A full custom trading platform, delivered to production standard, requires USD 4–10M in the first year and USD 1–3M annually in operating cost. A white-label deployment is typically USD 200–500k annually with a delivery time of months rather than years.
The non-engineering cost of building is the executive time required, typically a substantial fraction of leadership attention over multiple years.
Summary
For most institutions, the appropriate path is to white-label the platform, build the OMS workflow that constitutes differentiation, and license the remaining layers.
Custom builds are appropriate where the workflow itself is the product.