Research Outputs Versus Trade Execution

Define what the output actually means

A score may rank instruments, estimate a risk measure or identify a condition for further review. It may omit position size, account eligibility, order type and exit behavior. Ask the provider to define the intended interpretation and the information needed before an investment decision. Avoid allowing a developer's convenient mapping from score to order to become an undocumented extension of the model's purpose or an assumed feature of the licence being evaluated.

The SEC's performance-claims bulletin distinguishes hypothetical backtests from actual results. Carry that distinction into the data model: a research observation, a simulated transaction and a broker-confirmed execution should have different record types. A real market price used in a simulation does not turn the simulation into a real trade. Keep research evidence useful on its own terms without implying that it demonstrates the operational behavior of an account connected to a broker.

Separate the event clocks

Record input availability, calculation completion, publication, receipt, decision and order times where relevant. A model may reference a market observation earlier than its publication time. A buyer cannot assume it acted at that earlier observation simply because the field is called signal time. State the precision and clock source, and account for the possibility that timestamps from different systems are not perfectly synchronized when assessing very short delays or event ordering.

Use order identifiers and execution identifiers to connect later events without collapsing them. Submission is an attempt to send an order; acknowledgement is not necessarily a fill; partial fills leave a remaining quantity. Rejections, cancellations and replacements need explicit states. These operational distinctions matter even for a slow-moving model, because an apparently minor retry or stale-message error can change position exposure independently of whether the original research calculation was correct.

Keep research price conventions intact

When reviewing a research result, obtain its declared observation window and price-selection rule. If the convention selects the best hypothetical price within a bounded window, reproduce that rule and label it accordingly. Do not silently replace it with a first-available price, and do not enlarge the window to obtain a more favorable observation. Comparability requires preserving the frozen method even when the buyer also wants a separate execution-oriented sensitivity analysis.

Actual execution evidence comes from the relevant operational records, not from choosing a more plausible simulation price. A separate simulation can incorporate latency, spread and partial fills under stated assumptions, but it remains hypothetical. Keep those assumptions beside its results and identify the account constraints it omits. Procurement should ask what evidence supports the intended use, rather than encourage a research supplier to relabel an existing backtest as execution history to satisfy an unsuitable checklist.

Work through a hypothetical order timeline

Consider a wholly hypothetical, same-day UTC timeline with synchronized clocks. A model publishes at 09:30:00, the buyer receives it at 09:30:05, submits an order at 09:30:08 and receives an acknowledgement at 09:30:09. The intervals are five seconds for delivery, three seconds from receipt to submission and one second to acknowledgement, totaling nine seconds. Those times are invented process inputs, not measured service levels or claims about a real trading system.

The synthetic order requests one hundred units, and a later simulated execution message reports eighty units filled. Twenty units remain unfilled, so the filled fraction is 80%. No price or investment performance follows from that quantity calculation. The buyer's test checks whether its position record contains eighty units and its open-order record contains twenty, while the original research event stays unchanged. It does not record one hundred executed units merely because the order was accepted.

Review authorization and controls separately

Ask whether the actual model agreement covers the proposed automated use and whether upstream data permissions also fit. Obtain jurisdiction-specific review of the buyer's activities and regulatory position. The SEC's market-access FAQ addresses risk controls for broker-dealers within Rule 15c3-5's scope; it is not a licence for a model buyer to trade and not a universal rule for every research user. Its scope illustrates why execution requires a distinct regulatory and operational assessment.

For a proposed implementation, identify owners for account limits, stale-data checks, duplicate suppression, order cancellation and position reconciliation. Test with isolated synthetic fixtures before any separately authorized live process. A trial should not quietly acquire trading credentials because an API makes that convenient. Keep the purchasing decision, operational readiness decision and any authorization to submit orders distinct, with evidence appropriate to each and no automatic promotion from research evaluation to real-account activity.

Report evidence without merging categories

The final assessment should separately describe research reproducibility, delivery behavior, simulation assumptions and any actual execution evidence supplied. Identify gaps without filling them with proxies. A reliable feed can remain unsuitable for a required deadline, while a useful research model can be entirely appropriate for manual analysis. The right procurement conclusion depends on the named use, not on an assumption that every quantitative output should eventually become an automated trading instruction.

Preserve model and operational versions alongside timestamps so later changes can be investigated. Do not overwrite a research reference when an order receives a different fill or is rejected. Both records explain different events. The buyer should be able to answer what the model indicated, what the decision process authorized and what the broker actually confirmed. Keeping those answers separate is essential to honest evaluation and to avoiding unsupported performance or execution claims.

• Acknowledge, fill and remaining quantity are separate states.
• Research acceptance does not authorize real-account execution.

Source notes

1. SEC Investor.gov: Investor Bulletin on Performance Claims
2. SEC staff: Market Access Risk Management Controls FAQ