Who this is for — Anyone who wants to test a trading pipeline, order handling and operating procedures with current data without putting capital at risk, while understanding which aspects of the real market remain outside the test.
Paper trading is an environment in which decisions and orders are recorded or simulated without creating the normal economic exposure of a real trade. It may use current quotes, replayed data or a broker's simulation engine. It is useful for checking that data, signals, sizing, instructions, alerts and the journal work in the intended sequence. Its P&L nevertheless remains hypothetical.
Three concepts must be kept separate. Out of sample describes the relationship between data and development: a sample is OOS if it did not guide model choices. Forward test describes a prospective evaluation of a frozen version. Paper, shadow and micro-live instead describe the environment or implementation mode. A forward test may be conducted in paper, shadow or micro-live; none of these terms is synonymous with the others.
Four environments, different questions
| Environment | Does an order reach the market? | What it observes well | Distinctive limitation |
|---|---|---|---|
| Paper trading | No | Rules, interface, simulated accounting, routines | Fills and P&L depend on the simulator |
| Shadow mode | No | Complete automated pipeline and real-time decisions | The shadow order does not interact with the book or broker |
| Micro-live | Yes, with reduced exposure | Routing, rejects, fills and real reconciliation | Impact and pressure do not scale linearly |
| Live at intended size | Yes | Economic process within the actual perimeter | Capital is at risk; this is no longer a cost-free test |
Shadow mode is often a more controlled paper variant: the system produces the order it would have sent and preserves timestamps and reasons, but a block prevents routing. Micro-live instead uses real orders of small size. Reducing size limits exposure, but may change priority, fill probability, impact and human behaviour; it does not automatically reproduce the target scale.
What it can test
A well-designed paper test can show whether:
- feeds arrive with the expected timestamps, frequency and symbols;
- features and signals are calculated without using data that were not yet available;
- sizing, limits, calendars and corporate actions are applied correctly;
- orders, cancellations and amendments follow the intended state machine;
- the system handles openings, closings, shortened sessions, missing data and restarts;
- logs, alerts, journal and reconciliation make every decision reconstructable;
- the operator follows the agreed checklists, authorisations and escalation paths.
It may also produce prospective evidence about strategy behaviour if the version was frozen before the test began. That evidence must not be confused with a complete measure of tradability.
What it does not faithfully reproduce
A simulator does not actually occupy a position in the order-book queue, and its order does not alter available liquidity. Partial fills, latency, market impact and adverse selection can be approximated, but not observed as they are for a real order. Broker rejects, credit limits, borrow availability, stock recalls, margin, settlement and funding costs may also be absent or simplified.
Virtual money does not fully recreate the incentives, pressure, responsibility and consequences of a real loss. NFA Interpretive Notice 9025 makes this explicit within its scope: hypothetical performance does not represent actual trading and cannot fully account for liquidity, slippage or the ability to adhere to a programme under losses. It is a rule concerning promotional communications by NFA members, not a universal validation criterion; its caution about simulated performance remains relevant.
Operating protocol
- Define the objective. Separate pipeline checking, procedural training and prospective evaluation of the strategy.
- Identify the version. Freeze code, configuration, universe, data, sizing, costs, benchmark and criteria for leaving the test.
- Describe the environment. Record provider, broker or simulator, fill model, latency, schedules, order types and known differences from live.
- Use the normal clock. No rewinding, retrospective deletion of signals or substitution of prices after observing the outcome.
- Test failures too. Simulate a stopped feed, stale price, rejected order, partial fill, restart, connection loss and duplicate messages when relevant.
- Reconcile every event. Compare intention, shadow order, simulated fill, position, cash, cost and logs; classify differences.
- Record deviations and interventions. A material correction creates a new version; the earlier period is not erased.
- Decide according to the objective. “Pipeline verified”, “insufficient evidence” and “open operational problem” are outcomes distinct from P&L.
No universal duration
Four weeks, a certain number of trades or one profitable month are not general criteria. Useful duration depends on signal frequency and the need to encounter relevant sessions, events and operational cases. An intraday strategy may produce many correlated orders without covering one rare event; a monthly strategy may require more calendar time while still producing few observations.
Completion criteria should be established beforehand: coverage of order types, reconciliations without material anomalies, successful handling of error cases and an amount of information compatible with the statistical question. Simulated profit alone is not a promotion criterion.
Illustrative example
Example explicitly stated as illustrative — A strategy produces limit orders in shadow mode. The test discovers that two signals are sent twice after a restart and that the simulator considers the full quantity filled as soon as price touches the limit. The first problem is operational; the second is a fill assumption. Both are corrected in a new version, while previous logs and results are retained. No duration or quantity in the example constitutes a validation threshold.
A subsequent micro-live stage may test routing and reconciliation with small real orders. It does not, however, demonstrate that costs and capacity will remain unchanged at the intended size: the transition requires limits, monitoring and escalation.
Limitations
- Quality depends on the simulator model and available data.
- The absence of capital at risk changes incentives and behaviour.
- Paper trading does not demonstrate liquidity, borrow or capacity at target scale.
- A prospective result remains vulnerable to a small sample, a particular regime and multiple testing.
- Correcting the pipeline does not authorise retrospective rewriting of the tested version's history.
Sources
- National Futures Association, Interpretive Notice 9025 — Use of Promotional Material Containing Hypothetical Performance Results — limitations involving hindsight, liquidity, slippage and the absence of real risk; applicable within its NFA scope.
- U.S. Securities and Exchange Commission, Staff Report on Algorithmic Trading in U.S. Capital Markets, 2020 — market structure, automation and the effects of real interaction between orders and venues.
- Joseph Simonian, CFA Institute Research Foundation, Investment Model Validation: A Guide for Practitioners, 2024 — proactive validation and the limitations of investment models.
- European Commission, Commission Delegated Regulation (EU) 2017/589, RTS 6 — testing, environments separated from production, controls and deployment within the applicable EU algorithmic-trading scope.