Skip to short-horizon research
Short Term Trading StrategiesCondition / Request / Position

How to specify a short-term strategy

Write an availability-aware short-term specification with session clocks, order states, costs and predeclared rejection conditions.

Original editorial revision: 10 October 2026. Expanded edition: 14 October 2026.

Educational execution planning. Invented costs and clocks are assumptions, not forecasts, broker instructions or a measured strategy record. Check the exact venue, instrument and account rules.

Define one observable question before choosing indicators

Begin with an instrument and a hypothesis that can fail. Name the information observed, the decision clock, eligible opportunities and future label. A scanner hit, completed-bar condition and quote event create different populations. Define whether the project concerns a research reference or a proposed executable strategy. Do not treat the availability of a fast indicator as evidence that the remaining decision chain is implementable.

Write a compatible baseline before evaluating the candidate. It needs the same eligible sample, observation access and accounting convention. Record abstentions and unavailable labels so difficult rows do not disappear. Short-term does not choose an appropriate personal holding horizon or risk budget; those require a separate authorised contract. The first deliverable is a precise research question with boundaries, not a list of fashionable signals selected from the best-looking historical charts.

Build the causal input and session contract

For each feature, retain instrument identity, event time, receipt time, source version and revision policy. A historical row timestamp is not automatically proof that a subscriber could observe the row then. Define warmup and missing-interval handling before computing the feature. A corrected historical series and an original received stream can be different information sets, even when their date labels match.

Use the relevant venue's session calendar and timezone rules. Record auctions, shortened sessions and periods where matching or order entry differs from normal trading. NYSE documentation supplies a bounded schedule reference, not a universal calendar for every instrument. Save the historical mapping used in a test. If the source lacks the precision needed for the proposed rule, narrow the claim or leave that part unresolved instead of inventing a finer event stream.

Specify the action as a complete state transition

Write the activation condition, instruction identifier, requested quantity, order type and expiry. Then describe acknowledgement, partial execution, amendment, cancellation and final reconciliation. A trigger condition becoming true is not already an account position. A submitted exit is not already a completed closure. Keep whole-position strategy intent distinct from individual partial fills caused by execution.

Use a fictional hundred-unit entry that fills forty units and cancels sixty. The final entered quantity is forty, not one hundred or zero. If a later close instruction fills thirty units, ten remain open. Keep those arithmetic states in fixtures before testing market data. The example does not prescribe a partial-exit trading policy; it checks whether the implementation can retain the actual execution states of a declared whole-position instruction.

Freeze entry identity and the native exit contract

Choose the entry objective and candidate-ranking process within the authorised training region. Freeze the selected entry before applying one predeclared exit to validate economics, exposure and holding behaviour. Stop, target where applicable, maximum hold and session response need explicit ownership. A completed-session clock update must not silently introduce a different daily stop-management rule into a minute-based native policy.

Keep the original research fill convention, including directional and observation-window rules, distinct from actual broker executions. Changing it to the first available or a later favourable price is a new experiment. Label a changed exit or discretionary override with its effective version. Repeatedly reopening exit selection after reserved outcomes are known does not create independent validation. Preserve rejected candidates and the information already consumed by each selection decision.

Join economics and constraints to the accepted quantity

Measure costs using a named price basis and genuinely additional charges. Specify increments, contract value, funding and existing exposure. A stop-budget calculation is not a guaranteed loss cap, and margin is not a maximum possible loss. Keep accepted quantity separate from requested quantity so an unfilled remainder cannot receive a hypothetical profit or use an account denominator it never funded.

Declare the response to stale prices, duplicate instructions, unavailable account state and reconciliation mismatch. A fictional delay or spread penalty is a sensitivity input, not an observed percentile. Save the unadjusted reference result next to the conditional implementation result. The latter should show exactly which assumed prices and costs changed it. This guide supplies no universal fee, latency allowance, recommended leverage or account-size threshold.

Reserve evaluation and deliver the replay package

Chronological splitting helps organise evaluation, but label spans and source availability need additional checks. A five-observation future label must not cross a forbidden boundary merely because its starting row is in training. Pin warmup, gaps, fit ownership and the candidate family. Software fixtures test those mechanics; genuinely reserved observations are needed for independent strategy evaluation.

The blank construction record holds these decisions before a real test. Attach input identity, runnable code, candidate history, frozen entry, native exit and complete opportunity ledger to any result. Production release is another layer requiring active cloud revision, route and durable execution evidence; a research live label is not sufficient. A valid research conclusion may be no acceptable model. This page describes the package to build, not a completed backtest or a promise of profitable short-term trading.

Review checkpoints

Evidence requests, not completed study results
ObjectPreserveDo not infer
ObservationAvailable inputs and historical calendarA timestamp alone proves real-time access
ActionRequested, executed and outstanding statesA submitted exit means flat exposure
EvaluationFrozen entry, native exit and reserved sampleA successful software test validates an edge

Download blank review record (CSV)

Headers only. No trades, subscriber outcomes or operating receipts are supplied.

Source documents

Documentation establishes the stated mechanism, not tested trading performance. Checked 10 October 2026; a later version can differ.

Detailed research

Use complete setup fields without claiming a model

The three added cards freeze prior reference, input basis, availability, eligibility and invalidation before classifying nine authored branches. Missing context remains unresolved; failed comparisons remain failed despite an attractive later chart. They assign no actual entry fill or financial exit. Real model work retains the raw causal objective, frozen identity, one selected native exit and independent master. A complete educational predicate is useful but cannot replace an instrument adapter, licensed observations or the full canonical research process.