What a market view must contain before a financial agent builds a position

Before position sizing, a financial agent should build one concrete object: an accepted forecast packet. It is not a research memo, a confidence adjective, or a proposed order. It is a dated conditional forecast with a defined outcome, an auditable evidence snapshot, a categorical admission result, and a controlled interface to the sizing policy. The packet becomes decision-ready only when another component can interpret and evaluate the forecast without asking what its author meant, reconstruct what the author could have known, see why the packet was admitted, and use it without inheriting a disguised position.

That sequence separates four questions that prose often collapses. What exactly is predicted? What information was admissible at the time? Which defects stop the packet? What may sizing receive after it passes? Each answer creates the premise for the next. A forecast cannot be audited until it has one meaning; an audit cannot govern action until defects have routes; and admission is not operational until the handoff draws a hard line between belief and portfolio choice.

The object can be sketched before those details are filled in. The following is a provisional architecture, not an industry-standard schema:

accepted_forecast_packet:
  forecast: {claim, as_of, horizon, resolution, uncertainty}
  evidence: {snapshot_ref, lineage_ref}
  admission: {status}
  use_envelope: {intended_use, scope, warnings, valid_until}

Those four slices are the minimum cross-domain core: one resolvable claim, point-in-time evidence, one categorical gate result, and no action fields. Exact source identifiers, transformations, uncertainty details, validation artifacts, thresholds, and warning vocabularies are application-dependent controls. The rest of the article builds and checks this same packet one slice at a time; it does not turn the sketch into a production wire protocol.

What exactly is the agent predicting?

The first task is to turn a thesis into a forecast that will resolve only one way. “Rates are bullish,” “inflation should cool,” and “the stock looks attractive” may guide inquiry, but none tells a later observer which realized quantity would confirm or contradict the view. A forecast contract must identify the target and reference series, market or geographic scope, unit, and transformation and aggregation convention. It must also specify the forecast origin, horizon, observation time, resolution source, treatment of missing or exceptional observations, and the assumptions under which the forecast applies.

These fields change meaning, not just presentation. The Federal Reserve's projections, for example, distinguish fourth-quarter-to-fourth-quarter changes in GDP and inflation, a fourth-quarter average unemployment rate, and a year-end midpoint for the federal funds target range; they also state that projections are conditioned on information available at the meeting and each participant's policy assumptions.[1] If an agent records only year-end unemployment or the policy rate, it has discarded distinctions that determine what eventually counts as the outcome.

The plain rule for uncertainty is: name what the number means, then choose an evaluation rule consistent with that meaning. A probability should name its event; an interval should name its coverage and construction; scenarios should have clear conditions and, when they are meant to be probabilistic, weights. A point forecast is not automatically defective. A full predictive distribution is useful when sizing needs the shape or tails of possible outcomes, but it is not universally mandatory. A point value can be well-defined when it names the statistical functional—such as a mean, median, or quantile—and is paired with a consistent score.[2] A reduced summary can also be sufficient when a predeclared downstream loss makes it so.[3] An unlabeled best estimate is the problem: squared error makes the mean optimal, absolute error points to a median, and a tail-sensitive decision may need information neither contains.

The practical test is resolvability. Give the packet to someone who did not write it and ask: “When the horizon ends, can you identify the observation, apply the resolution rule, and score the original claim without reinterpretation?” If two competent reviewers can choose different observations or success criteria, return the packet for clarification. Later market moves must not decide what the author “must have meant.”

The forecast slice of the provisional packet therefore includes:

• as_of and forecast_created_at;
• target, reference_series, market_scope, unit, and transformation;
• horizon, resolve_at, resolution_source, and resolution_rule;
• conditions and declared out-of-scope regimes;
• forecast_value, forecast_type, and its statistical functional or probability meaning; and
• evaluation_rule or the named decision loss that justifies a reduced summary.

Complete these fields first, then run the independent-reader test and reject ambiguity before collecting more persuasive evidence. Once the claim has only one resolvable meaning, the next question is whether every material input actually belonged to the information set available at as_of.

What evidence was admissible at the cutoff?

A precise forecast can still be invalid because it used revised data, a late publication, a stale capture, or a transformation nobody can reproduce. The cure is not a bibliography. It is a frozen evidence snapshot that records the state of each material input and the path from those inputs to the forecast.

Time needs more than one field. The date an observation describes is different from the time that version was released and from the time the agent retrieved it. Official real-time systems make this distinction explicit. Statistics Canada uses a vintage dimension linked to the official release date.[4] The St. Louis Fed API separately distinguishes observation dates from real-time periods and vintage dates.[5] A current link to a macro series may therefore be an inadequate record because it can return a revision that did not exist at the forecast cutoff.

For every material input, store its source and series identity, captured value or durable content reference, observation date, and official publication or release timestamp. Also retain the retrieval timestamp and vintage or revision state. Freeze enough content to reproduce what was seen; a URL alone can change. Record missingness and substitutions instead of silently filling them. Point-in-time controls matter because current or end-of-sample vintages can exaggerate predictive power and alter model comparisons.[6] This does not mean every revision changes the decision. It means the agent cannot know whether one mattered unless the historical information set is recoverable.

The snapshot must also preserve derivation. Link inputs and outputs through the material activities that selected, cleaned, normalized, aggregated, filtered, or modeled them. Record who or what performed those activities and when. The W3C provenance model's entity–activity–agent structure is a useful general pattern for representing these transformations and responsibilities.[7] For a financial agent, the record should also retain variable definitions, model and prompt versions where material, parameter choices, missing-data treatment, exclusions, assumption changes, and the code or rule version that produced the forecast.

Finally, freeze challenge material, not just support. Store material adverse observations, plausible alternative explanations, excluded inputs and reasons, and any change from the predeclared analysis plan. A protocol-to-publication study in another regulated research setting shows why this control matters: statistically significant outcomes were more likely to be fully reported, and primary outcomes were often changed, introduced, or omitted.[8] This is not a finance mandate, but it demonstrates the general failure mode. A complete trail of favorable inputs can still be selectively constructed.

Run the snapshot audit input by input. Confirm that each release preceded as_of; retrieve the actual vintage rather than today's value; replay material transformations; account for missing and excluded data; and ask whether credible counterevidence would change the forecast, its conditions, or its eligibility. If a critical input cannot be reconstructed, mark the defect. The packet's evidence slice now contains a point-in-time snapshot and lineage, not a changeable bibliography. That snapshot makes defects visible. Deciding whether a defect is fatal, repairable, or a reason not to predict is the job of admission.

When must the agent stop instead of shrinking the position?

The admission gate answers whether the packet has earned access to sizing. It must operate before the sizing policy and return a category, not a quality score that quietly becomes a tiny trade. Smaller exposure can manage residual risk in a valid forecast. It cannot remove look-ahead leakage, restore missing provenance, repair an out-of-scope model, resolve a contradiction, or calibrate an uninterpretable confidence value.

One proposed admission gate uses four explicit outcomes:

1. rejected means the packet's evidentiary basis or intended use is invalid. Examples include future-information leakage, a fundamental model or implementation error, unsupported market or regime scope, a broken integrity check, or an evaluation process the agent could manipulate. Banking model-risk guidance identifies both fundamental error and inappropriate use or misunderstood limitations as sources of model risk, and emphasizes objective, informed challenge.[9] The guidance is a governance pattern here, not a universal legal rule for trading agents.
2. needs-evidence means the blocker is potentially curable: an unavailable vintage, stale or sparse support, missing validation, or a material contradiction that additional verified evidence might resolve. The packet stays outside sizing while the named evidence is acquired, then returns to the gate. Evidence-acquisition research in selective prediction shows that some low-confidence cases can be recovered without increasing error when the added evidence is reliable and the underlying confidence is calibrated.[10]
3. abstain means the claim and evidence are admissible, but the system cannot meet a validated error tolerance at the required coverage for this case. This is not the same as rejection: nothing necessarily invalidates the packet's construction, yet the agent should decline to issue a usable forecast at the current operating point. The risk–coverage threshold must be validated for the actual domain and failure slices. In financial regime classification, headline selective accuracy can largely reflect label persistence even when performance fails where transitions matter, so raw confidence is not enough.[11]
4. accepted means the packet passed validity, intended-use, evidence-integrity, and calibrated-uncertainty checks. Acceptance authorizes a handoff; it does not determine exposure.

These distinctions prevent category drift. Suppose a critical series cannot be tied to a historical vintage. That is needs-evidence if the correct capture can still be obtained, but rejected if the packet was evaluated with known future information and the claim can no longer be reconstructed honestly. Suppose the data are admissible and the model is in scope, but uncertainty exceeds the predeclared operating limit. That is abstain. “Accepted with lower conviction” is appropriate only after the checks pass and the remaining uncertainty is both interpretable and permitted by the sizing policy.

For every gate run, record the check identifier, tested input or condition, result, evidence used, and threshold and validation version. Also retain the disposition, restrictions, unresolved warnings, and reviewer or service identity. Tests should cover resolvability, temporal leakage, source integrity, transformation replay, applicability conditions, counterevidence disposition, uncertainty interpretation, calibration at the chosen operating point, and packet expiry. A gate whose thresholds are chosen after seeing the desired position is not a gate; it is a justification device.

Apply the gate and stop on any non-accepted outcome. Write that categorical result into the packet's admission slice; do not encode it as a quality score. Only after the packet is accepted may residual quantified uncertainty influence position size. That leaves one final question: how can sizing consume the belief without inheriting either hidden defects or a hidden trade?

What may cross into position sizing?

The handoff should contain the accepted belief and its use envelope. This final step completes the provisional packet rather than creating a second object. The forecast, evidence, and admission slices pass unchanged—or by an immutable reference—while the use_envelope states intended use, applicability limits, known limitations, unresolved warnings, relevant gate and threshold versions, and finite validity metadata.

Enforce three receiving rules. First, sizing accepts only status: accepted. Second, it rejects expired packets and packets outside their declared use envelope. Third, it rejects action fields in the forecast payload. Model-risk guidance supports conditioning use on intended purpose, understood limitations, validation, and continuing performance assessment; portfolio guidance separately combines capital-market expectations with investor objectives and constraints to produce allocation.[12, 13] Those sources support the conceptual boundary, not the exact serialization proposed here.

Sizing must not invent the missing link between a domain forecast and portfolio returns. An accepted forecast about inflation, policy, an industry, or another non-tradable target may still sit upstream of sizing. Unless it is already expressed in an admitted risk-factor, return, or scenario representation that the receiving policy has been validated to use, it needs another pre-sizing forecast stage. That translation must pass the same controls: a resolvable claim, point-in-time evidence, a categorical admission result, and a bounded use envelope. Only then may sizing consume it.

Instrument selection, current holdings, covariance with the portfolio, risk budget, mandate constraints, liquidity, taxes, concentration limits, target exposure, leverage, rebalancing, transaction costs, order type, and execution timing all belong downstream. The sizing policy combines those facts with the accepted packet. If the market-view component emits buy, weight, units, or leverage, it has hidden an allocation decision inside analysis. A later loss can no longer be cleanly assigned to the forecast, sizing rule, or execution.

The implementation boundary remains intentionally modest. This evidence set supports a separated, auditable forecast-to-allocation interface, but it did not verify a maintained released system that enforces this exact typed packet in code.

Make the negative boundary machine-checkable. Allow belief, provenance, admission, warnings, scope, and validity fields; deny action fields; log the packet hash received by sizing; and make the sizing decision reference that immutable packet. The completed object still has the minimum invariant shown at the start: one resolvable claim, point-in-time evidence, one categorical gate result, and no action fields. The market-view component says what outcome it expects and within what bounds; the sizing component decides what, if anything, the portfolio should do.

Conclusion

Before a financial agent sizes a position, its market judgment should be an accepted forecast packet, not persuasive prose or an early trade. Specify one observable claim and its evaluative meaning. Freeze the information set and transformation trail that existed at the forecast cutoff, including counterevidence. Route invalid, repairable, and high-uncertainty cases to rejected, needs-evidence, or abstain; do not disguise them as small positions. If a domain forecast still needs translation into an admitted risk-factor, return, or scenario representation, perform and validate that translation before sizing. Then hand sizing only the accepted belief and its use envelope, leaving instruments, exposure, leverage, and execution to a separate policy.

This procedure creates a testable boundary. A reader can specify the forecast object, reconstruct the admissible evidence, run the admission gate, and define what crosses into sizing. When any one of those operations fails, the agent does not yet possess a decision-ready market view.

Limitations

The packet fields are a conservative synthesis, not a universal financial-industry standard. This evidence set did not verify the exact typed interface against a maintained released implementation. Numerical calibration thresholds, acceptable coverage, scoring rules, and decision losses are application-specific; they must be validated for the target, horizon, market regime, and cost of error. Point-in-time data controls are essential for reconstruction, but their measured impact varies by dataset and model. Several governance lessons come from banking supervision or other regulated research settings. They support audit mechanisms rather than impose a legal requirement on every financial agent.

Finite valid_until metadata prevents an obviously stale packet from entering sizing, but it does not define the full lifecycle. When an accepted view should be reviewed, superseded, or retired—and how those events should be audited—is the next question, not part of this initial handoff.