Give waiting decisions useful feedback labels

“Should have traded” is a weak label for a waiting decision. It blends the owner's preferences, the information available at the time and the price path that arrived later. We propose labeling abstentions by their decision reason and the evidence needed to reconsider them.

The DXAP activity guide explains that a completed turn may decide to wait, while an action rejection or execution failure requires a different interpretation. The policy reference places configured checks on the path after the model requests an action. A useful label must preserve that distinction.

Label the missing condition

Consider three illustrative turns with no resulting fill. In the first, the strategy requires a completed candle above a threshold and the candle remains open. In the second, the model proposes an entry that breaches a configured position limit. In the third, an allowed order is submitted but remains unfilled. The economic observation is the same, but only the first is a model decision to wait. The second is policy rejection; the third is execution status.

For the first case, we would ask a reviewer to record the stated missing condition, verify it against the saved input and identify the earliest future evidence that could satisfy it. The label could be “strategy condition absent, supported.” If the condition was present and the agent overlooked it, the label becomes “available condition missed.” If the snapshot is incomplete, the case stays unresolved.

A separate owner-preference field captures whether the owner likes the strategy's waiting rule. Disliking the rule is useful product feedback and might motivate a reviewed strategy proposal. It should leave the historical compliance label intact. The chat guide describes how persistent instructions require review and confirmation before becoming active.

Resist the hindsight shortcut

Suppose the price rises from an illustrative 100 to 105 after the first turn. That later move can supply a forecast outcome if the question and horizon were defined in advance. It cannot establish that the candle was complete at the earlier decision. A model can obey the mandate and miss a favorable move; it can violate the mandate and happen to profit.

Our proposed annotation form would therefore separate reason support, mandate consistency and a fixed-horizon forecast outcome. Economic evaluation would add a declared entry policy and execution costs. Reviewers would first label the information-time record without seeing the future price, then receive outcomes in a separate pass. This allows us to measure whether future outcomes change their compliance judgments.

For an illustrative audit of 30 no-fill turns, report how many were genuine abstentions before calculating abstention-label agreement. If ten were rejected requests and five were pending orders, the abstention population contains fifteen cases. Reporting agreement over all thirty would obscure what was evaluated.

The practical result would be a dataset of answerable questions: which conditions were missing, which were overlooked and which judgments remain uncertain. That dataset could guide better context, clearer explanations or a candidate model intervention. Its immediate test is improved diagnosis of waiting behavior, with forecasting and economic gains evaluated on their own terms.