From Score to Action: What the Model Can and Cannot Decide

zhulinsen/daily_stock_analysis is most persuasive when its model turns a bundle of market observations into a clear recommendation. That is also the moment when a reader most needs to slow down. The key question is what the dashboard’s score, confidence, action, and conditions each mean—and which of those meanings are still controlled by the surrounding pipeline. Understanding that contract lets the reader use a result as a bounded recommendation rather than mistake polished language for an order or a guarantee.

The project does not treat a dashboard as one field with several aliases. It carries multiple layers of interpretation because a direction, a confidence level, and an executable-looking plan are different claims. The model proposes a synthesis inside that structure. Parsing, normalization, phase checks, data-quality rules, and persistence determine what survives as the system’s final result.

A dashboard has several kinds of truth

The headline score is a useful starting point, but it is not the whole decision. DSA keeps a compatibility decision_type with three values—buy, hold, and sell—while also supporting a richer action vocabulary: buy, add, hold, reduce, sell, watch, avoid, and alert. The three-state field is convenient for legacy statistics and broad reporting. The eight-state action expresses a more precise relationship to a possible decision.

A watch action, for example, can mean that the evidence is interesting but a trigger has not arrived. Reduce can mean that risk has risen enough to lower exposure without necessarily claiming that every position must be closed immediately. Avoid can express a stronger boundary than a neutral hold. These distinctions are lost if every recommendation is forced into a single buy/hold/sell word.

The score and confidence answer different questions as well. The project’s canonical score scale maps high scores to stronger bullish actions, middle scores to watch or hold conditions, and low scores to reduce or sell actions. A score is therefore a normalized summary of the model’s directional assessment. Confidence expresses how strongly the conclusion is held given the available evidence and constraints. A high score with degraded core data is not automatically a high-confidence decision. If a final action does not line up with its score band, the dashboard can retain a guardrail reason explaining why the action was downgraded or otherwise constrained.

A reader can use the following distinction:

The rest of the dashboard supplies the missing conditions. The coreconclusion gives a short statement and separates guidance for a person with no position from guidance for someone already holding the security.