排行榜數字看似簡單，定義卻不簡單。本文說明評分規則、回測含義，以及命中率被誤用的常見方式。
評分前先寫清：標的、方向、起訖日、基準（絕對、板塊 ETF 或指數）。缺任一要素則分數未定義。
區間是從錨定日 T 起算的交易日數。看完結果再改區間即為挑選樣本。
方向命中可能是「報酬 > 0」；相對命中可能是「報酬減板塊 ETF > 0」。同一評論員在不同定義下排名可差很多。
若下市標的被靜默刪除，歷史準確率會被高估。我們力求按方法論處理未解決或下市結果。
回測依你選擇的日期與信號來源匯總歷史指標，供方向性參考，並非扣除成本稅費後的可交易獲利證明。
Free 可能僅顯示摘要；Pro 解鎖更深表格與匯出。評分規則不因方案而改變。
Vague: “ABC looks cheap and should rip.” Testable: “ABC’s total return exceeds the sector ETF XLY over the next 90 trading days starting 2026-03-01.” Only the second sentence can be scored without inventing hidden assumptions.
Vague: “Smart money loves DEF.” Testable (different object): “Manager M’s reported share count in DEF rose quarter-over-quarter as of 2025-12-31 (13F accepted mid-February).” Notice this is a holdings statement, not a forecast score—and it still needs lag labels.
A 70% hit rate on ten calls is a coin-flip story wearing a suit. Always publish N beside the rate. Prefer longer samples even when the rate looks less flattering. Education rewards calibration more than swagger.
Variance across regimes matters: the same forecaster can look brilliant in a one-way bull tape and ordinary in a choppy year. Segment notes by period without silently dropping embarrassing windows.
Journal the claim on day one, freeze horizon and benchmark, wait, mark hit/miss/void, then review monthly. Link your journal fields to InsightMeter methodology vocabulary so you do not invent private definitions that cannot be audited.
When a live leaderboard and a backtest disagree, write the universe, horizon, and filter differences before preferring the prettier chart. See also the backtest-versus-live guide.
Scores cannot certify character, cannot see undisclosed hedges, and cannot replace primary filings. They summarize history under rules. Keep the risk notice visible: educational materials are not investment advice.
Some claims should be marked unscored rather than forced into hit/miss. Missing horizon, missing benchmark, purely qualitative language (“could be interesting”), or contradictory clauses are refusal cases. Refusing to score is a feature of honest systems.
Delistings and corporate actions need pre-written void or resolution rules. If you improvise after the fact, you recreate survivorship bias under a new name. Methodology should state the rule; your journal should follow it.
Lead with definition: “55% directional hits vs SPY over 90 trading days, N=40, sample 2023–2025, delistings retained as misses per rule.” That sentence is longer and safer than “this guru is elite.”
Ban leaderboard screenshots without the definition caption. If your charting UI cannot show N and horizon, add them in text. Education is allergic to orphan percentages.
Automated matching of claims to outcomes can help at scale, but automation does not remove definitional choices. If the claim parser guesses a horizon you did not intend, the score is about the guess, not the pundit. Review ambiguous parses; document overrides.
InsightMeter describes scoring as rule-based accountability for education. It is not a promise that models found secret alpha. Keep that sentence near any product demo.
只有能用一句話重複其定義時，才應相信數字。另見預測紀錄指南與方法論頁。
如何公平評估預測成績 · Backtest vs live forecast scoring pitfalls · 研究方法
Scoring is a contract with your future self: freeze definitions, wait, mark outcomes, publish N. Break the contract and you are doing theater with spreadsheets.
Use Start here, the forecast track-record guide, and the journal template as companion reading. Optional Pro tables do not change the contract; they only change how much history you can browse.
If a score cannot be restated in one precise sentence, do not share it. Precision is the product of education-first forecast work.
依主題學習: Form 4／內部人素養 · Form 8-K 與基本面 · Form D／豁免發行 · 13F／機構快照