Scoring Management Tone and Evasiveness

An analyst asks a CEO directly: "are you seeing any weakness in demand this quarter?" Two very different answers can look similarly positive on a naive sentiment score if both avoid explicitly negative words: "No, demand remains strong across all our segments" versus "We're monitoring the demand environment closely and remain focused on executing our strategy regardless of near-term conditions." The first is a direct, confident answer. The second answers a different question than the one asked, using hedged, vague language that says nothing concrete — a classic evasive non-answer. Scoring what was said (sentiment) misses this distinction entirely; scoring how directly it was said (evasiveness) is a separate and often more informative signal.

What evasiveness scoring measures

Evasiveness scoring works from a handful of linguistic markers rather than a single word list, because evasion is a pattern of speech, not a vocabulary of "evasive words." Common markers include: hedging language — "may," "could," "we believe," "it's possible that" — used at unusually high frequency; question deflection — the answer's topic doesn't match the question's topic, often detectable by comparing the semantic similarity between question and answer; non-answers — a response with normal sentence length and structure but no concrete quantitative or factual content where one was requested; and repetition of prepared-remarks language — falling back on a scripted phrase instead of engaging with the specific question, often literally repeating a sentence used earlier on the same call.

Worked example: scoring three answers to the same question

Three CFOs are asked the same analyst question — "what's driving the gross margin decline this quarter?" Answer 1: "Input costs rose 8% and we didn't fully pass that through in pricing, so margin came down about 140 basis points from that alone." Answer 2: "There are a number of puts and takes this quarter, and we're always looking at ways to improve efficiency across the business." Answer 3: "We don't break out margin drivers by line item, but overall we feel good about our cost discipline." A scoring pipeline would flag Answer 1 as low-evasiveness — it directly answers with a specific cause and a number. Answer 2 would score high on evasiveness: it contains no concrete cause, no number, and its semantic content barely overlaps the question asked. Answer 3 sits in between: it explicitly declines to answer the specific question (itself a marker) but offers a generic positive framing in its place — a pattern worth flagging on its own, since an explicit refusal to disclose is often more informative than vague hedging.

What this means in practice

Evasiveness scores are typically used as a feature alongside — not instead of — sentiment and financial data, since a rise in evasive answering around a specific topic (margins, a particular segment, a pending investigation) can be an early signal that management doesn't have a good answer yet, well before that shows up in reported numbers. Scoring pipelines usually need to be built per-question-per-answer pair, matching each analyst question to the response that follows, which requires reasonably clean speaker-turn segmentation of the transcript first.