From research idea to panel

Use this workflow when you want to turn a topic, mechanism, or empirical question into an auditable call-level panel. The worked example walks through every step with a concrete construct.

Start with a topic

Write the research idea as a short topic statement before opening the tools. Define the firms, countries, sectors, and time period that matter for the analysis.

This first step matters because NL Analytics measures are query-dependent. A broad query can increase recall but add false positives. A narrow phrase can improve precision but miss relevant language.

Translate the idea into search terms

List the terms, phrases, synonyms, and exclusions that represent the concept. Use the Keyword Tool to find related terms and validate whether suggested terms fit the intended construct.

Record the final query string. Small changes to terms, plurals, exclusions, or Boolean logic can change every downstream metric.

Run an exploratory search

Use the Risk Tool to search earnings-call transcripts over the relevant date range.

Review whether the topic appears often enough and varies across the firms, countries, sectors, and periods that matter for the project — the coverage-check procedure shows how. Do this before interpreting trends or group differences.

Inspect matched sentences

Open matched text in the Snippet Tool. Check whether the sentences actually describe the intended concept.

Look for:

• False positives: terms that match but do not represent the construct.
• False negatives: relevant language that the query misses.
• Ambiguous terms that need exclusions.
• Useful synonyms, phrases, or plural forms to add.

Then refine the query and rerun until the matches support the construct.

Decide whether to continue

After the exploratory round, make an explicit decision:

• Refine the query if matches are mostly right but noisy or incomplete.
• Narrow the research question if the topic is real in the corpus but only well covered for certain firms, sectors, countries, or periods.
• Abandon or postpone the topic if it is too rare in earnings-call language to support the design — better to find out now than after a semester of analysis.

Define the measure

Interpret the output using the metric definitions. Exposure, Risk, Positive Sentiment, Negative Sentiment, and Sentiment are raw sentence-count measures — not causal estimates, forecasts, investment recommendations, classifier labels, or complete summaries of business risk.

If the search used section, speaker, or adjacent-sentence options, note how they change the counts.

Export the panel

When the query is defensible, export the research panel. Use earningscallID as the unique document identifier for joins and deduplication.

Join and document

Join the export to the downstream datasets and decide how to handle normalization by transcript length.

Finally, record the full definition — query, options, export date, coverage check, join key, normalization choice — per the reproducibility checklist, so the measure can be explained to coauthors, seminar audiences, and referees.