Export a research panel

Export a research panel after the query has been reviewed and the matched sentences support the intended measure. This page lists the core files, their columns, and the identifiers for downstream joins.

Before exporting

Confirm that:

The query matches the intended construct.
The date range matches the analysis sample.
The topic has enough sample availability and variation for the intended use.
Matched sentences have been inspected in the Snippet Tool.

Download output

Open the completed Risk Tool dataset or find it in My Datasets.
Download the main call-level file, firmlevel.csv.
Download unconditional.csv if the analysis needs overall risk or sentiment counts.
Download gvkeys.csv if the downstream workflow needs a Compustat join.
Download the dataset metadata where available.

The downloadable cuts depend on the dataset, context, and plan. Keep every downloaded file together with the query and dataset settings listed in the reproducibility checklist.

firmlevel.csv

The main export. One row per earnings call in the selected date range, identified by earningscallID, including calls with zero matches.

Column	Definition
`earningscallID`	Unique identifier for the earnings call. Use it for joins, deduplication, and citations.
`company_name`	English company name.
`date`	Date of the earnings call, based on UTC time.
`event_name`	Name of the earnings-call event, usually including the fiscal period discussed.
`start_time`	ISO 8601 timestamp for when the call started, in UTC. Use it when precise timing matters.
`exposure`, `risk`, `positive`, `negative`, `sentiment`	The topic-conditioned measures. See measure definitions.
`nr_of_sentences`	Total number of sentences in the full transcript.
`tickersymbol`	Ticker symbol.
`permid`	Permanent entity identifier, where available in the corpus metadata.
`ric`	Instrument code, where available in the corpus metadata.
`isin`	International Securities Identification Number.
`cik`	SEC Central Index Key, mainly for U.S. entities.
`date_q`	First day of the calendar quarter of the call. Calendar quarter, not fiscal quarter.
`headquarterscountry`	Registered headquarters country.
`trbc`	Business classification code.
`economic_sector`	Label for the first two digits of `trbc`.
`business_sector`	Label for the third and fourth digits of `trbc`.
`nr_of_sentences_filtered`	Number of sentences passing the section, speaker-affiliation, or speaker-title restrictions. Present only when such a restriction was used.
`matched_speaker_names`	Names of speakers whose sentences matched, when a speaker-title restriction was used.

The file can also contain provenance columns describing the scoring mode and columns for measures under development, which these docs do not define.

Identifier columns other than earningscallID can be empty for some calls. Check identifier completeness in your target sample before designing a join around one of them.

unconditional.csv

Overall risk and sentiment counts, not restricted to the query. One row per earnings call.

Column	Definition
`earningscallID`	Unique identifier for the earnings call.
`nr_of_sentences`	Total number of sentences in the full transcript.
`unconditional_positive`	Sentences with at least one positive sentiment word.
`unconditional_negative`	Sentences with at least one negative sentiment word.
`unconditional_risk`	Sentences with at least one risk or uncertainty synonym.
`unconditional_sentiment`	`unconditional_positive - unconditional_negative`, included when sentiment is among the selected measures.

Join it to firmlevel.csv on earningscallID to express topic-conditioned measures relative to all risk or sentiment language in the call. See Overall risk and sentiment for the definitions.

gvkeys.csv

A convenience crosswalk for Compustat joins. One row per earnings call.

Column	Definition
`earningscallID`	Unique identifier for the earnings call.
`gvkey_compustat`	Compustat GVKey, where a match exists.
`sic_compustat`	Four-digit SIC code, where available.
`sic_compustat_description`	Label for the SIC code.

Treat the crosswalk as an input to verify, not as a guaranteed match: check match quality in the downstream dataset and document the checks you performed.

keyword_breakdown.csv

Produced when the dataset export includes the per-keyword breakdown option. One row per keyword and earnings call.

Column	Definition
`keyword`	The query keyword or phrase.
`earningscallID`	Unique identifier for the earnings call.
`exposure`, `risk`, `positive`, `negative`, `sentiment`	The measures counted over only this keyword's matched sentences in this call.

Use it to see which terms drive a measure. For example, check whether a single broad keyword is responsible for most of the variation, or re-derive the measure after dropping a term without running another search.

Dataset metadata

Export the query and dataset settings as JSON where available. Store them with the exported files so a coauthor or referee can see which query, settings, and identifier produced the panel.

Weighted dictionary artifacts

A dataset created from a weighted dictionary query additionally exports its dictionary:

Column	Definition
`rank`	The term's rank by weight, 1 = highest.
`term`	The dictionary term.
`raw_weight`	The term's raw tf-idf weight.
`normalized_weight`	The weight actually used in scoring (median term = 1.0, clamped to 0.25–4.0).

One row per term. Keep this file with the export: it is the full, auditable lexicon behind the weighted measure. In weighted output, the metric_kind, weighting_source, and weighting_normalization columns identify the measure's unit and normalization.

Keep zero rows

firmlevel.csv includes all earnings calls in the selected date range, including calls with zero Exposure or other zero measure values.

Zero rows preserve absence in the selected sample. Removing them changes averages, combinations, and aggregate statistics. See Query quality and zero rows.

Next step

Join the export to the rest of the empirical design: Join and normalize outputs.