Measure definitions
NL Analytics core measures are integer sentence counts calculated at the earnings-call level.
Core measures
| Measure | Output variable | Definition | Calculation |
|---|---|---|---|
| Exposure | exposure | Number of sentences in an earnings call that contain at least one keyword from the user's query. | Count each sentence once if it contains at least one query keyword. |
| Risk | risk | Number of topic-matched sentences that also contain at least one risk or uncertainty synonym. | Count each sentence once if it contains at least one query keyword and at least one synonym for risk, risky, uncertain, or uncertainty. |
| Positive Sentiment | positive | Number of topic-matched sentences that also contain at least one positive sentiment word. | Count each sentence once if it contains at least one query keyword and at least one positive sentiment word. |
| Negative Sentiment | negative | Number of topic-matched sentences that also contain at least one negative sentiment word. | Count each sentence once if it contains at least one query keyword and at least one negative sentiment word. |
| Sentiment | sentiment | Net tone conditional on topic discussion. | positive - negative. |
Positive and negative sentiment words come from the Loughran-McDonald financial sentiment dictionary, with question and questions excluded from the negative word list. The risk synonym list is the deduplicated union of Oxford Thesaurus synonyms for risk, risky, uncertain, and uncertainty, excluding question, questions, and venture.
Calculation rules
The unit of analysis is a sentence inside an earnings-call transcript.
A sentence contributes to Exposure if it contains at least one query keyword. A sentence can contribute to Risk, Positive Sentiment, or Negative Sentiment only after it first matches the topic query.
In the main Risk Tool counts, Positive Sentiment and Negative Sentiment are counted independently. A sentence that contains both positive and negative sentiment words contributes positive = 1, negative = 1, and net sentiment = 0.
How query settings change the counts
The definitions above describe the default query, where the query string and the dictionary words are evaluated on the same single sentence.
Query settings change what is counted:
- The section, speaker-affiliation, and speaker-title filters restrict which sentences are eligible to match the query. The counts then cover only the eligible part of each call, while
nr_of_sentencesstill measures the full transcript. See the normalization caveat. - The adjacent-sentences option does not change query matching, which always operates on single sentences. It widens where risk and sentiment words are detected to a three-sentence window around each matched sentence (one sentence before and after, by the same speaker, where available). Risk and sentiment counts are usually higher with this option on.
Record the full query, including filters and adjacent-sentence behavior. Two datasets created from different queries can contain different measure values even when their query strings match.
Overall risk and sentiment
A separate output reports unconditional counts that are not restricted to the user's query, for each call:
| Measure | Output variable | Calculation |
|---|---|---|
| Overall Risk | unconditional_risk | Sentences with at least one risk or uncertainty synonym. |
| Overall Positive Sentiment | unconditional_positive | Sentences with at least one positive sentiment word. |
| Overall Negative Sentiment | unconditional_negative | Sentences with at least one negative sentiment word. |
| Overall Sentiment | unconditional_sentiment | unconditional_positive - unconditional_negative. |
These counts support analyses that compare topic-linked risk or sentiment with all risk or sentiment language in the same call. They do not replace the topic-conditioned measures. See Export a research panel for the file that contains them.
Weighted measures
Running a weighted dictionary query replaces the plain sentence count with a weighted sum, reported in the unit weighted_matched_sentences.
Each sentence that matches a dictionary term contributes the weight of its matched terms, and the call-level value sums those contributions over the call's matched sentences. Weights are tf-idf term weights from the source measure's construction, normalized so that a median-weight term counts like one ordinary matched sentence: weights are log-scaled, divided by the median, and clamped between 0.25 and 4.0. Weighted values are therefore decimals, not integer counts.
The built-in dictionaries reproduce the firm-level political risk measures of Hassan, Hollander, van Lent, and Tahoun ("Firm-Level Political Risk: Measurement and Effects") and the country risk measures of Hassan, Schreger, Schwedeler, and Tahoun ("Sources and Transmission of Country Risk"). Cite the original papers when using these measures in research.
Every weighted dataset exports its full dictionary with raw and normalized weights, so the lexicon behind the measure can be audited term by term. See weighted dictionary artifacts.
The political-risk and country-risk measure families behind these dictionaries are also available as maintained datasets in Curated Measures.
Interpretation caveats:
- Weighted values are not comparable to count measures or across different dictionaries.
- Weighted datasets cannot be combined or compared with other datasets.
- The
metric_kind,weighting_source, andweighting_normalizationcolumns in the export identify weighted output and the normalization used.
What the measures do not mean
The core measures are raw counts. Do not interpret them as percentages, probabilities, classifier labels, causal estimates, forecasts, investment recommendations, or complete measures of real-world risk.
Exposure measures how much a call discusses the query topic. It does not prove the topic is important outside the selected corpus and query design.
Risk measures topic-linked risk discussion. It is not a compliance determination or a full risk assessment.
Sentiment is conditional on topic-matched sentences. It is not a standalone positive or negative classification of the whole call.
Filter caveat
The main Risk Tool counts and the Snippet Tool's sentiment measure filters do not use identical sentiment logic.
In the main Risk Tool counts, a mixed positive-and-negative sentence counts toward both positive and negative.
The Snippet Tool's sentiment measure filters are stricter: the positive filter shows positive-only sentences, the negative filter shows negative-only sentences, and the sentiment filter shows sentences that are one-sided in either direction. Mixed positive-and-negative sentences are excluded by those filters, so the sentences visible under a sentiment filter can be fewer than the corresponding counts suggest.