The measures follow the sentence-level construction documented by NL Analytics and used for the corresponding topic-specific Risk measures in Hassan et al. (2025).
1. Define the topic query
Each measure has a documented query containing keywords and phrases associated with its topic.
For a given topic k, define:
> Topic match(s, k) = 1 when sentence s contains at least one term matching topic query k.
Otherwise:
> Topic match(s, k) = 0
The complete query for each measure is available on its individual page.
2. Identify risk-related sentences
The Risk dictionary is the deduplicated union of Oxford Thesaurus synonyms for:
- risk
- risky
- uncertain
- uncertainty
The terms question, questions, and venture are excluded.
For these measures, the topic match and Risk-dictionary language must occur in the same sentence.
Define:
> Risk match(s) = 1 when sentence s contains at least one Risk-dictionary term.
Otherwise:
> Risk match(s) = 0
3. Calculate the call-level Risk measure
For earnings call c and topic k:
> Risk(c, k) = Σ Topic match(s, k) × Risk match(s)
> for all sentences s in call c
In plain language:
> Topic Risk is the number of sentences in an earnings call that contain at least one topic-query match and at least one risk or uncertainty synonym.
Each sentence is counted at most once for a given topic, even if it contains multiple query terms or multiple Risk-dictionary terms.
Overall Risk is calculated as:
> Overall Risk(c) = Σ Risk match(s)
> for all sentences s in call c
In plain language:
> Overall Risk is the number of sentences in an earnings call that contain at least one risk or uncertainty synonym, regardless of topic.
4. Calculate the JEP-style risk decomposition
For a selected firm, group of firms, or period:
> Topic share of risk (%) = 100 × Topic Risk ÷ Overall Risk
The numerator and denominator must be aggregated over the same calls and period.
This is the construction used for the cardinal risk decomposition in Figure 3 of Hassan et al. (2025). That figure compares AI, Brexit, COVID-19, Inflation, Russia, Supply Chain, and Trade.
Research considerations
The measures capture topic-linked risk discussion. They do not estimate the probability of an event, its objective severity, or its expected financial cost.
Results depend on the topic queries and selected sample. Researchers should inspect matched sentences, document the queries used, retain calls with zero matches, and account for changes in sample composition.
To reproduce the JEP figure, researchers must also match the paper’s sample, date range, aggregation, and smoothing choices.
The maintained dataset extends the measures to currently available data. It should therefore be described as a maintained implementation of the paper’s measures, not as an exact replication of the published figure.
Academic foundation
All seven measures in this curated dataset are used in the cardinal risk decomposition in:
Tarek A. Hassan, Stephan Hollander, Aakash Kalyani, Laurence van Lent, Markus Schwedeler, and Ahmed Tahoun. “Text as Data in Economic Analysis.” Journal of Economic Perspectives 39(3), 2025, pp. 193–220.
Figure 3 of the paper compares AI, Brexit, COVID-19, Inflation, Russia, Supply Chain, and Trade as shares of overall corporate risk discussion.
For exact replication of the published results, use the paper’s replication package and follow its sample and aggregation choices.