---
title: README
emoji: ⚡
colorFrom: red
colorTo: blue
sdk: static
pinned: false
---

# Extreme Weather Event Impacts: Data and Models

This is the model and data description for the paper "[What Firms Actually Lose (and Gain) from Extreme Weather Event Impacts](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6035794)".

## Data
The simplest form of the project is the 13,277 firm-event impacts. This is created by analyzing over 1.7 million filings (3.5 billion paragraphs) of all publicly listed firms in the US. We analyse filing types: event-based Form 8-K, quarterly reported Form 10-Q, and annually reported Form 10-K. We upload all paragraphs that contain at least a mention of an extreme weather event / physical risk in a more fine-grained dataset.

### Event-Impact Data
This is the final data, including the 13,277 firm-event impacts (**name: [event_impact_data](https://huggingface.co/datasets/extreme-weather-impacts/event_impact_data)**). It has the following structure:
- event_id: ID of the event in the NOAA billion-dollar dataset
- cik: unique company identifier [If this combination is present, then there was a firm-event impact]
- asset, economic_flows, none: classification dimensions of the impact channel classifier; shows whether at least one document indicated an asset impact (asset==1), a pure economic flow impact, i.e. at least one document was only addressing an impact through economic flows and NOT through assets as well (economic_flows==1), or none if at least one document mentioned a none impact (neither asset, nor economic flows; none==1)
- neutral, negative, positive, reimbursement: classification dimensions of the impact dimensionality classifier; shows whether at least one document indicated a neutral, negative, or positive impact, or a reimbursement 
- name: event name according to NOAA
- event_type: disaster event type according to NOAA
- begin_date: event begin date according to NOAA
- cpi_adjusted_cost: CPI-adjusted cost of the event according to NOAA

As a complementary dataset, we also upload the NOAA extreme weather event dataset (**name: [NOAA_event_with_summary](https://huggingface.co/datasets/extreme-weather-impacts/NOAA_event_with_summary)**), [obtained here](https://www.ncei.noaa.gov/access/billions/):
- Name: event name according to NOAA
- Disaster: disaster event type according to NOAA
- Begin Date: event begin date according to NOAA
- End Date: event end date according to NOAA
- CPI-Adjusted Cost: CPI-adjusted cost of the event according to NOAA
- Unadjusted Cost: non-adjusted cost of the event according to NOAA
- Deaths: deaths caused by the event according to NOAA
- Event Duration: event duration according to NOAA
- Event ID: ID of the event in the NOAA billion-dollar dataset
- Summary: event description / summary provided by NOAA

### File-based Data
We classify over 1.7 million files and 3.5 paragraphs with our models. We analyse event-based Form 8-K, quarterly reported Form 10-Q, and annually reported Form 10-K. We upload paragraphs per file classified at least as extreme weather relevant. If a file didn't contain at least one extreme-weather-related paragraph, we leave the "paragraph" column empty (**names: [classified_data_8K](https://huggingface.co/datasets/extreme-weather-impacts/classified_data_8K), [classified_data_10Q](https://huggingface.co/datasets/extreme-weather-impacts/classified_data_10Q), [classified_data_10K](https://huggingface.co/datasets/extreme-weather-impacts/classified_data_10K)**):
- path: filepath and unique identifier of the file
- company: unique company identifier
- year: year of the filing
- filename: filename of the 8-K, 10-K, 10-Q given by us in the downloading process
- date: date of the filing
- paragraph: text paragraph that is analyzed with the LLMs
- num_paragraphs: number of paragraphs that the entire filing contained
- num_words: number of words that the entire filing contained
- Storm, Flood, Heatwave, Drought, Wildfire, Coldwave, physical risk: Indicates 1 if the text mentions any of the above categories
- item1a, item7, item8: indicates 1 if the text was in the item mentioned; Form 8-K filings contain item7 and item8, Form 10-K filings contain item1a (not used in the analysis)
- impact: indicates “Yes” if the text mentions that the company was impacted to any extreme weather event
- impact_YY: indicates “Yes” if the text mentions that the company was impacted by the specific extreme weather event with the event_id YY in the NOAA data
- impact_directionality: indicates “Neutral”, “Negative”, “Positive”, or “Reimbursement” according to the impact directionality classification
- impact_channel: indicates “asset”, “only_economic_flows”, “none” according to the impact channel classification;  “only_economic_flows” means that there was no asset impact co-detected with economic flows

## Models
We upload all models, and training data in this repository. Model usage is described in the corresponding model pages.

## Repository under construction
The repository is currently being constructed. If you have any questions, please reach out to tobias.schimanski@df.uzh.ch.

## Paper
If you make use of the data, please cite [the paper](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6035794):

```shell
@article{schimanski2026extremeweatherimpacts,
  title        = {What Firms Actually Lose (and Gain) from Extreme Weather Event Impacts},
  author       = {Schimanski, Tobias and Gostlow, Glen and Toetzke, Malte and Leippold, Markus},
  url          = {https://ssrn.com/abstract=6035794},
  doi          = {10.2139/ssrn.6035794}
}
```