*** Mannheim, Germany on 24 September 2018 (updated on 13 December 2019) ***

This is the supplementary data for our 2018 ECONLP paper "Word Embeddings-Based Uncertainty Detection in Financial Disclosures".

The folder "data/" contains the two data sets "regression.csv" and "classification.csv" which are used in the paper:
- The regression dataset features the following columns:
	- file: the filename that can be used to retrieve the document from SEC EDGAR
	- date: filing date in YYYYMMDD format
	- year: year dummy
	- ff48: the FF48 industry dummy
	- rmse_post: future volatility (i.e. the dependent variable)
	- alpha: historic performance
	- rmse_prior: historic volatility
	- ret_abn: filing abnormal return
	- mkt_cap: market capitalization
	- btm: book-to-market ratio
	- nasdaq: NASDAQ dummy
	- unc: cummulative tf-idf according to the UNC dictionary
- The first column of the classification dataset contains the sentences and the second column contains the label (c: "certain", u: "uncertain")

The folder "dicts/" contains four TXT files storing the newly added terms according to our four dictionary expansions: 
- UncGen
- UncGen_exp
- UncSpec
- UncSpec_exp
The lists have to be merged with the 297-term uncertainty dictionary published by Loughran & McDonald (2011).

The folder "docu/" contains five PDF files:
- "classification_results" presents the full classification results according to both tf and tf-idf normalization and all ML algorithms
- "examples" contains randomly sampled terms from Unc together with their 20 similarity candidates according to the general and the specific embedding model
- "guidelines" contains our annotation guidelines
- "parsing" describes our data screening and parsing procedure in greater detail
- "regression_results" presents the full regression results with all control variables