CoARiJ: Corpus of Annual Reports in Japan

We organized Japanese financial reports to encourage applying NLP techniques to financial analytics.

Dataset

The corpora are separated to each financial years.

master version.

fiscal_year	Raw file version (F)	Text extracted version (E)
2014	.zip (9.3GB)	.zip (269.9MB)
2015	.zip (9.8GB)	.zip (291.1MB)
2016	.zip (10.2GB)	.zip (334.7MB)
2017	.zip (9.1GB)	.zip (309.4MB)
2018	.zip (10.5GB)	.zip (260.9MB)

• financial data is from 決算短信情報.
  • We use non-cosolidated data if it exist.
• stock data is from 月間相場表（内国株式）.
  • close is fiscal period end and open is 1 year before of it.

Past release

• v1.0

Statistics

fiscal_year	number_of_reports	has_csr_reports	has_financial_data	has_stock_data
2014	3,724	92	3,583	3,595
2015	3,870	96	3,725	3,751
2016	4,066	97	3,924	3,941
2017	3,578	89	3,441	3,472
2018	3,513	70	2,893	3,413

File structure

Raw file version (--kind F)

The structure of dataset is following.

chakki_esg_financial_{year}.zip
└──{year}
     ├── documents.csv
     └── docs/

docs includes XBRL and PDF file.

• XBRL file of annual reports (files are retrieved from EDINET).
• PDF file of CSR reports (additional content).

documents.csv has metadata like following. Please refer the detail at Wiki.

• edinet_code: E0000X
• filer_name: XXX株式会社
• fiscal_year: 201X
• fiscal_period: FY
• doc_path: docs/S000000X.xbrl
• csr_path: docs/E0000X_201X_JP_36.pdf

Text extracted version (--kind E)

Text extracted version includes txt files that match each part of an annual report.
The extracted parts are defined at xbrr.

chakki_esg_financial_{year}_extracted.zip
└──{year}
     ├── documents.csv
     └── docs/

Tool

You can download dataset by command line tool.

pip install coarij

Please refer the usage by -- (using fire).

coarij --

Example command.

# Download raw file version dataset of 2014.
coarij download --kind F --year 2014

# Extract business.overview_of_result part of TIS.Inc (sec code=3626).
coarij extract business.overview_of_result --sec_code 3626

# Tokenize text by Janome (`janome` or `sudachi` is supported).
pip install janome
coarij tokenize --tokenizer janome

# Show tokenized result (words are separated by \t).
head -n 5 data/processed/2014/docs/S100552V_business_overview_of_result_tokenized.txt
1       【      業績    等      の      概要    】
(       1       )               業績
当      連結    会計    年度    における        我が国  経済    は      、     消費    税率    引上げ  に      伴う    駆け込み        需要    の      反動   や      海外    景気    動向    に対する        先行き  懸念    等      から   弱い    動き    も      見      られ    まし    た      が      、      企業   収益    の      改善    等      により  全体  ...

If you want to download latest dataset, please specify --version master when download the data.

• About the parsable part, please refer the xbrr.

You can use Ledger to select your necessary file from overall CoARiJ dataset.

from coarij.storage import Storage


storage = Storage("your/data/directory")
ledger = storage.get_ledger()
collected = ledger.collect(edinet_code="E00021")