StockNet dataset

StockNet’s defined task is binary next-day movement prediction — Rise or Fall — from text plus price history, and the repository ships both halves aligned. ./price holds raw Yahoo Finance CSVs (Date,Open,High,Low,Close,Adj Close,Volume) and a preprocessed form giving date, movement percent and normalized open, high, low, close and volume; ./tweet holds raw Twitter JSON and a preprocessed form reduced to text, user_id_str and created_at. The universe is all 8 stocks in the Conglomerates sector plus the top 10 by capital size in each of the other 8 sectors. Price data covers all 88 tickers, but tweet data covers only 87 — both tweet/raw and tweet/preprocessed have 87 stock subdirectories against 88 price files.

The README’s 2014-01-01 to 2016-01-01 window describes the paper’s experimental split rather than the file contents: the shipped raw price CSVs run wider (AAPL spans 2012-09-04 to 2017-09-01) and AAPL’s raw tweets continue to 2016-03-31. There is no installer — a git clone or direct raw fetches over HTTPS, no authentication. The derived task is also mirrored on Hugging Face as TheFinAI/flare-sm-acl, which reports 27,056 rows; PIXIU’s own README lists the same ACL18 stock-movement task at 27,053 samples. The Hugging Face mirror states no license.

The repository is dormant — the last commit to master is dated 2018-11-02 and it is not archived — which is expected for a frozen benchmark corpus but means the data will not be extended. The MIT LICENSE (Copyright (c) 2018 Yumo Xu) is the authors’ grant over their compiled collection; the repo redistributes full raw Twitter JSON and carries no separate data-terms file addressing platform terms for that content. The accompanying paper is Stock Movement Prediction from Tweets and Historical Prices, ACL 2018, pages 1970–1979 (P18-1183).