Stock Movement Prediction from Tweets and Historical Prices (StockNet)

The model is a deep generative architecture that exploits text and price signals jointly, built on recurrent continuous latent variables with neural variational inference and a hybrid objective with a temporal auxiliary. The authors collected the dataset it is evaluated on: two years of price movements, 1 January 2014 to 1 January 2016, for 88 stocks — all eight in the Conglomerates sector plus the ten largest by capital size in each of eight other sectors — with tweets from Twitter and prices from Yahoo Finance. Coverage is asymmetric: 88 tickers have preprocessed price files but only 87 have tweet directories, GMRE being the gap.

Both repositories have been dormant since November 2018. The reference implementation lives separately in yumoxu/stocknet-code, which selects among the paper’s four variants through a config file but targets Python 2.7.11 and TensorFlow 1.4.0 — a historical artifact rather than runnable code. The data clones directly under MIT, though the tweet content it redistributes remains subject to Twitter/X terms.

The dataset outlived the model. It is indexed as the ACL18 stock-movement task in PIXIU/FLARE and in FinBen, and mirrored on Hugging Face as TheFinAI/flare-sm-acl, where the task is framed as predicting “Rise” or “Fall” from price history plus social-media context; that mirror’s card declares no license of its own. Sample counts differ between those sources — PIXIU’s README lists 27,053, the Hugging Face viewer 27,056 rows — so cite whichever with its source. The paper itself is open access on the ACL Anthology under CC BY 4.0.