PriceSeer Dataset
Introduction
PriceSeer was designed to provide a professional framework for LLMs stock prediction, including external and internal data. For detailed information about how to collect it and how to use it, please refer to: https://arxiv.org/abs/2601.06088
Stock Coverage
This dataset collected totally 110 stocks across 11 sectors in the U.S. stock market, 10 stocks per sector, covers all industries in the market. The stocks are:
"Technology": ["NVDA", "MSFT", "AAPL", "AVGO", "ORCL", "PLTR", "AMD", "CSCO", "IBM", "CRM"],
"Financial Services": ["BRK-B", "JPM", "BAC", "MA", "V", "MS", "C", "WFC", "GS", "AXP"],
"Consumer Cyclical": ["AMZN", "TSLA", "HD", "MCD", "NKE", "LOW", "BKNG", "TJX", "DASH", "MELI"],
"Communication Services": ["GOOG", "META", "NFLX", "CMCSA", "DIS", "APP", "SPOT", "T", "VZ", "TMUS"],
"Healthcare": ["UNH", "JNJ", "ISRG", "MRK", "LLY", "ABT", "DHR", "ABBV", "TMO", "AMGN"],
"Industrials": ["UNP", "GEV", "CAT", "HON", "RTX", "DE", "LMT", "GE", "ETN", "BA"],
"Consumer Defensive": ["PG", "KO", "PEP", "COST", "WMT", "MDLZ", "CL", "MO", "PM", "MNST"],
"Energy": ["XOM", "CVX", "COP", "WMB", "EPD", "EOG", "PSX", "KMI", "ET", "MPC"],
"Basic Materials": ["LIN", "SHW", "FCX", "SCCO", "NEM", "CRH", "APD", "ECL", "CTVA", "VMC"],
"Utilities": ["NEE", "DUK", "SO", "EXC", "AEP", "SRE", "XEL", "VST", "D", "CEG"],
"Real Estate": ["WELL", "PLD", "AMT", "EQIX", "SPG", "DLR", "O", "PSA", "CBRE", "CCI"]
Data Type
There are 4 data types in the dataset:
Raw data: the historical stock data in the past 1 year were directly downloaded from open source financial website, including opening price, closing price, daily highest price, daily lowest price, and trading volume.
Financial indicators: professional financial indicators that can reflect the price trend are calculated based on the raw data. The indicators used are: Simple Return & Log Return, Simple Moving Average (SMA), Relative Strength Index (RSI), Moving Average Convergence/Divergence (MACD) & Signal Line, and Bollinger Bands (BB).
News data: top 10 the most relavant news on Google research about each stock were downloaded. It consists of the publish date, source, news title, and content.
Fake news data: we generate fake news by manipulating numbers, introducing fictional elements, and employing superlative language in the news data, to introduce perturbations.
Citation
If you find our work interesting, please feel free to cite our dataset:
@misc{liang2025priceseerevaluatinglargelanguage,
title={PriceSeer: Evaluating Large Language Models in Real-Time Stock Prediction},
author={Bohan Liang and Zijian Chen and Qi Jia and Kaiwei Zhang and Kaiyuan Ji and Guangtao Zhai},
year={2025},
eprint={2601.06088},
archivePrefix={arXiv},
primaryClass={q-fin.ST},
url={https://arxiv.org/abs/2601.06088},
}