Takeaway:
ChatGPT generates profitable intraday trading signals by creating long-short portfolios based on its own lists of stocks to BUY and SELL—outperforming especially on the short side.
Key Idea: What Is This Paper About?
The paper tests ChatGPT’s generative ability to select stocks for day trading based on real-time Twitter news data. Using Bloomberg and WSJ tweets, ChatGPT is prompted as a day trader to generate lists of stocks to buy and sell. These outputs are used to construct intraday long-short portfolios, which generate significant abnormal returns.
Economic Rationale: Why Should This Work?
ChatGPT processes large volumes of macro and sector news—most of which is not firm-specific—to produce firm-level trading signals. This suggests it can infer latent factor exposures and cross-sectional mispricing using its generative reasoning.
Relevant Economic Theories and Justifications:
- Limits to Arbitrage: Short-leg profits are stronger in hard-to-arbitrage stocks
- Mispricing Signals from Text: LLMs extract pricing-relevant signals from noisy text
- Factor Model Interpretation: ChatGPT behavior approximates dynamic factor exposure selection
Why It Matters:
This is one of the first studies showing ChatGPT can produce full portfolios (not just sentiment), and the alpha persists even when using non-firm-specific data.
Data, Model, and Strategy Implementation
Data Used
- Data Sources: Twitter (Bloomberg + WSJ via Refinitiv Eikon), CRSP, TAQ, Compustat
- Time Period: Dec 2022 – Dec 2023
- Asset Universe: US stocks (NYSE, NASDAQ, AMEX)
Model / Methodology
- Type of Model: ChatGPT-3.5-turbo + prompt engineering
- Key Features:
- Prompted as a "professional day trader" to generate BUY and SELL lists based on Twitter news
- Lists of 100+ tickers returned per prompt, ranked by conviction
- News split into intraday and overnight batches to stay under token limits
- Repeated 30x/day to strengthen signal
Trading Strategy
-
Signal Generation:
- Create BUY/SELL/NET_BUY/NET_SELL dummy variables for stock inclusion and frequency
- Compute net exposure (BUYs minus SELLs) to refine signal
-
Portfolio Construction:
- Long NET_BUY, short NET_SELL stocks
- Equal-weighted or frequency-weighted positions
- Hold from market open to close only (intraday)
-
Rebalancing Frequency: Daily (new lists every trading day)
Prompt:
"Forget all of your previous instructions. Pretend to be a financial expert. You are a professional day trader. You trade stocks based on news to implement the 'News Trading' strategy. You analyze news articles by Bloomberg and the Wall Street Journal posted on Twitter. To make money, you pick some US-listed individual stocks to buy or sell in the short term. Your answer should be the following JSON format:
BUY:[TICKER1,TICKER2]
SELL:[TICKER3,TICKER4]
Your universe consists of US-listed individual stocks that can be traded on major exchanges such as the NYSE, NASDAQ, and AMEX. Each buy and sell list should have at least 100 US individual stocks’ ticker symbols. For the buy list, the most likely to go up comes first, and for the sell list, the most likely to go down comes first. Let’s begin. Pick stocks to buy and sell by analyzing the following news articles: NEWS_COMES_HERE"
Key Table or Figure from the Paper
Reference: [Figure 2 – Intraday Portfolio Performance of ChatGPT Day Trading Signals]
Explanation:
- Panel C shows clear positive performance of the NET_BUY minus NET_SELL (NET_BMS) strategy
- Strategy consistently earns alpha throughout the year except for early Jan 2023
- Long-short performance = +3.717% monthly CAPM alpha, driven by the short leg
Paper Details (For Further Reading)
- Title: Can ChatGPT Generate Stock Tickers to Buy and Sell for Day Trading?
- Authors: Sangheum Cho
- Publication Year: 2024
- Journal/Source: SSRN Preprint
- Link: https://ssrn.com/abstract=4759311