Takeaway:
ChatGPT3.5-derived sentiment from Twitter outperforms other sentiment models (BERT, VADER) in predicting Bitcoin returns. This introduces a new alpha source for crypto traders.
Key Idea: What Is This Paper About?
The paper proposes a novel framework that uses ChatGPT3.5 to evaluate sentiment in Twitter posts by key crypto influencers. These sentiment labels ("Bullish", "Neutral", "Bearish") are turned into indicators that significantly explain Bitcoin returns. Compared to other NLP models like BERT or rule-based systems like VADER, ChatGPT consistently produces more predictive signals.
Economic Rationale: Why Should This Work?
Bitcoin lacks standard valuation models, making it more sentiment-driven than traditional assets. Twitter is a major information channel for the crypto market, and ChatGPT’s ability to extract nuanced sentiment gives it an edge over fixed dictionaries or linear models.
Relevant Economic Theories and Justifications:
- Noise Trader Theory (Black, 1986): Investors act on sentiment, not fundamentals.
- Behavioral Finance: Emotion and narrative drive crypto trading more than fundamentals.
- Information Frictions: Most sentiment models fail to capture nuance; ChatGPT overcomes this.
Why It Matters:
Crypto traders can use LLMs like ChatGPT to extract predictive sentiment signals directly from social media—without coding or training complex models.
Data, Model, and Strategy Implementation
Data Used (If Applicable)
- Data Sources:
- Twitter API (24,196 tweets from top crypto influencers)
- Bitcoin prices (Bitstamp via Refinitiv Eikon)
- Crypto-specific and macro sentiment indices (VIX, BW, F&G, etc.)
- Time Period: Jan 2018 – Jun 2023
- Asset Universe: Bitcoin
Model / Methodology (If Applicable)
- Type of Model: Machine Learning (LLM + Linear Regression)
- Key Features:
- Sentiment classified by ChatGPT into Bearish / Neutral / Bullish
- Tweets labeled using zero-shot classification
- Indicators built: Bullishness Index (BI), Variation (VA), Agreement (AG)
Prompt Used in the Paper:
“Think from the point of view from Bitcoin investors. You are reading tweets from Twitter and want to decide whether you want to invest (buy, sell, or hold) your Bitcoin. Can you help me to identify the following tweets from Twitter by categorizing those tweets into one of the 3 groups ‘Bearish’, ‘Neutral’, ‘Bullish’? Just give me the total numbers in each category, you don’t have to show the tweet again. You should be able to recognize each tweet because the tweets will be within the quotation mark (‘’) and after each tweet there will be a semi colon (;). Type ‘I got it’ if you are clear about my request.”
Key Equations:
- Bullishness Index:
BIₜ = ln((#Bullₜ + 1) / (#Bearₜ + 1)) - Agreement Index:
AGₜ = 1 − sqrt(1 − ((#Bullₜ − #Bearₜ) / (#Bullₜ + #Bearₜ))²) - Regression Model:
BitcoinReturnₜ = β₀ + β₁·Sentimentₜ + γ·Controlsₜ + εₜ
Trading Strategy (If Applicable)
-
Signal Generation:
- Use ChatGPT to classify daily tweets as Bullish/Bearish
- Construct sentiment scores (BI, VA, AG)
- Buy BTC on bullish signal (e.g., BI↑), reduce or short on bearish
-
Portfolio Construction:
- Long-only Bitcoin position
- Optional: risk-adjusted with stop-loss or BTC volatility control
-
Rebalancing Frequency:
- Daily or weekly depending on signal stability
Key Table or Figure from the Paper
Reference: [Table 3 – ChatGPT Sentiment and Bitcoin Return]
Explanation:
- Shows that ChatGPT’s Bullishness Index (BIₜ), Variation (VAₜ), and Agreement (AGₜ) each significantly predict Bitcoin returns.
- Coefficients are statistically significant at 1–5% level, even when controlling for market variables.
- Models using BERT and VADER do not show significance—highlighting ChatGPT’s edge.
Final Thought
💡 ChatGPT turns crypto Twitter noise into alpha. It beats BERT, VADER, and even traditional sentiment proxies. 🧠📈
Paper Details (For Further Reading)
- Title: ChatGPT, Twitter Sentiment and Bitcoin Return
- Authors: Binh Nguyen Thanh, Anh Nguyen Tuan, Tuan-Thanh Chu, Son Ha Xuan
- Publication Year: 2023
- Journal/Source: SSRN Preprint
- Link: https://ssrn.com/abstract=4628097