The initial wave of scholarly research using ChatGPT in finance is emerging, and early outcomes indicate the preceding months’ excitement is well-founded. This month, two novel papers were published that utilized the AI chatbot for market-related tasks: interpreting Federal Reserve statements as hawkish or dovish and determining if headlines were favorable or unfavorable for a stock. ChatGPT excelled in both tasks, implying a significant advancement in employing technology to transform vast textual content, from news articles to tweets and speeches, into trading signals.
While language models have long informed Wall Street strategies, recent discoveries reveal OpenAI’s technology reaching new heights in understanding subtlety and context. Slavi Marinov, Man AHL’s head of machine learning, confirms the hype is legitimate, as the company has used natural language processing (NLP) to analyze texts like earnings transcripts and Reddit posts for years.
In the first paper, researchers from the Fed discovered ChatGPT closely mirrored human performance in deciphering the central bank’s statements. The AI outperformed a prevalent Google model called BERT and dictionary-based classifications. ChatGPT even demonstrated an ability to explain its classifications, resembling the central bank’s own analyst.
The second study tasked ChatGPT with acting as a financial expert and interpreting corporate news headlines. The researchers found a statistical connection between ChatGPT’s responses and the stock’s subsequent movements, suggesting the technology correctly interpreted the news’ implications.
Although many quants already use NLP to assess a stock’s popularity on Twitter or incorporate recent company headlines, ChatGPT’s advancements appear to unlock new information sources and broaden accessibility for finance professionals. Marinov believes ChatGPT can accelerate the entire process.
Man AHL initially built models by manually labeling sentences as positive or negative for an asset, providing a blueprint for machines to interpret language. They gamified the process, ranking participants and calculating agreement levels to involve all employees. The two new papers indicate that ChatGPT can accomplish similar tasks without explicit training. This zero-shot learning surpasses previous technologies, but fine-tuning it with specific examples improves it even further. Marinov suggests complementing manual data labeling with designing suitable prompts for ChatGPT.