Solving ChatGPT’s “Garbage In, Garbage Out” Problem

February 28, 2023

ChatGPT is having a moment, but it isn’t meeting everyone’s expectations.  

The AI chatbot, predicated on a neural network machine learning model, seems ubiquitous. It has essentially reached household status, yet most are unfamiliar with how large language models such as ChatGPT actually work. 

As a result of this confusion, many companies have tried to benefit from ChatGPT’s human-like conversational abilities but struggled to get consistently useful output. In some cases, companies have even banned the use of ChatGPT due to risks posed by inaccurate output.

This is a classic “garbage in, garbage out” problem. 

Overcoming Weaknesses in Language Models

Large language models, like GPT and BERT, are fantastic for generalist use cases but tend to struggle when it comes to understanding the nuances of domain-specific language, such as that used in finance. In other words, ChatGPT is a blunt instrument, but it can be made much more precise. 

ChatGPT is trained on a large dataset of general conversational text and thus excels at producing conversational, “human-sounding” output. In order to get specific results that are more suited to technical applications, it needs to be trained on domain-specific language and fed timely, well curated data at the input stage. And that is our current approach at MarketReader.

In order to generate succinct, useful explanations of why asset prices are moving in the market, we are sending ChatGPT the combined output of our proprietary fundamental models, which analyze potential volatility drivers ranging from calendar events to social media posts. 

This approach plays to ChatGPT’s strengths by allowing it to produce conversational, easily digestible explanations of price moves underpinned by our system’s highly technical model output coupled with carefully curated news and social media content. These customized and succinct summaries generated by ChatGPT have already proven themselves to be extremely high-quality. 

Automated Market Explanations

After receiving input from MarketReader, ChatGPT generated a single paragraph summarizing why Walt Disney Co ($DIS) stock was moving after-hours between February 8th and 9th, after the company had released its quarterly earnings at the close of the previous trading day:

In a more recent example, our system summarized the reason behind the unusual volatility of Tegna Inc ($TGNA) stock on February 27th:

“Tegna Inc has come under scrutiny from the Federal Communications Commission, leading to a delay in Standard General’s proposed $5.4 billion takeover of Tegna and an administrative law judge hearing on material concerns surrounding potential price hikes for consumers or job losses…”

This explanation is possible only because MarketReader sent ChatGPT a combination of heavily filtered, high-frequency output from our specialized machine learning modeling and data processing system.

Redefining Financial News Coverage

Enterprise clients are already noticing the power of this combination of high quality bespoke input + conversational output. We expect AI-assisted MarketReader insights to become the industry standard for financial news coverage, assisting millions of users around the world to better understand why the market is moving in real time.

Sign up for our Insights & Updates list to receive MarketReader news to your inbox, or enroll in our Early Access Community to get the first look at our upcoming product.

Get Email Updates

Sign up to receive MarketReader news and technical + financial insights from our expert team.

If you are interested in investing in MarketReader Inc, please reach out to ir@marketreader.com.