Large Language Models (LLMs) have become increasingly relevant for financial clients. For example, Bloomberg has all the necessary resources and data to create custom-tailored models like ChatGPT. Unlike ChatGPT models trained on general knowledge from books and Wikipedia, custom-tailored models are trained on real-world financial data, enabling them to help clients navigate their company data and solve a wide range of financial problems.
For instance, a custom-tailored ChatGPT can be integrated into a company's marketing department, allowing on-the-fly drafting of marketing materials for clients. It can also be connected to a company's analytics process to provide valuable insights.
To learn more about the use of LLMs and models like ChatGPT in the finance industry, we spoke with DataArt's AI/ML expert, Yury Zaryadov. During our conversation, we explored how these models can benefit financial companies and their impact on the industry as a whole.
Q&A with Yury Zaryadov
Q: Some general materials about ChatGPT models suggest that these models can generate financial headlines and analyze financial vocabulary and features. Can these models provide insights on which shares to buy or sell?
Yes, it is possible to do so with a Large Language Model trained on real-world financial data. By incorporating smart formulas that use initial input data from the market, such as prices, the LLM can generate new figures as indicators for deciding whether to buy or sell shares. If the LLM understands how these formulas work, it can perform these calculations on the fly without changes to the code.
Moreover, the LLM can assist with data collection and analysis tasks. For example, if an analyst in a company is asked to gather data on a small company, they would need to collect data sheets and analyze the company’s history. Similarly, an LLM can interact with unstructured content and perform tasks that require analyzing a large amount of data.
Overall, the ability of LLMs to perform complex financial calculations and analyze unstructured data makes them valuable tools for financial companies seeking to improve their decision-making processes.
Q: Can you identify areas or processes in financial corporations where a Large Language Model could be applied to facilitate these processes or eliminate them somehow?
Financial corporations of any size often deal with unstructured content such as documents, an area where the Large Language Model can be extremely useful. The model bridges the system and the user, enabling queries to be processed easily in a natural language format.
For example, if you need to generate reports or perform analytics, the Large Language Model can be used as a layer to perform these tasks. Working with unstructured content is a complex and time-consuming process, and people often hire separate companies to process documents manually. Moreover, people who process these documents require a financial background to understand the data.
A user interface (UI) with windows, tables, and other features is typically used to interact with the Large Language Model. A system like Copilot, which is integrated into all office programs, is a potential option. Some financial corporations still prefer to use terminals that display live data in real-time, but there is a growing trend toward using a copilot-style interface.
It's worth noting that implementing a Large Language Model doesn't necessarily mean that financial corporations will replace the human analysts. Instead, the tool can benefit analysts by helping them work more efficiently and quickly while reducing cost per hour. Analysts can perform tasks faster and more efficiently than without using the model.
Q: Can such a model trade on its own?
No, a model like ChatGPT cannot function as a standalone trader. ChatGPT is a generative model that can continue text based on a given prompt. It is an advanced autocomplete tool. Therefore, ChatGPT requires human input to provide context and decision-making capabilities to generate responses.
Q: Is it possible that in the future, ChatGPT could become so advanced that it won’t require any additional data sets for analysis to generate its responses independently?
This would require the development of artificial general intelligence (AGI), which would involve creating a conscious machine that can react to various external stimuli. While it is theoretically possible, it is currently beyond the scope of ChatGPT's capabilities.
Q: So, the system can predict or offer potential events based on the input data, depending on the query sent to it. But is there a level of analytics and immersion, like what Bloomberg GPT offers, that could predict cases like SVB and Credit Suisse?
If a model is made to predict these kinds of events, it is designed to predict big market events and the horizon of these events. However, if such a model exists, it will generate events built on the assumption that it exists.
For instance, high-frequency trading was invented based on the idea of reacting to market inertia. However, news spreads at a certain speed, and you can catch a trend of "it's going down, we have to play on it," or "it's going up, we have to play on it." If you can react to it before someone does, you can do well. But having such a system automatically leads to the fact that there are many systems, and these systems start trading with each other. They affect each other differently, causing a chain reaction effect. Therefore, assuming whether Bloomberg GPT can predict something like that is uncertain. It could be, but not in its current form.
It may help identify signals for someone who has already analyzed the data and made a conclusion about an upcoming crisis at SVB or Credit Suisse. Thus, it can work as a tool, but it is unlikely to give alerts about the urgent need to sell SVB.
Q: What is DataArt currently doing in the area of Large Language Models?
We are focusing more on the application of LLMs rather than the models themselves. While many models are already available, we specialize in creating custom models and developing processes to train and tune them properly. Our approach involves utilizing Machine Learning Operations (MLOps) to ensure the best results.
However, working with LLMs directly can be challenging due to the enormous amount of data required for training. Companies often crowdsource data for LLM training, but it needs to be processed in a specific way to prepare it for the custom form on which the LLM will learn. This is one of the main tasks we face.
We are also working with more specific datasets, such as financial LLMs, that are trained in financial data and understand the related terminology. As new datasets and LLMs emerge, we are well-positioned to integrate them.
Q: Apart from the need for a large amount of data, what other challenges do clients face when working with Large Language Models? Is there a significant storage requirement for all the data?
Firstly, deploying an LLM typically only involves storing part of the dataset unless you're a large company like Bloomberg.
Secondly, capacity is indeed a consideration. It also depends on the architecture of the model itself.
Creating an LLM is not a straightforward process, and it's more complex than starting from scratch. It requires specific skills and expertise. Currently, it doesn't work in an automated mode where you input a few commands, and the model instantly knows all your documents and information. It still involves a data science process. Developing such a model requires careful handling and a certain level of skills to navigate the complexities involved.
Q: What does the concept of architecture mean in the context of such a model?
The architecture model refers to the underlying mechanism or algorithm used to process the input. Programs always operate based on functions. In the case of language models, the input can be text, an image, or a combination of both. Generative AI models now combine different modalities to gather more contextual information. The architecture defines how the data is processed within the model, the order in which operations are performed, and how the information is combined. Creating and working with such models is a highly complex and computationally demanding task that requires a semi-scientific approach.
Q: So, if I understand correctly, even with the same dataset, the results can vary depending on how the architecture is set up?
Yes, that's correct. If you don't choose the right model, it can auto-compile and generate results differently. The models themselves can vary in size, meaning they have a different number of parameters. For example, GPT-3 has around three and a half billion parameters, while GPT-4 may have trillions of parameters. The size of the model affects the amount of memory it requires and also complicates its internal structure. Therefore, the choice of architecture plays a crucial role in determining the output and behavior of the model, even with the same dataset.
Q: What are some use cases where these models can be applied?
These models have many applications, particularly in word processing and document work. They can also be used for various types of translation, converting syntax from one form to another. A notable example is text-to-SQL conversion, where the model can understand the context and generate SQL queries that can be executed to retrieve information from a database. By leveraging the model's capabilities, text-to-SQL solutions built on GPT are significantly faster than traditional methods.
Another important application is efficient document parsing. We have explored different ways to integrate GPT into various systems and workflows, finding ways to utilize its capabilities effectively.
However, there are still areas we haven't explored but are eager to try, such as creating custom Large Language Models. We are looking forward to exploring and harnessing this exciting opportunity.
Conclusion
In conclusion, Yury's insights shared in this discussion highlight the transformative potential of Large Language Models (LLMs) in the finance industry. LLMs offer the capacity to enhance financial operations, streamline data processing, and augment decision-making processes. However, the finance sector must address security, compliance, and technical integration challenges to harness the power of Generative AI fully.
To delve deeper into these topics, we encourage you to watch the webinar "Beyond the Hype: Is Generative AI Actually Ready for Finance?" where experts delve into the practical applications and considerations of Generative AI in finance.
