Commentary: Data to Alpha — How large language models can reshape investment strategies

The financial services industry and just about every industry on Earth are set to undergo a series of major transformations brought about by the generative proliferation of artificial intelligence, and specifically large language models (or LLMs). Education, research and practically every data-driven workflow are susceptible to rapid change as the full flexibility and nature of this technological breakthrough are realized. Made possible by ever-increasing semiconductor prowess and open-sourced efforts, LLMs have the potential to transform global markets with significantly enhanced efficiency and access.

Indeed, such breakthroughs will provide all institutions with the ability to expeditiously obtain sophisticated real-time data analysis without prohibitive costs, thereby leveling the playing field between larger and smaller institutions globally. Imagine having access to automated transcript analysis that provides real-time sentiment and keyword reporting of all publicly reporting entities, or having the ability to decipher predictive relationships of gross margin progression across every publicly traded asset globally at a price entry point that is no longer only the domain of the mega institutions.

So, what are the risks, and how should investors approach integration of LLMs into their investment processes?

Before continuing, it is important to highlight what an LLM is and how recent advancements differ from previous AI frameworks. LLMs are computerized models that utilize millions to hundreds of billions of layers of statistical connections that are trained on large computer networks. These layers, known as parameters, set statistical weights and connections for the language components contributed into the model. While humans can influence the importance of these inputs, many of the outward behaviors between LLM parameters are self-created by the contributed language data. Advancements in hardware have allowed these systems to grow exponentially in complexity and while we still don't understand exactly why, the ability for these large models to acquire implicit knowledge about syntax, semantics and ontology within a language is what differentiates LLMs from human-trained systems that came before.

Certainly, there are many naysayers who are quick to call out that LLMs can produce a number of data accuracy errors. I won't argue that there are hurdles to overcome regarding data confidence, or that accuracy is speckled with a touch of unexplained behaviors. However, while imperfect given their infancy, LLMs are good enough at present to have positive impacts in the short-term. And, compared to human endeavors, the reflections of imperfections are quite relatable and manageable.

Current implementations of LLMs, coupled with their open-source nature, have the potential to drive new strategies for real-time investing and analysis. As such, it is important that investors today focus on leveraging LLMs to increase efficiency and deploy small- to large-parameter LLMs as replacements for human capital in data analytics, reconciliation and accounting roles. In other words, investors should seek to leverage AI as a margin enhancement tool that drives lower-cost strategies and increased access for investors, while allowing investing personnel to focus on continuous refinement of strategies and adaptation methods in a quickly changing landscape.

Longer term, AI is only limited by the ingenuity of its human creators, and the extent to which they are willing to expand and regulate it. LLMs' potential to advance and adapt human productivity will open up new avenues for human work and redefine what societal contribution looks like. When Adobe Photoshop was initially released, people feared the death of photography. Fear is a rational reaction when facing something new and unknown. Yet, humans found a way, and creative and productive endeavors soon flourished as these new tools enabled more to be done with less.

As LLMs evolve, we anticipate larger and more complex modeling capabilities will lead to several operational efficiencies. Using a simple prompt, investors will be able to:

• Drive differentiation in performance by utilizing large data set analyses and tools previously only available to the largest entities. LLMs can efficiently synthesize any medium of data, including text, imagery and sound. With this capability, any data source can become a language for analysis. For example, investors will be able to instantly complete many tasks from predicting spread compression or expansion based on financial performance of assets and recent rating change activity to visualizing retail commentary by segment and overlaying it with average consumer borrowing balances and default rates to project spending and margins by sector.
  
• Access bespoke real-time performance visualizations of complex and time-intensive research directly into digestible formats. With analytics at your fingertips, users can focus on bringing data and decision-making to life, greatly shortening the time to answer while enhancing predictive abilities. Through easy-to-use plug-ins, users can analyze data, create charts, edit files and perform complex calculations to create visual outputs in seconds, instead of hours or days. With access to LLMs, an analyst looking into an industrial manufacturing business will be able to visualize leverage trends across the entire sector over the last 12 months in seconds, and save the output in any file type or format desired.
  
• Find correlations and trends currently hidden from view due to unstructured data sets and enhanced ability to collapse information silos. Finding new sources of data to feed LLMs will drive new observations and broader understanding of the interconnected nature of markets. As an example, LLMs trained on the language of Wi-Fi router data can detect and predict object locations within their sphere of influence, opening up significant amounts of previously siloed information for deeper consideration and analysis.
  
• Receive information from advanced and predictive analytics at a significantly lower cost. While initial model training can be expensive and time consuming, once trained, task-specific models can operate on the equivalent of a smartphone, with more complex and larger models operating within traditional server environments. The cost of entry is low and likely to grow cheaper in the coming years.

With LLMs, investors will be able to easily create smartphone applications that provide notifications for material sentiment score changes of communicated text and audio transcripts for a universe of companies, and notify users of material deviations of sentiment compared to recent historical averages.

These advancements are exciting, potential game changers for the financial industry. And still, one very important, looming question remains: How will AI's prevalence impact humans' job functions?

As AI infiltrates deeper into traditional research and analytics processes, it will likely shift humans' role to one of steward and teacher, tasked with maintaining and building new models and combining task-specific LLMs to drive even greater capabilities. Subsequently, efficiencies within traditionally manual financial processes will free up investment teams to focus on maintenance and expansion of model knowledge as drivers of enhanced predictive accuracy.

As a result, LLMs' direct access to a significant amount of financial data and expertise in the nomenclature of financial transactions has the potential to provide a narrow but important focus on the outward influencers of financial performance. We expect that task-specific training on sentiment vs. performance and interconnected information sources will drive quicker adoption of LLMs and remove an important initial layer of LLM training. Indeed, specialized LLMs are likely to enable even more immediate adaptation and analysis vs. comprehensive models.

We believe healthy skepticism surrounding AI is required today as the technology matures at a speed that should be uncomfortable when compared to other great enhancements in human productivity historically. Foundation of knowledge in electricity occurred in the mid-1700s, however widespread adoptions still took over 100 years. With AI and LLMs, we are facing a similar societal impact, unfolding exponentially over the next five years. Further, healthy skepticism typically leads to healthy debate, and ultimately successful outcomes.

While AI certainly has roadblocks that lie directly in its path ahead, adoption by financial institutions globally is all but inevitable. The value proposition of LLMs in deep predictive and analytical capabilities, coupled with the cost-savings financial institutions can achieve are impossible to ignore. As such, financial institutions would be wise to begin embracing AI as early as possible in its development lifecycle. This next phase of AI and LLM development represents an exciting attempt to expand human productivity while challenging the limitations of technology. As we enter the next chapter, we should endeavor to know the limitations of AI to truly understand its ability.

Philip Raciti is head of performing credit at Bardin Hill Investment Partners, based in New York. This content represents the views of the author. It was submitted and edited under Pensions & Investments guidelines but is not a product of P&I's editorial team.