AI Decoded: NLP the new language of investing
Natural Language Processing (NLP) is transforming investment strategies, turning unstructured text like earnings calls and filings into actionable insights.
Today’s rapidly evolving technological landscape offers a variety of advanced Artificial Intelligence (AI) techniques with the potential to enhance investment strategies. The Systematic Equity team at Allianz Global Investors has been at the forefront of this innovation, leveraging cutting-edge tools to stay ahead in an increasingly data-driven world. As the data universe expands beyond traditional structured data to include vast amounts of unstructured text – such as news articles, social media, and financial reports – investors should adapt by embracing more sophisticated techniques. Among these, Natural Language Processing (NLP) has emerged as a powerful solution, enabling us to extract valuable insights from text data at scale. The rapid advancements in large language models (LLMs) such as ChatGPT and DeepSeek underscore the growing impact of NLP, further expanding the possibilities for AI-driven insights in investment management.
With the success of ChatGPT – a generative AI tool – the technology behind it, NLP, has taken centre stage in the public eye. More recently, the launch of DeepSeek has further highlighted the power of LLMs, even driving notable movements in the stock market. However, generative AI and NLP did not emerge overnight. Their roots can be traced back to foundational techniques like word counts and dictionaries, like the well-known Loughran-McDonald dictionary1. While these early methods provided valuable insights, they had certain limitations in capturing the deeper nuances of language, spurring the development of more sophisticated and advanced techniques like GloVe2 and word vectors, which focused on capturing word relationships in a mathematical space. As NLP techniques advanced, LLMs such as BERT3, GPT4, and Llama5 elevated NLP to new heights moving beyond language comprehension to generation – creating coherent text, answering questions, and even writing code by learning complex patterns from vast datasets.
Exhibit 1: A brief history of Generative AI and LLMs
Source: Databricks, Allianz Global Investors, Systematic Equity team. March 2025.
Enhancing operational efficiency with LLMs and NLP
LLMs and NLP have improved operational efficiencies significantly across various domains, particularly in coding, where these technologies assist with code completion, generation, explanation, and translation, as well as debugging, optimization, and documentation. In the asset management industry, this automation largely frees portfolio managers and researchers from repetitive tasks, allowing them to focus on higher-value activities, such as researching investment signals with greater alpha potential – an essential priority for quantitative investment professionals like us.
However, operational efficiency is just one dimension of what LLMs and NLP bring to the investment management industry. Beyond streamlining workflows, a key question, particularly for quant investment professionals, is whether NLP can go beyond automation to directly extract meaningful insights from unstructured text to enhance systematic investment strategies.
Exhibit 2: Boosting operational efficiency with LLMs and NLP
Source: Allianz Global Investors, Systematic Equity team.
Leveraging unstructured data in investment strategies
The answer is yes. Fortunately, the finance and investment industry offer a wealth of unstructured data sources that can be analyzed and leveraged by investment professionals. Key examples include earnings call transcripts, company filings, news, and other corporate disclosures. In this section, we will focus on two crucial document types: earnings call transcripts and 10-K filings.
Earnings call transcripts
Earnings calls and their transcripts play a crucial role in conveying information from company management to investors. They provide valuable insights for two key reasons. First, the Q&A session during an earnings call allows equity analysts to ask questions and challenge management, offering a deeper understanding of the company’s position. Second, earnings calls occur several weeks after the reporting period begins, during which
10-K filings
A 10-K filing is a comprehensive annual report required by the U.S. Securities and Exchange Commission (SEC) that provides a detailed overview of a company’s financial performance. These filings are publicly available through the SEC’s EDGAR system. One of the most important sections is Risk Factors, which outlines potential risks that could impact the company’s financial health and operations. While much of the content remains unchanged from year to year, research suggests that newly emerging risk factors may signal potential weaknesses in future performance.
Traditionally, these texts from data sources like earnings calls and 10-K filing were manually analyzed, structured, and incorporated into investment decisions. However, for a quantitative equity strategy covering thousands of stocks, and with databases like FactSet containing over 250,000 earnings call transcripts, manually processing such vast amounts of information is impractical. Recent advancements in text processing now allow us to bypass this manual step, extracting valuable insights directly from raw text.
Exhibit 3: Word cloud of Apple earnings call
Source: Allianz Global Investors, Systematic Equity Research. March 2025. For illustrative purposes only.
Exhibit 4: Word cloud of Apple 10-K Risk Factor section
Source: Allianz Global Investors, Systematic Equity Research. March 2025. For illustrative purposes only.
Various approaches to utilizing text data
Yet, despite rapid advancements in language modeling technology, the alpha potential of these diverse methods remains uncertain. The Systematic Equity team has explored various approaches, including word-count-based techniques, approaches based on word embeddings such as GloVe, and transformer-based models such as BERT and LLaMA. Our findings highlight the distinct workflows and development processes between classical NLP models and LLM models. For classical models, the entire training process is conducted and controlled by the investment team. In contrast, LLMs are an example of transfer learning, leveraging the concept of Foundation Models. The LLM is pre-trained elsewhere, and the user focuses on fine-tuning and prompt engineering. Fine-tuning involves adapting the LLM to a specific use case, while prompt engineering entails crafting the most effective instructions. While the abilities of LLMs in many domains are spectacular, surprisingly, our results show that strategies derived from smaller, GloVe-based models are at least on par with those built using much larger models with significantly more parameters. This raises an interesting discussion about whether the focus should be on training and fine-tuning massive models or if there is greater value in smaller, specialized expert models.
In signal research one goal is to associate information on a company with future performance of its stock in equity markets. A popular approach tries to extract sentiment from earnings calls and then links positive sentiment to potential alpha opportunities in a separate stage. While this approach is intuitive and applied by several teams on AllianzGI’s investment platform, the GloVe based approach aims to directly associate text with potential future outperformance. The Systematic Equity toolbox comprises both approaches, this episode will focus on a smaller model based on the GloVe algorithm, and the LLM approach will be discussed in a later publication.
What is GloVe?
GloVe is a word-vector-based model that builds word embeddings based on word co-occurrence statistics from large text corpora. GloVe captures both global and local contexts, creating rich and precise word representations. A word vector, in its simplest form, is a numerical representation of a word’s meaning, enabling machines to understand and process language in a more sophisticated way.
In linguistics a semantic relationship refers to a meaningful association between words, while syntactic relationships define the grammar of a language. Rather than treating words as isolated units, word vectors place them in a multidimensional embedding space respecting semantic and syntactical relationships. The location of word embeddings represents similarities of words, as an example closest to “cat” are “dog” and “rabbit”, while the distance between “cat” and “computer” is rather large, as illustrated by Exihibit 5.
However, vector arithmetic can uncover even more nuanced relationships. The question “king is to man as queen is to _ ?” can be answered by finding the word closest to “man”-“king”+”queen” which actually turns out to be “woman”. Hence, GloVe captures the concept of gender. In a real world application GloVe can associate companies with their CEOs as “Allianz”-“Bäte” ~ “Cook”-“Apple” as illustrated by Exhibit 6. Similarly, the vector closets “small”-“smaller”+”big” is resolved as “bigger” proving that GloVe captures the syntactical concept of comparative forms, as demonstrated by the third chart in Exhibit 5. The crucial takeaway is that mathematical operation on the embedding space have a meaning in text space.
Exhibit 5: The essence of GloVe: capturing meaning through word juxtaposition
Source: Allianz Global Investors, Systematic Equity team. March 2025.
GloVe application in the Systematic Equity team
When we began our research on GloVe in 2016, we quickly realized that the model could capture similar relationships relevant in financial contexts. For example, as mentioned above and also illustrated below, the relationship between companies and their CEOs, which reflects a semantic relationship, is encoded by the relative positioning of their representatives in the vector space, allowing us to extract meaningful patterns important for financial analysis.
For the Systematic Equity team, the goal is to leverage this kind of data source and technique to develop a signal that enhances our factor definitions, which is a cornerstone of our multi-factor investment approach. One research project, for instance, focuses on predicting the likelihood that a company hosting an earnings call will outperform its peers in the near term based on its transcript. So, how do we approach this?
Our training pipeline for earnings call transcripts consists of two major steps. The first step involves training the GloVe model, an unsupervised machine learning task that uses our proprietary database of financial texts as input. To prepare the necessary input for the second step, consider an earnings call transcript as a set of words and use the GloVe model to represent each word as a vector. Since this naturally results in many vectors as we have many words in an article, aggregation is required to condense them into a single vector representing the entire transcript for further processing. The simplest form of aggregation is the computation of elementwise averages, which turns the transcript into one vector that is ready for downstream processing by machine learning methods. The second step of the pipeline involves fitting a machine learning model, more specifically a logistic regression model, that links the embeddings of earnings call transcripts to potential future outperformance relative to a suitable peer group.
Exhibit 6: A real-world GloVe example in finance
Source: Allianz Global Investors, Systematic Equity team. March 2025.
Exhibit 7: Training workflow
Source: Databricks, Allianz Global Investors, Systematic Equity team. March 2025.
Once the models are fitted and validated, they are ready for inference tasks, which are performed in four steps as soon as a new transcript arrives shortly after the earnings call is completed: Tokenization, Embedding (with GloVe model), Aggregation, and Prediction (with logistic regression model), as illustrated in the below exhibit.
These types of pipelines processing textual data have become foundational for various NLP tasks we perform, such as our proprietary earnings call transcripts score. Since sell side analysts participate in Earnings Calls and update their earnings estimates partially based on information from earnings calls, it is natural to seek improving the Revisions score with information extracted as described above. We have augmented the Revisions factor, one of the key factors in the Best Styles strategy, as early as 2018. As illustrated in Exhibit 9, our analysis shows that leveraging the capabilities of GloVe (and other NLP techniques) can improve the information ratio by 9%. This highlights the predictive power embedded in earnings call transcripts.
Exhibit 8: Inference pipeline
Source: Allianz Global Investors, Systematic Equity team. March 2025.
Exhibit 9: Backtest has shown that Earnings Call Transcript score has improved the IR significantly
Source: Allianz Global Investors, Systematic Equity team. Valuation period: 31/12/1998 – 31/12/2024. The hypothetical performance and simulations shown are for illustrative purposes only and do not represent actual performance; they do not predict future returns. Please see important information regarding back-testings and hypothetical or simulated performance data at the end of this document. For illustrative purposes only and does not represent actual performance of any client account. The information should not be relied upon as an indicator of future results.
Conclusion
Advancements in NLP and LLMs have significantly improved the way we process and extract insights from financial text. This article illustrates an example of how the Systematic Equity team leverages GloVe embeddings to efficiently transform earnings call transcripts and other corporate disclosures into structured data for quantitative analysis. Most importantly, backed by simulations our approach has demonstrated a potential tangible improvement, enhancing the information ratio of our Revisions factor. By integrating these textual insights into our systematic investment process, we have strengthened our ability to identify signals that drive excess returns.
As our research progresses, we will continue to refine these techniques and explore their broader applications in systematic investing. Future publications will delve deeper into LLM-based approaches, further evaluating their effectiveness in extracting alpha from unstructured financial data.
Opportunities
- High return potential of stocks in the long run
- Particular opportunities through Systematic Equity approach
- Broad diversification across numerous securities
- Possible extra returns through single security analysis and active management
- Currency gains possible in unit classes not hedged against investor currency
Risks
- High volatility of stocks, losses possible
- The volatility of fund unit prices may be strongly increased
- Underperformance of the investment approach possible
- Underperformance of Systematic Equity approach possible at times
- Limited participation in the yield potential of single securities
- Success of single security analysis and active management not guaranteed
- Currency losses possible in unit classes not hedged against investor currency
