How AI can open new avenues for systematic investment research

The second quarter was marked by another sustained risk‑on rally and strength in momentum stocks. Key drivers included continued enthusiasm for the artificial intelligence (AI) capital expenditure cycle and a strong earnings season, both of which seemed to supersede concerns about the U.S.‑Iran war. Valuations and quality, on the other hand, were largely detractors (Figure 1).

Factor performance trends were persistent in the U.S. (Figure 2A) and even more consistent with emerging markets (Figure 2B).

We see fundamental and technical reasons for this alignment in U.S. and emerging markets factor performance:

• Fundamental: Emerging markets’ exposure to memory and hardware companies that have benefited from AI resulted in similar performance to U.S. equities. The AI infrastructure trade, in other words, expanded beyond the U.S.
• Technical: Significant retail investor participation, combined with exchange‑traded funds that offer leveraged exposure to a stock or theme, contributed to this effect.

This convergence in factor performance between U.S. equities and emerging markets raises important questions for asset allocators: chiefly, whether emerging market exposure still provides the level of diversification it has in the past, and whether the AI trade going global will exacerbate any volatility, if and when expectations disappoint. Given that the MSCI Emerging Markets Index is more concentrated, it may perform like a high-beta version of the Russell 1000 and the Russell 2500 indexes.

How we are using AI to enhance quantitative investment research

Our research has made extensive use of AI. We have found that these tools appear to add value, especially in situations where traditional quantitative models have exhibited shortcomings.

When designing analyses that incorporate large language models (LLM),¹ we seek to integrate our strength in quantitative analysis with our fundamental research analysts’ deep understanding of individual companies and industries.

The significant underperformance of the quality factor and software stocks over the past year prompted us to use AI to explore two key debates in the market:

• What can we learn from a systematic approach to analyzing definitions of quality that are more qualitative in nature?
• Which software companies might be better positioned to withstand AI disruption?
• What did our analyses uncover?
• The LLM version of quality also reflected recent underperformance for the factor; however, qualitative measures of quality outperformed quantitative metrics.
• The LLM prompt that we engineered did a reasonable job of independently capturing the market’s view of which software stocks could be more at risk of AI disruption.

More broadly, the analyses reinforced the value of bringing both quantitative and fundamental expertise together in leveraging AI tools.

---

Case study 1: Does the definition of quality matter in small‑ to mid‑cap stocks?

Many investment professionals have attributed the quality factor’s run of underperformance in the small‑ and mid‑cap universe to investor speculation.

We have explored alternative explanations, including technological disruption, the option‑like nature of equity valuations, and changes in market structure and market participants.²

To be clear: We believe that quality companies should outperform over a full market cycle. However, we also recognize that traditional measures of quality can and will lag, sometimes for extended periods.

Why and how we built an LLM framework for business quality

Traditional quantitative measures of quality—commonly used by investment managers, risk models, and data providers—typically focus on financial characteristics that indicate whether a company has been stable and growing. Frequent inputs in quantitative quality ratings include:

Profitability—e.g., earnings and free cash flow generation;

Earnings stability—e.g., measures of historical profitability; and

Balance sheet strength—e.g., low leverage.

Fundamental investors often take a broader view of quality, incorporating assessments of a company’s industry structure, its competitive advantages, pricing power, management quality, and, increasingly, whether AI could be a tailwind or headwind to the business. Many of these qualitative attributes have been difficult to measure systematically.

LLMs help to make qualitative analyses more practical by extracting information from all manner of unstructured textual sources, from company filings and earnings call transcripts to investment research reports.

We do not view this approach as a replacement for traditional financial metrics. Rather, the intent is to complement them with a systematic framework for evaluating qualitative aspects of business quality that have historically depended on fundamental analysts’ judgment.

To explore this opportunity, we developed an LLM quality framework that aims to assess whether a business appears to have durable earnings growth and be well positioned from a competitive standpoint. The process involved several steps:

1. Developing a detailed definition of fundamental quality and refining it through iterative feedback from both our analysts and the LLM itself;
2. Building a supporting framework around the model to reduce hallucinations and improve consistency of outputs;
3. Integrating the model with our internal databases so that our analysts’ proprietary research could be incorporated alongside publicly available information; and
4. Creating a batch‑processing system that uses the model to score the members of the small‑ and mid‑cap Russell 2500 Index on a weekly basis.

Key findings from our LLM quality framework

The initial results have been encouraging. As a robustness check, the correlation between our quantitative quality score and the LLM quality score stood at 74% as of June 30, 2026.³ This level of correlation is high enough to suggest that both approaches measure a similar underlying concept, yet low enough to indicate that the information is not redundant. Put differently, the two approaches appear to capture quality through different lenses.

Consistent with this conclusion, about 79% of the Russell 2500 Index received the same quality classification from both methodologies (Figure 3).

We can also use these quantitative and LLM scores to analyze the recent rally in low‑quality stocks. Looking at trailing 12‑month returns across the Russell 2500 Index, stocks that both methodologies classified as low quality outperformed stocks that both systems placed in the high‑quality bucket by 5.6% (Figure 3). More interestingly, among stocks where the two methodologies disagreed in their classifications, companies rated as high quality only by the LLM outperformed companies rated as high quality only by the quantitative model by 9.1%.

While this analysis should be viewed as preliminary, it suggests two potential conclusions:

1. The LLM framework also reflects the substantial outperformance of lower‑quality stocks over the past year.
2. Qualitative dimensions of quality may have outperformed traditional quantitative measures of quality. This divergence could reflect AI exposure, resilience to disruption, stronger pricing power, or other business characteristics that are difficult to capture using conventional financial metrics alone.

In fact, when we asked the LLM to analyze structural differences between its view of quality and the traditional quantitative view of quality, it attributed the narrow disagreements that cropped up to:

• AI penalizing “currently good but structurally threatened companies”—high current ratios, but AI doubts sustainability; and
• AI crediting “structurally good but currently ugly companies”—volatile financials, but AI sees a durable franchise underneath.

Caveats to our analysis of business quality

However, there are two important caveats to our case study.

Look‑ahead bias: We do not currently have a point‑in‑time LLM quality history extending back one year. Although the model was instructed to focus on fundamental quality characteristics, its assessments may nevertheless have been influenced by knowledge of the subsequent market environment.

Potential overfitting: Both the prompt design and scoring framework were refined through multiple iterations, raising the possibility that the methodology may be tailored to the current market environment. Although the aim of this research project was to improve consistency and economic interpretability rather than maximize historical performance, additional out‑of‑sample testing will be necessary to assess the robustness and generalizability of the framework.

Nevertheless, the LLM classifications were directionally consistent with the broader observation that lower-quality stocks outperformed. We therefore believe the analysis may be informative, although the magnitude and persistence of the results require further out-of-sample testing.

---

Case study 2: Which software companies appear better positioned against AI disruption?

Software stocks underperformed significantly over the past year, driven by various AI concerns, including outright product replacement, reduced pricing power, and lower demand for seat‑based licenses.

Is it possible to distinguish between software businesses that are relatively resilient to AI disruption and those that are more exposed?

How we used an LLM to analyze software companies’ resilience

Developing and applying an LLM framework for assessing software companies’ potential to resist AI disruption involved the following steps.

• Collaborating with our software analyst team to identify 12 business characteristics that could contribute to greater resilience and defensibility in an AI‑driven environment;
• Developing a detailed, multifactor prompt that evaluates companies across these dimensions (Figure 4);
• Building a supporting framework to improve consistency and reduce LLM hallucinations that result in falsities; and
• Generating resilience scores across software companies in the small‑ and mid‑cap Russell 2500 Index and iterating with our fundamental analysts to refine the methodology and outputs.

After multiple rounds of refinement, we ranked small‑ and mid‑cap software companies into quintiles based on their estimated resilience to AI disruption and examined their performance since June 30, 2025.

Key findings from our LLM analysis of software resilience

During periods of software weakness—particularly the drawdown from late 2025 through early 2026—the companies identified as most resilient meaningfully outperformed those identified as least resilient (Figure 5).

Beginning in May 2026, however, the relationship reversed alongside a sharp recovery in software stocks, with the less resilient cohort rallying hardest and ultimately closing the 12 months ended June 30, 2026, ahead of the more resilient group. Even so, the pattern of the most resilient quintile leading during selloffs held consistently over the period analyzed.

This pattern is consistent with what one might expect from a defensiveness framework. Software companies perceived as more insulated from AI disruption may help to mitigate downside during periods of elevated concern, while companies viewed as more exposed to this risk may benefit disproportionately when those worries recede.

While the observation period remains short and returns over any given window will be shaped by shifting sentiment, the early results suggest that resilience to AI disruption is becoming an increasingly important differentiator within the software sector.

Over longer investment horizons, we believe software companies with stronger competitive positioning, greater pricing power, more defensible customer relationships, and lower AI disruption risk should be better positioned to create shareholder value. We believe making these distinctions is important for small-and mid-cap investors.

---

Conclusions

LLMs have the potential to expand the set of investment questions that can be addressed systematically, particularly in areas where traditional quantitative models historically have struggled to measure qualitative business characteristics. In our experience, the most effective applications tend to arise not from fundamental or quantitative expertise but from combining the two. Incorporating insights from these two disciplines allows economic intuition and technological sophistication to enhance each other.

The examples we presented illustrate this integrated approach. Our LLM framework suggests that systematic measures of qualitative business quality may capture information that is distinct from conventional quality metrics. Similarly, our software resilience framework demonstrates how LLMs can help evaluate complex competitive dynamics that historically have depended on judgment. While both applications remain in the early stages of development and require additional validation, the initial results suggest that LLMs can provide meaningful signals that complement existing research tools.

As these technologies evolve, we believe the greatest opportunity lies not in replacing established investment processes but in extending them. By combining structured financial data with systematic analysis of unstructured information, investors may be able to measure important business characteristics that historically have been difficult to evaluate at scale. We expect integrating AI tools to become an increasingly important component of quantitative investment research.

 

Appendix

Factors are our internally constructed metrics, defined as follows:

Valuation: Proprietary composite of valuation metrics based on earnings, sales, book value, and dividends. Specific value factor weighting may vary by region and sector.

Growth: Proprietary composite of growth metrics based on historical and forward‑looking earnings and sales growth. Factor selection and weighting vary by region and industry.

Momentum: Proprietary measure of medium‑term price momentum.

Quality: Proprietary measure of quality based on fundamental and stock price stability; balance sheet strength; and measures of profitability, capital usage, and earnings quality.

Profitability: Return on equity.

Risk: In this paper, risk is the standard deviation of trailing 12‑month returns. High‑risk stocks exhibit higher standard deviations, indicating a wider degree of variation or dispersion in their historical return.

Size: Market capitalization (positive return means larger stocks outperform smaller stocks).

Quintile spread: Also referred to as long‑short returns, a quintile spread is calculated by sorting securities based on a specific characteristic or factor criterion, dividing them into five groups (or quintiles), equal‑weighting the securities within each quintile, and then subtracting the bottom‑quintile returns (lowest 20%) from the top‑quintile returns (highest 20%).

Factors and indices cannot be invested in directly and are shown for illustrative purposes only. They do not reflect performance of actual investments nor do they reflect the reduction of fees associated with an actual investment, such as trading costs and management fees.

All investments are subject to market risk, including the possible loss of principal. Past favorable company characteristics may not persist into the future.

Investment Risks: Equities can lose value rapidly for a variety of reasons and can remain at low prices indefinitely. Growth stocks are subject to the volatility inherent in common stock investing, and their share price may fluctuate more than that of income‑oriented stocks.  Small‑cap stocks have generally been more volatile in price than large‑cap stocks. Mid‑caps generally have been more volatile than stocks of large, well‑established companies. Investing in technology stocks entails specific risks, including the potential for wide variations in performance and usually wide price swings, up and down. Technology companies can be affected by, among other things, intense competition, government regulation, earnings disappointments, dependency on patent protection, and rapid obsolescence of products and services due to technological innovations or changing consumer preferences.