How AI is transforming investment management: State Street’s strategic approach

• AI and machine learning have shaped State Street’s investment process for years, driving data-driven strategies and deeper insights
• Specialized models and natural language processing tools are tailored for financial tasks, enhancing forecasting and risk analysis with robust, transparent methods
• Human expertise remains central, guiding disciplined integration of AI to improve outcomes while maintaining clarity and control

Artificial intelligence (AI) is transforming investment decision-making, moving traditional approaches toward data-driven, technology-enhanced strategies.

At State Street Investment Management, machine learning (ML) and AI have become integral to our investment process, supporting use cases from trading to forecasting. We apply a range of ML tools and advanced techniques—non-linear methods, natural language processing, generative AI—through the lens of our core investment principles.

While these tools bring powerful capabilities, integration requires discernment and discipline. Some tools fit seamlessly into existing workflows; others demand new approaches and pose unique challenges. Ultimately, however, AI is not replacing our human judgement—it’s amplifying it to enable deeper insights and better outcomes for investors.

Our foundational principles for investment modeling

At State Street Investment Management, our investment process is structured, deliberate. It is also a reflection of our investment philosophy. Four principles guide our investment models:

1. Foundational clarity: Model logic should be supported by well-articulated economic or behavioral perspectives
2. Transparency: Model outcomes should be interpretable (linked to some rationale), and the model itself should be parsimonious to an extent possible (to make that linkage visible)
3. Robustness: When model logic is tested on historical data, trivial changes in inputs should not drastically affect outcomes
4. Efficiency of process: Workflows around modeling should be flexible, cost-effective, and robust to support continuous and efficient innovation

How these principles shape our models

These principles are reflected in our modeling choices. Many State Street Investment Management models use linear structures, where changes in factor variables lead to proportional changes in recommended active positioning.

We routinely prefer:

• Rankings over raw signals
• Risk budgeting over full portfolio optimization
• Constant factor allocations over dynamic variety

These choices improve model robustness, support treating outliers as noise rather than signal, and greatly enhance model transparency.

Financial markets are complex. Regime shifts and unstructured data don’t always conform neatly to linear frameworks. A significant amount of information is now coming to us in an unstructured, textual or visual form. These and other complexities compel State Street Investment Management to adopt additional modeling paradigms and methodologies, ones that typically fall under the umbrella of ML.

Systems integration and operational efficiency are key to sustainable success. That’s why we design quantitative investment platforms with high levels of flexibility, scalability, and automation at their core. And at State Street, we are continuously enhancing these platforms—drawing on emerging technologies like generative AI to super-power these efforts.

Using machine learning in the investment process

Machine learning spans a broad and continuously evolving set of capabilities. Here, we briefly outline key use cases, challenges associated with using ML in investment modeling, and examples of how State Street Investment Management applies these methods today.

Early ML methods

While you won’t find these methods in modern ML textbooks, they are part of any serious data scientist’s toolkit, in our view:

• Regularized regressions
• Markov chains
• Kalman filtering
• Noise suppression tools, and
• Stochastic volatility models

These approaches help manage noise and model regime-driven, dynamic, and non-linear relationships. We use them extensively in our quantitative investment processes. They also fit very neatly within the foundational principles we outlined before. Resulting models have clear economic logic, and are sufficiently transparent and robust. These models have long joined the proverbial portfolio of modeling ideas used within State Street Investment Management.

Here are three examples of how we’ve applied early ML methods:

Tactical asset allocation

The relationship between asset prices and underlying factors is complicated and it often shifts with market regimes. Our Investment Solutions Group (ISG) developed proprietary techniques using the Hidden Markov Model (HMM) approach for tactical trading in commodities, credit,¹ and broader markets—methods still in use by ISG today.

Dynamic linear modeling

ISG has also been using dynamic linear modeling for situations in which the time variation of asset price-to-factor relationship is material, but gradual. ISG has used early ML techniques like the Kalman filter since 2014 for hedge fund replication strategies.

Noise management

Another example is the extensive use of noise management techniques within ISG investment practice, including random matrix-based correlation cleaning, selective risk budgeting, fuzzy mathematics,² and other approaches.

Core ML methods

From early ML approaches, core ML methods (ML) have emerged, including:

• Neural networks
• Clustering
• Non-linear classifiers (e.g., random forest, boosting)

ML is a paradigm shift from “traditional” science. Unlike the traditional scientific method where you form a hypothesis and test it to draw conclusions, ML enables us to directly uncover patterns in data without a hypothesis or an explicit understanding of the system being analyzed. This makes ML ideal for incredibly complex, multifaceted problems.

However, the same elements that make it ideal in some circumstances make it less effective in others. ML success over time depends on two conditions being met:

1. There is plenty of data to train the model
2. The relationship we are trying to uncover is stable over time

For example, ML methods are doing a fantastic job in image recognition. You can take a picture of your cat, upload it online, and the model would tell you that this is indeed a cat, what breed it is, and so on. That’s because there are, by conservative estimates, more than 6.5 billion cat images on the internet. Cats also invariably remain cats as time passes.

The financial modeling world is much less forgiving. Financial markets change frequently, with outcomes driven by dynamic interaction among market participants. Variables in the financial system and relationships between them often experience structural shifts.

Financial data also varies dramatically in terms of quantity. Asset allocation models, for example, deal with extremely limited amounts of data which are usually counted in hundreds or, at best, thousands. State Street Investment Management models rarely trade more frequently than weekly, given the size of most of our portfolios. Twenty years of weekly data for a single market yields a small (by ML standards) set of 1040 data points—not nearly enough for ML algorithms.

With stock and bond forecasting, however, availability of data is much better. There are more than 20,000 tradable stocks in the global universe and more than 25,000 tradable bonds in global sovereign and corporate universes alone. For these universes, the same 20 years of weekly data yields a significantly richer dataset, ranging from several million to tens of millions of data points.

This is why when it comes to “core” ML methods State Street Investment Management uses those methods primarily for securities selection.

Where we apply core ML methods

State Street Investment Management’s Systematic Equity team has developed an XGBoost-based method to forecast company fundamentals such as earnings and cash flows—essential inputs for estimating expected stock returns.

This method—which models complex relationships like trailing firm fundamentals, fast-moving market variables, analyst estimates, and macro indicators—outperformed plain sell-side analysts across regions in terms of accuracy. While the method has not been implemented in live portfolios as overlapping information was already present in other parts of the Systematic Equity’s model, machine-learned company fundamentals can serve as useful inputs in other contexts, providing in-house estimates with better coverage and less biases than conventional analyst estimates.

One of the key problems in bond portfolio optimization is risk estimation. Traditionally, bond portfolios are managed through a set of constraints on duration, sector allocations, etc. This approach, however, has significant limitations for concentrated bond portfolios. Developing a risk model—or, as ML specialist would call it, a “similarity measure”—for bonds that combines multiple features, some of the numeric (like duration) and some represent a category (like sector) is a potentially valuable application of ML methods such as Random Forest. State Street Investment Management Systematic Credit researchers are working on it as we speak.

Core ML methods could only be used for financial modeling if they are carefully designed with investment expertise and executed within a rigorous research discipline. An entirely data-driven approach—as performed by typical ML applications—is unlikely to succeed when tackling investment problems. Unless carefully calibrated and controlled, a ML method would “overfit” investment data and would not help us achieve desired investment outcomes.

Non-generative large language models in finance

Natural language processing (NLP) is an area where ML has been highly effective and one of the most impactful applications of ML—it has substantially transformed how we extract insights from texts in financial modeling.

Why NLP matters

The availability of unstructured data has increased in recent decades, with over 80% of data available on the Internet being unstructured. In finance, unstructured textual data such as earnings call transcripts or regulatory filings have been a relatively untapped source of information but can offer valuable insights into a company's operations and business prospects.

How we apply NLP in our investment process

State Street Investment Management Systematic Equity team employs a suite of factors based on NLP. These factors are designed to capture return-influencing features and insights from unstructured textual data. For earnings call transcripts, for instance, they apply a variety of techniques—from traditional linguistic processing to advanced ML—to assess a wide range of measures such as the following:

• Tone and complexity of the language
• Subtle sentiment behind executive statements
• Management behavior during the call

These measures have proved to enhance the team’s ability to gauge overall sentiment and quality beyond the headline financial results released. Other useful textual data our Systematic Equity team processes include regulatory filings, patents, job postings, and other data.

Our use of neural-network-based embedding

Another powerful technique we use is a neural-network-based “embedding” that converts text into lower-dimensional numerical vectors, while preserving meaning and structure, or semantic and syntactic information. Once the texts are converted, companies that discussed similar topics will appear clustered together in the resulting numerical vector space (Figure 1). This can serve as a useful byproduct to identify novel peer groups beyond the traditional sector and industry classifications.

NLP is an ideal way to apply ML in finance, as one can train the model using a large corpus of textual data without relying on noisy market data or even borrow pre-trained language models directly from the mature field of ML research. And the core language model will remain stable even when its outcomes may need to be used differently in response to changing market conditions.

Generative AI, a powerful tool with limits

The 2022 release of ChatGPT, the artificial intelligence chatbot from OpenAI, has been a global game changer. Since that moment, generative AI modeling and implementation has become a transformative wave affecting many facets of human existence, including investment management.

State Street Investment Management is among those companies who’ve embraced this innovation, and we have made an AI-powered chatbot available to our investment teams. Our researchers use it extensively as a research assistant, helping to:

• Browse academic literature and compile organized reference lists
• Generate ready-to-use code suggestions
• Organize documentation and draft routine communications

While we find generative AI capabilities incredibly useful for improving workflows, enhancing individual productivity, and boosting efficiency across research and operational tasks, we’re not yet comfortable using it for direct investment modeling.

Generative AI’s mysterious ability to never refuse a question, no matter how complicated—and to provide confident answers, without nuance or clear reasoning—raises concerns for us.

When applied directly to investment forecasting, Generative AI tests three out of four foundational principles of financial modeling that we set out in the beginning. Yes, it is very efficient. But the economic logic of its financial prediction is unknown, stability of those predictions is poor (the same question asked differently or at different times may yield different results), and transparency of how it arrives at the conclusion is absent (to put it mildly). These obstacles need to be addressed before we could seriously rely on generative AI in financial modeling.

Strategic integration of AI at State Street

Even with its noted limitations, artificial intelligence is rapidly becoming essential in investment processes across the industry.

AI is quickly becoming essential in investment processes at State Street Investment Management and other financial institutions. Industry leaders must focus on both implementing AI and maintaining a competitive edge by adapting quickly, building strong advantages, and positioning well in the market.

State Street Investment Management is forging our own path through this new and exciting terrain. Data remain the foundation of our quantitative investment processes. Over the past decade, we’ve enhanced our datasets by incorporating unstructured information, proprietary analyst insight, and high-value third-party resources. Today we are:

• Fine-tuning our models with these datasets to create tailored solutions
• Developing new workflows with the potential to become proprietary intellectual property

The rise of AI enables human insight and machine discipline to enhance each other. At State Street Investment Management, part of our quantitative investment involves integrating qualitative insights from analysts and portfolio managers into our robust modeling process. This approach aims to consistently capture alpha, improve returns, and maintain a competitive advantage.

The ongoing evolution of artificial intelligence necessitates the implementation of intentional strategies to maintain a competitive edge. Organizations can achieve industry leadership by consistently investing in new technologies and establishing themselves as pioneers in innovation. State Street Investment Management is actively following this strategic approach.

When context is everything, strong investment expertise matters

We don’t treat “AI” as a single construct and instead prefer a differentiated approach. ML methods can offer a path forward in areas that have historically been difficult to tackle, including nonlinear relationships, noise management, and multi-category similarity. Large language models can be fruitfully used to process unstructured data. Generative AI is extensively used by us to improve workflow and efficiency, even though we do not yet include it directly into models because of lack of transparency and robustness.

When deploying or exploring AI tools, it is important to understand the context. ML is best when combined with the knowledge of strong investment teams that can control the structures, contexts and training, and determine the questions that AI will address. When used thoughtfully, AI can be a powerful tool in investing, and it already plays a significant role within State Street Investment Management quantitative investment process.

Have more questions about how we use AI in our investment models or workflows? Please contact us or reach out to your State Street representative.