1. Investment decision feasibility: multi-agent AI for institutional investors

Institutional investors often have to compare competing investment proposals against risk limits, portfolio fit and benchmarks. This challenge uses multi-agent AI to test whether that assessment can be automated reliably, and to produce explanations that a human investment team can check.

• This project explores whether a multi-agent collaboration (such as specialized risk analyst and portfolio manager agents) can assess competing investment proposals, identify conflicts, recommend feasible adjustments, and evaluate benchmark suitability.
• Using a synthetic portfolio of public and private assets, the student will design a multi-agent workflow incorporating verification, self-critique, or human checkpoints to prevent compound errors, produce explainable investment assessments, and test their financial validity.

• LLM agents, RAG, tool use, rigorous evaluation design, Python.
• Interest in financial research workflows helpful.

2. Indexing versus active management and market functioning

Passive investing has grown rapidly, and the balance between indexing and active management is shifting within many asset classes. This challenge uses AI to measure that trend and to understand what it means for price discovery and the functioning of markets.

• What is the trend in indexing versus active management within an asset class, and how can AI measure it from fund, holdings and flow data?
• What does that trend mean for price discovery and market functioning, including liquidity, price efficiency and how quickly information is reflected in prices?
• Can AI identify turning points in the trend, and how closely do those turning points match observed changes in market behavior?
• How do markets respond as the balance shifts? Suggested approach: scenario testing or agent-based modeling.

• Financial markets and market microstructure knowledge, time-series analysis and econometrics, Python or R, AI or NLP methods for extracting signals from fund and holdings data.
• Scenario testing or agent-based modelling helpful.

3. Multimodal AI for information integrity and automated research assurance

Disclosures, charts and datasets often disagree with one another, and checking them by hand is slow. This challenge tests whether multimodal AI can detect inconsistencies and unsupported claims automatically, and how far the results can be trusted.

• Can multimodal models detect inconsistencies across tables, charts, text and filings, with calibrated confidence?
• Can a multi-agent system check numerical consistency between text, tables, charts and datasets, and verify that cited sources support the claims made?
• How can AI detect omissions and contradictions across disclosures and flag unsupported conclusions, and what does that mean for data reliability?

• Multimodal ML (vision-language, table and chart understanding), calibration and selective prediction, multi-agent systems, Python.
• Accounting or disclosure familiarity helpful.

4. AI-coded internal solutions: standards and measured return

Non-engineers are building internal solutions with AI coding tools. This project examines what standards and controls make them reliable, and how to measure their return.

• What framework of infrastructure, build standards and control gates allows non-engineers to build AI-coded solutions to the same standard as engineered software?
• What instrument measures the return and output of AI-built internal solutions in a way that is valid, comparable across functions and credible to Finance?

• Software engineering fundamentals (software development life cycle, testing, code review, CI/CD, Git)
• Hands-on use of AI coding tools and agentic workflows (e.g. Claude Code, Copilot, Cursor), including writing reusable skills or rule files
• Secure coding and LLM-specific risks; evaluation and monitoring of AI-built software
• Python or TypeScript. Helpful: empirical measurement of productivity and quality (metrics design, statistics)
• Technical writing and communication.

5. Energy resources for global data centers

Data center growth raises questions about which energy sources will power it. This project maps the planned energy resources behind global data centers. MSCI data on about 13,000 data center locations can be made available.

• What energy resources (gas, batteries, nuclear) are planned for global data centers, based on public disclosures?
• Are there also documented instances of social opposition at these locations?

• Data pipelines, APIs and web scraping, NLP and LLMs for classification and information extraction

6. Self-evolving event intelligence system

A multi-agent AI system that continuously finds, classifies and updates events against a taxonomy, and proposes taxonomy changes over time.

• Can a multi-agent AI system continuously find and classify events against a taxonomy while keeping it consistent?
• How can the system propose taxonomy updates, and generate evidence-grounded charts and summaries from the event data?

• Strong proficiency in Python, LLMs, agentic AI frameworks (e.g., LangGraph, AutoGen), and retrieval-augmented generation (RAG).
• Hands-on experience in automated event detection, information extraction, semantic classification, entity resolution, and deduplication.
• Experience developing AI evaluation frameworks to ensure classification accuracy, temporal consistency, evidence traceability, and controlled taxonomy evolution.
• Ability to design automated analytical workflows that generate source-grounded visualizations, trend analyses, and narrative summaries.
• Experience building scalable, production-ready AI pipelines with human-in-the-loop validation and monitoring.

7. Real-time detection of extreme-weather impacts in company disclosures

Extreme-weather impacts disclosed by companies are spread across regulatory filings and hard to track. This project uses AI to identify and structure these disclosures in near real time, creating a continuously updated record of realized financial and operational impacts.

• Can major extreme-weather events be identified from public weather monitoring or news, and used to trigger monitoring of company regulatory filings on their effects over the following days, weeks and months?
• What financial and operational impacts (such as revenue loss, input cost impacts, share price movements and insurance effects) can be consistently extracted and quantified from these disclosures?

• Data pipelines, APIs and web scraping, NLP and LLMs for classification and information extraction.
• Familiarity with financial and regulatory filings beneficial.

8. Fund intelligence at scale: extracting and validating thematic exposure

Fund documents, including prospectuses, factsheets, annual reports, and holdings files, show what a fund says it does and what it actually holds. Extracting that information reliably is difficult. Documents differ in format, language is vague, and sources rarely agree on names or identifiers. This challenge proposes building and validating a general framework for doing this well, using AI-fund exposure as the worked example.

• How can an AI system reliably extract structured information from fund documents across source types, and how should its accuracy be validated?
• The AI-fund case study tests whether the framework can identify funds claiming AI thematic exposure, how closely those claims match reported holdings, and whether funds can be classified into different buckets using these insights (we can leverage the new MSCI AI value chain taxonomy)

• NLP and LLM extraction and classification.
• Document parsing (PDF, HTML, tables).
• Multimodal methods, evaluation design and calibration, APIs and web scraping, Python.
• Knowledge of investment themes and fund documentation helpful.