Can AI Decode Not Just What Management Says, but How They Say It?

In 2026, the CFA Institute launched its first AI Investment Challenge, a national competition where university teams apply artificial intelligence to real-world investment problems. It helps students build practical skills in investment analysis, data, and responsible AI use.

The CFA Institute AI Investment Challenge is designed to simulate how investment teams use technology to analyze data, generate insights, and present recommendations. Students participated in teams, working through a structured, multi-stage competition to develop an AI-enabled solution to a real investment problem. Throughout the challenge, participants received guidance from faculty advisors, industry professionals, and judges with expertise in investment practice and technology.

In our first AI Challenge, we fielded 28 teams and had to make difficult decisions, as all of the competitors submitted interesting and innovative solutions. The competition culminated in a final round where five top teams presented their solutions to a panel of judges to win a spot on the podium.

Over three consecutive weeks, we are showcasing our top three finalists. This AI innovation comes from our third-place winners from Heriot-Watt University and University of Glasgow: William Armstrong, Archie Clark, James Hillan, and Zac Moraghan.

• Earnings calls contain valuable signals beyond the transcript. By combining financial language analysis with vocal delivery metrics, investors can gain deeper insight into management confidence, consistency, and communication style.
• AI can make qualitative analysis more scalable and comparable. EarningsIQ demonstrates how tasks that traditionally require hours of manual review can be standardized across companies and reporting periods, helping analysts focus on interpretation rather than information gathering.
• The most effective investment applications of AI augment human judgment. Transparent, explainable tools can surface potential risks and opportunities while leaving investment decisions in the hands of analysts, improving efficiency without sacrificing oversight.

Earnings calls remain one of the most important sources of information for investors. Beyond the financial results themselves, they provide insight into management sentiment, confidence, and strategic direction. Analyzing these calls, however, traditionally requires hours of manual review, with analysts reading lengthy transcripts and listening to recordings in search of subtle signals that may influence investment decisions.

For firms covering dozens of companies each quarter, this creates a significant bottleneck. Analysts must prioritize coverage, meaning lower-priority companies often receive less attention. Furthermore, qualitative assessments can vary considerably between analysts, leading to inconsistent conclusions and making it difficult to compare management communication across firms and reporting periods.

As part of the CFA Institute AI Investment Challenge, our team set out to address a problem familiar to investment professionals: extracting meaningful insights from corporate earnings calls is time-consuming, inconsistent, and difficult to scale.

To address these challenges, we developed EarningsIQ, a multimodal AI platform designed to automate and standardize earnings call analysis by combining natural language processing and audio analytics.

Our methodology centered on the idea that valuable information exists not only in what management says, but also how they said it. As a result, EarningsIQ processes both the transcript and audio recording of an earnings call simultaneously.

For textual analysis, we used FinBERT, a finance-specific language model trained on financial documents and earnings call data. FinBERT analyzes segments of a transcript and classifies each according to sentiment, generating a score that reflects the positivity or negativity of management commentary. Unlike general-purpose language models, FinBERT is specifically designed to understand financial terminology and context.

For audio analysis, we employed Wav2Vec2, a deep learning model capable of extracting speech characteristics such as pitch, articulation, energy, and delivery consistency. Rather than focusing on speech recognition, we used these audio embeddings to measure variability in management delivery, providing an additional layer of insight beyond the transcript alone.

We then combined these signals to create several proprietary metrics. The flagship metric, the Management Confidence Index (MCI), integrates text sentiment, vocal delivery characteristics, pause behavior, and pitch variability into a single score ranging from 0 to 100. The aim is to provide investors with a standardized measure of management confidence that can be tracked over time and compared across companies.

Alongside MCI, we developed the Tone-Text Divergence (TTD) metric, which measures the gap between the positivity of management language and the confidence reflected in vocal delivery. A significant divergence may indicate that management's spoken confidence does not fully align with the optimistic tone of prepared remarks.

Additional supporting metrics included the Narrative Shift Index (NSI), which identifies meaningful changes in sentiment relative to previous quarters, and a Q&A Stress Score, designed to highlight differences between prepared remarks and responses during analyst questioning.

One of the most valuable insights from the project was the importance of combining multiple data modalities. Text analysis alone provides useful information, but adding audio signals offers a deeper understanding of management communication. Earnings calls are ultimately spoken events, and vocal delivery often contains information that cannot be fully captured through transcript analysis alone.

The project also reinforced the value of explainable AI within investment processes. Rather than producing opaque outputs, every metric within EarningsIQ can be traced back to underlying transcript segments or audio characteristics. This transparency allows analysts to verify results and maintain human oversight, ensuring the system functions as a decision-support tool rather than an automated decision-maker.

From a practical perspective, the prototype successfully demonstrated that a process which would traditionally require several hours of analyst time can be standardized and largely automated. Through our current implementation, we process a typical one-hour earnings call in approximately 30 minutes, while identifying clear pathways for further optimization through cloud infrastructure, GPU acceleration, and institutional data integrations.

Ultimately, EarningsIQ highlights how artificial intelligence can augment the efficiency of, rather than replace, investment professionals. By automating repetitive analysis and surfacing potentially valuable signals, analysts can spend less time reviewing transcripts and more time interpreting insights, challenging assumptions, and making informed investment decisions.

Through EarningsIQ, we demonstrated how multimodal AI can be applied to a real-world investment problem, creating a scalable framework for analyzing one of the market's most important sources of qualitative information.