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 second-place winners from University of Exeter: Maksim Kitikov, Harrison Maxwell and Kirill Papka.
Every day, large institutional investors execute billions of dollars in trades. Even when an investment decision is correct, money can still be lost through poor execution.
Sell too quickly and the market may move against you, increasing execution costs through slippage or adverse market impact. Sell too slowly and changing market conditions can create even larger losses. On a $5 million order, for example, these costs may appear small in percentage terms, but they can translate into thousands of dollars and compound across hundreds of trades.
This trade-off is known as the optimal execution problem.
To address it, we built an AI-powered trade execution platform that adapts to market conditions throughout the life of an order. The challenge was not teaching a model to outperform a benchmark. The real challenge was measuring that advantage and explaining its decisions in a way that clients, compliance teams, and regulators could understand.
Why Traditional Approaches Fall Short
Most institutional trading desks rely on predefined execution schedules. A time-weighted average price (TWAP) strategy divides an order into equal pieces and executes them over a fixed period. A volume-weighted average price (VWAP) strategy follows a similar approach but allocates trades according to expected market volume.
These methods remain popular because they are simple, predictable, and easy to explain. Both, however, share an important weakness: they cannot react to changing market conditions. Once the schedule is set, it will continue executing even if prices are falling or liquidity conditions deteriorate.
Building an Adaptive Alternative
To overcome this limitation, we used reinforcement learning, a form of AI that learns through trial and error by receiving rewards for good decisions and penalties for poor ones.
Our model was trained using market factors including volatility, liquidity, and broader trading conditions. Rather than following a fixed schedule, it continuously adjusts execution decisions as conditions change.
This adaptability is what distinguishes the model from traditional strategies and many existing adaptive systems, which typically optimize only at the start of a trade. Our approach reassesses conditions throughout execution while remaining transparent enough for its decisions to be explained.
Testing the Model
To evaluate performance fairly, we compared the reinforcement learning agent against a standard TWAP strategy on identical price paths. This removed the influence of differing market conditions and isolated the value added by each approach.
The results were encouraging. The model outperformed TWAP in roughly two-thirds of simulations, achieving an average improvement of around 20 basis points. On just one $5 million sell order of an S&P 500 ETF, that equates to nearly $10,000 in reduced execution costs.
The performance advantage remained statistically significant after accounting for noise and became more pronounced during declining markets, precisely when static schedules are most vulnerable.
The Importance of Explainability
Performance alone is not enough for deployment in real-world trading. Execution strategies must demonstrate that they pursued the best available outcome for clients. As AI regulation evolves, firms must increasingly demonstrate not only that a model works, but also how and why it reaches its decisions.
Simply stating that an "algorithm decided" is unlikely to satisfy a regulator, a compliance officer, or a client. For this reason, explainability was built into the model from the outset. Each decision is linked to the market conditions that triggered it and is accompanied by a concise justification. For example: "Sold 8% less than the benchmark schedule because market conditions were stable and liquidity remained strong." Such explanations allow decisions to be reviewed, verified, and communicated clearly to stakeholders.
Limitations of the Model
Our results should be viewed in context. The model tracks nine market factors, whereas production-grade systems may incorporate dozens of variables, high-frequency data feeds, and latency considerations. These enhancements could improve performance but would also increase complexity and operational cost.
In addition, our system executes a single instrument once per day in simulation. Although the observed edge appears genuine, it remains relatively small and noisy. A live trading environment would introduce trading fees, borrowing constraints, execution latency, and market regimes that differ substantially from the training data.
Looking Ahead
This project reinforced two important questions that should be asked before deploying any AI system that interacts with client capital:
- How would the model have behaved if market conditions had evolved differently?
- Can its decisions be explained clearly enough that a client can easily understand them?
A model that cannot answer the first question may be unreliable. A model that cannot answer the second may never be trusted. Success in financial AI requires both performance and explainability. Get those right, and the performance is worth having. Get them wrong, and an impressive number is just a rising market in a lab coat.
We would like to acknowledge our mentor, Sunil Shukla, CFA, whose guidance was invaluable throughout the project.
If you liked this post, don’t forget to subscribe to the Enterprising Investor.
All posts are the opinion of the author. As such, they should not be construed as investment advice, nor do the opinions expressed necessarily reflect the views of CFA Institute or the author’s employer.
Image credit: ©Getty Images