Weldon Rice

Weldon Rice heads up AI ETFs at QRAFT Technologies, a Seoul, Korea-based investment technology company founded in 2016. Weldon's path to finance was unconventional: he moved to Asia right after college, lived in Korea for over 12 years, and studied under quantitative hedge fund managers and Goldman Sachs's Korea operation. QRAFT received a $140 million investment from SoftBank, which accelerated the firm's expansion into new asset classes and jurisdictions. The majority of QRAFT's employees are researchers and developers who build AI models entirely in-house.

On this episode, recorded live at the Exchange ETF conference, Weldon talks with Brad about how QRAFT uses AI for stock selection (not as a thematic bet on AI companies), the difference between analytical and generative AI in portfolio management, and their two ETFs: QRFT (US Large Cap) and AMOM (Large Cap Momentum).

What QRAFT Actually Does with AI

Weldon starts with an important distinction: QRAFT is not an AI-thematic fund that buys AI companies. It's a fund that uses AI to pick the stocks in the portfolio. The firm positions itself as "quant 2.0," taking academically researched factors that have been around for decades (value, momentum, quality) and using machine learning and deep learning to build more reliable exposure to those factors with better risk-adjusted returns.

QRAFT's AI sits on two sides. The analytical side uses pattern recognition on structured numerical data: price, market data, financial ratios, and other quantitative inputs. This is the core of their stock selection engine. Traditional quant models rely on linear regressions, but QRAFT is building non-linear models because, as Weldon puts it, "we know that markets aren't linear." The deep learning side adds more complexity, with neural networks that can identify patterns across larger datasets and more variables than traditional machine learning approaches.

On the generative AI side (large language models for text and sentiment analysis), QRAFT has some capabilities but has been more cautious about deployment. They recently partnered with LGI Research specifically because of LGI's strength in large language models, launching a fund that incorporates sentiment data from news sources into the AI prediction model. Weldon emphasizes they don't add AI features just for marketing purposes. Every AI input has to earn its place by demonstrably improving the model's predictive accuracy.

QRFT and AMOM: The Two ETFs

QRFT is QRAFT's US large cap ETF. It uses the full AI engine to select stocks from the large cap universe, rebalancing on a regular cadence as the models update their predictions. The fund doesn't make predictions beyond about one month, which Weldon explains is a deliberate constraint. Their research found that prediction accuracy degrades rapidly beyond that horizon, so they keep the window tight and update frequently rather than making long-dated bets.

AMOM is the large cap momentum ETF. It uses QRAFT's AI to identify momentum opportunities within the large cap space. The key innovation here is that the AI can identify when momentum signals are likely to work and when they're likely to reverse, which is the classic problem with momentum strategies. Pure momentum gets crushed during momentum crashes (like Q4 2018 or the COVID rotation in late 2020), and QRAFT's model aims to reduce exposure before those reversals happen by reading early warning signals in the data that traditional momentum screens miss.

The SoftBank investment allowed QRAFT to expand across asset classes and geographic markets, extend the frequency range of their strategies from intraday to monthly, and build out more sophisticated capabilities like long-short strategies. Beyond ETFs, QRAFT operates as a B2B solution provider, building AI-powered model portfolios, robo-advisors, and portfolio construction tools for large financial institutions in Korea and increasingly in other markets.

Guardrails and Continuous Improvement

Weldon addresses the natural question about AI guardrails. The models are constantly updated with new research and data inputs, but there are constraints on what the AI can do. Position limits, sector concentration limits, and other risk management rules are coded in as hard constraints that the AI cannot override. The team of researchers continuously evaluates new data sources and model architectures, but any changes go through a rigorous validation process before being deployed in live portfolios. It's a balance between letting the AI learn and adapt while preventing it from making outsized bets that could blow up.

Key Takeaways

• QRAFT uses AI to pick stocks, not to invest in AI companies. Their models combine machine learning and deep learning to build non-linear factor exposure, which they call "quant 2.0."
• The firm received $140 million from SoftBank, growing from 40 to significantly more employees, with the majority being researchers and developers who build all AI models in-house.
• Predictions are capped at roughly one month because accuracy degrades rapidly beyond that horizon. The models update frequently rather than making long-dated bets.
• AMOM uses AI to identify when momentum signals are likely to work or reverse, aiming to reduce exposure before momentum crashes that destroy pure momentum strategies.
• QRAFT also operates as a B2B provider, building AI-powered portfolio tools for large Korean financial institutions, making ETFs just one part of their broader technology business.

Listen to the full conversation on Spotify, Apple Podcasts, or YouTube.