Scott Sherwood at the FT Global Commodities Summit: Unlocking the next phase of AI in trading

Awareness of AI has never been greater thanks to ChatGPT. However, while the leap forward in Large Language Models (LLMs) has seen mass market adoption, we should not forget the decades of critical ML and AI models that remain cutting-edge today.

At Kpler, we have been working with natural language processing, time series modelling, and deep learning since the company’s inception. Some of the areas where we apply these techniques include understanding market sentiment, predicting commodity flows, modelling supply and demand, tracking inventory levels, and identifying dark shipping practices.

When the world talks about AI today, it typically refers to a very small part of the field - focusing on Generative AI and specifically LLMs. Every week, we see new models and improved versions of existing models being released. Just this week, we saw a major announcement by Deepseek - not only releasing a drastically upgraded model but also making it available under the MIT open-source license (source). Given the commoditisation of models such as GPT-4, Grok, Claude, Deepseek, Gemini, etc., the big question is: what comes next?

I firmly believe that the next phase of AI will involve pairing these generalised reasoning engines with highly specialised models and deep proprietary datasets—combining their broad knowledge with focused domain expertise. My favourite example of this can be seen in chess (source). Many existing LLMs are incredibly poor at chess (source), yet Stockfish, an ML model, has been beating humans for more than two decades. In this case, by pairing the generalised knowledge of an LLM with the specialised capabilities of a bespoke ML model, you get an agent that can interact like a human - providing history, context, and conversation - while offering the deepest possible expertise in the game.

In the trading world, pairing these general reasoning engines with our proprietary data, exposing our algorithms as tools, and enriching them with the context of a user’s workflow will drive the next major area of advancement. Coupling this with the ability to launch hundreds or even thousands of agents - each with their own unique viewpoints—to generate millions of possible scenarios that can be mined to find the best candidate strategies in minutes will be a game-changer for our industry.

Here’s to the future!