(SeaPRwire) –
By: Oliver Hawthorne
Most quant shops treat large language models like a magic box. You feed in market data, the model spits out a trading signal, and nobody can really explain what happened in between. That’s the core anxiety. Firms are racing to deploy AI, but they are also terrified of models that hallucinate and cannot be audited. Waton Financial’s latest preprint on AlphaSchema is a deliberate attempt to break that cycle. It is not a product. It is a research manifesto that says: force the AI to write down its hypothesis before it writes a single line of code.
The paper, titled “AlphaSchema: Exploring the Space of Trading Semantics for LLM-Based Alpha Mining,” proposes a five-field structure for every candidate trading plan. Event, Context, Qualities, Direction, and Output. The model cannot just generate a Python script and call it a day. It must first articulate the market phenomenon, the conditions under which it happens, the filters to apply, the expected trading interpretation, and the numerical shape of the signal. Code generation is delayed until a plan survives the semantic search. The search itself combines broad exploration with surrogate-guided selection and local mutation. That is a very different workflow from the typical agent-based systems that jump straight to execution.
The empirical work uses historical Chinese equity data for the CSI 300 universe. Training data runs from 2016 to 2020. Validation from 2021 to 2022. The held-out test period covers 2023 to 2025. The authors report five independent discovery runs. They compare against representative machine-learning methods, deep sequence models, factor-mining libraries, and agent systems. Under that specific protocol, the strongest results appear on certain predictive and portfolio metrics. The paper also reports CSI 500 experiments. But the language is careful. The outcomes have not been independently validated. They are based on a limited number of runs over a single historical test period. No live execution. No real-account results. The public GitHub repository contains the core mining workflow, schema library, prompts, and configuration examples. Market data and API credentials are not included.
Now let’s talk about the commercial loop. Waton is a publicly traded company. It has a subsidiary, Waton Securities International, that operates under regulated infrastructure in Hong Kong. It has a separate product called MoTA, the Manager of Trading Agents workbench. AlphaSchema is explicitly not a current MoTA feature. It is not a live trading system. The company is drawing a hard line between research collaboration, product development, and regulated operating capabilities. That is smart. Publishing a preprint and open-source code costs credibility if the backtest evaporates. But it also opens the door to future integration. Chairman Zhou Kai stated clearly that historical testing is not live performance and that any production use would require further validation, controls, and review. The potential next research priorities include evaluation across market regimes, sensitivity to data definitions, reproducibility, robustness, and clearer human-oversight procedures. Those are research objectives, not deployment commitments.
The industry end-game here is straightforward. Every quant fund that uses AI faces the same regulatory and reputational risk. If a model blows up a portfolio, the regulator will ask why. If the answer is “the LLM decided to,” that is a career-ending moment. Waton is betting that the firm which can produce an inspectable, auditable, hypothesis-driven research pipeline will win the trust of both investors and regulators. The separation of plan from code is the weapon. The public code and the arXiv preprint are the ammunition. The live trading account is the target. I do not know if AlphaSchema will produce consistent alpha in production. But I do know that the strategy of making assumptions explicit before implementation is the only sustainable path for AI in finance. Everything else is just a fancy random number generator.
Author bio: Oliver Hawthorne, a Principal Correspondent permanently stationed at an international technology review, covering the intersection of quantitative finance and artificial intelligence.