AI investing, with risk in mind
The risks of AI investing
AI can help you invest. It can also fail quietly and expensively. The difference is how you think.
AI is genuinely useful in investing. It can read filings and news at scale, find structure in messy data, and turn a research idea into something testable in hours rather than weeks. The trouble is that the ways it fails are quiet. A model that is wrong does not look wrong; it looks like a clean backtest and a confident number, right up until it loses money in live markets.
So the guard is not a tool, it is a set of habits of thought. What follows is the mental models I use, not a build manual. If you want the how-to, the tools and the agent patterns, that is the AI for investing guide. This page is about keeping it honest.
The mental models
- Assume you have overfit until proven otherwise. Test enough strategies and one of them looks brilliant by pure chance. The more ideas you tried, the more a great backtest tells you about luck and the less about skill. Treat a strong result as a suspect, not a discovery.
- Hunt for look-ahead and data leakage. A model that quietly uses information it would not have had at the time looks like genius in testing and falls apart the moment it is live. Trace every input back to when you would actually have known it.
- The market is non-stationary. The regime you trained on is not the one you will trade in. A model is a bet that tomorrow resembles yesterday, so ask what breaks it, and what you do when the break comes.
- Correlation is cheap, real signal is rare. AI will always find a pattern. Before you trust one, demand a reason it should exist, not just the fact that it fit the data.
- Keep a human on the decisions you cannot monitor. Autonomy you cannot see into is risk you cannot manage. The more an AI acts on its own, the more you need a way to watch it and a point at which a person decides.
- An AI investing model is still a model. It carries model risk like any other. Validate it independently, monitor it for drift, document it, and size its use to how much harm a failure would do. Under the MAS AIRG, it is in scope like anything else.
- With agents, errors compound along the chain. When an AI agent takes a sequence of actions, a small mistake early becomes a large one by the end. Bound what the agent may do, and test the guardrails, not only the final output.
Where this fits
- AI for investing - how to build an AI investing agent: the tools, the patterns, and the production gap.
- The complete guide to AI risk management in finance in Singapore - where this sits in the wider picture.
- The AIRG explained - an AI investing model is a model the guidelines cover.
- The research behind this, on Google Scholar.
I did a PhD building AI models for investment and risk management, and I developed the MAS AIRG while leading AI risk supervision there. This is the risk-aware way to think about AI in investing. I am publishing more in the coming weeks; subscribe to get it.