AI for Forecasting
AI for forecasting: how AI actually forecasts in finance
Narrower, and more useful, than the sales pitch.
AI forecasting is the use of AI to predict the future values of a time series: sales, demand, prices, risk. Someone will try to sell you a chatbot that does this. It will sound confident, and it will be wrong more often than it lets on, without ever telling you which times those were.
A language model cannot forecast on its own. It was trained on text, and a time series is a different kind of sequence, with structure and drift that text does not have. Ask a language model to forecast and it produces something that reads like a forecast, a number and a confident paragraph, generated the same way it generates any text. You cannot tell the good from the bad by looking. What works is to let an AI drive the real forecasting tools rather than pretend to be one.
There is a ladder of methods that genuinely forecast, from classical statistical models that still win competitions, through machine learning and deep learning, to the time-series foundation models now trying to become general-purpose. No rung is reliably best. The right choice depends on the series in front of you. So a forecasting system needs something to inspect the series, pick a method, run it, and check the result against reality. That is a job for an agent: a language model given real tools, a loop, and memory, where the model handles the orchestration and the tools do the computation.
I did a PhD building models that forecast financial markets from prices, news, and networks that evolve over time, and I wrote Singapore's first AI risk management guidelines for the financial sector. This set of pages is how to think about it.
Read more
- Why a language model cannot forecast on its own - what makes time different from text.
- The forecasting ladder - classical models to deep learning, and why the old methods still win.
- Time-series foundation models - the frontier, and why it is not finished.
- Why agents, not prompts - forecasting as a loop, not a single step.
- The tools a forecasting agent needs - data, model, and evaluation tools.
For the PDF, go to the AI Agents for Forecasting primer. See also the sibling guide on AI for investing.
The research behind this
My PhD was on deep learning for time series, networks and multimodal data, with financial forecasting as a main application, published in peer-reviewed venues:
- Guided Attention Multimodal Multitask Financial Forecasting with Inter-Company Relationships and Global and Local News - ACL 2022.
- Learning Dynamic Multimodal Implicit and Explicit Networks for Multiple Financial Tasks - IEEE Big Data 2022.
- Learning Dynamic Multimodal Network Slot Concepts from the Web for Forecasting ESG Ratings - ACM Transactions on the Web.
More on Google Scholar.
Get more
I am publishing more on AI in finance in the coming weeks. Subscribe to get it.