Member-only story
When Tables Learned to Forecast
How tabular foundation models are turning time series prediction into an ordinary regression problem
Most of the recent attention in artificial intelligence has gone to language models.
That makes sense. A system that writes paragraphs and answers questions is considerably easier to demonstrate than one that predicts customer churn from a spreadsheet.
Yet a large part of the world’s useful data does not look like language.
It looks like a table.
A bank has rows of customers and columns describing their balances, income, payment history, and loan status. A factory records temperature, pressure, vibration, and equipment failures. A retailer tracks prices, promotions, inventory, holidays, and daily sales.
Financial researchers work with returns, valuation ratios, volatility measures, macroeconomic variables, and technical features.
This type of information is called tabular data. It remains the standard format for an enormous share of business, scientific, and financial modelling.
For years, deep learning struggled to dominate this area.
Neural networks transformed image recognition, speech processing, and natural language…
