TabFDM
A foundation model for tabular data that keeps its accuracy when the regime shifts.
Fig. 4. When the regime shifts, a model fitted to the old one keeps forecasting the past. The series switches between calm and stormy regimes, the way geomagnetic activity does. Illustration, not real data.
- observed
- forecast, TabFDM, with 80% interval
- forecast, model fitted to the past regime
- now
Notes
Most models for tabular data assume that tomorrow follows yesterday's rules. In space weather, in markets and in many other places it doesn't.
TabFDM is built for data whose rules change. Where the regime shifts, it keeps its accuracy while LightGBM and current tabular foundation models lose theirs. It is a research project; we share results privately.
Access
If your data changes its rules and you need to forecast it anyway, write to us.
Request access on Telegram