Economic Validity in Tabular Foundation Models for Choice Prediction
A recent study introduces a two-phase adapter designed to address economic logic discrepancies in tabular foundation models employed for discrete choice forecasting. These models frequently yield economically nonsensical outcomes, like demand rising alongside price or negative willingness-to-pay figures. The adapter integrates foundation model predictions into a utility-maximization context: initially estimating a constrained choice model, followed by freezing parameters and training a corrective term. This approach ensures consistent price-demand correlations and measurable trade-offs. In tests on two transportation datasets, the adapter improves accuracy by as much as 13 percentage points while maintaining economic validity.
Key facts
- Tabular foundation models violate economic logic in choice prediction tasks.
- Raising a price sometimes increases predicted demand.
- Implied willingness-to-pay estimates are frequently negative or implausible.
- A two-stage adapter embeds predictions within a utility-maximization framework.
- First stage: estimate standard choice model with economically constrained parameters.
- Second stage: freeze parameters and train a correction term using foundation model predictions.
- Guarantees monotonic price-demand relationships under policy perturbation.
- Produces analytically computable trade-off measures.
- On two transportation datasets, the adapter recovers up to 13 percentage points of accuracy.
Entities
—