This paper shows that a large pretrained time‑series model can predict consumer demand better if it is first taught basic economic logic. The authors fine‑tune Amazon Chronos‑2, a transformer‑based probabilistic forecasting model, on synthetic histories that were generated to obey a simple economic condition called the Generalized Axiom of Revealed Preference (GARP). After this fine‑tuning, the model made substantially better demand forecasts on a real experimental panel.

To create useful synthetic training data the researchers used Afriat’s theorem. Afriat’s theorem says that a set of observed price–quantity choices is consistent with utility maximization—people choosing the best affordable option—if and only if it satisfies GARP. GARP is easy to check and so the team used it to generate many price–quantity time series that are coherent with utility‑maximizing behavior. They then fine‑tuned Chronos‑2 on those synthetic GARP‑consistent histories. The synthetic data were not mixed with the real test data; they were used upstream to shape the model’s internal representations before the model saw the real panel at test time.

The real data used for evaluation come from an experimental portfolio‑choice panel by Ahn et al. (2014). That panel contains choices from 154 consumers over 50 budget scenarios. Each scenario has three goods and three prices, so the forecasting problem is multivariate. In this dataset only 20 of the 154 consumers satisfy GARP on their full 50‑period histories, which means most observed behavior is not perfectly rational by this standard. Despite that, fine‑tuning on GARP‑consistent synthetic histories improved Chronos‑2’s forecasts.

Why this helps: Chronos‑2, like other foundation models, is pretrained to capture generic temporal patterns such as trends and seasonality. It does not automatically encode economic links between prices and quantities caused by a budget constraint. By training on GARP‑consistent series, the model learns qualitative price–quantity relationships and substitution patterns that come from utility maximization. In practice this produced sizeable gains in accuracy: bundle prediction error fell by about 17–18% for forecast horizons H = 5, 10, and 15, and by 31% at H = 1, all relative to zero‑shot Chronos‑2.