We develop a novel framework that applies generative artificial intelligence to numerical data. Empirically, our approach substantially outperforms existing benchmarks in both predictive performance and economic interpretability. We further propose a real-time methodology to identify ex ante whether a factor or factor model is “pricing on” or “pricing off.” Our results show that factor premia are concentrated almost entirely in active states and are largely absent otherwise. Conditioning on these states yields economically and statistically significant out-of-sample performance gains. In addition, we show that difference-in-differences (DID) does not necessarily deliver causal identification when individual securities are aggregated into portfolios, and propose a protocol that restores inference.