A generative model learns how the data themselves arise
Slide 5 in the PDFA discriminative model estimates \(p(y\mid\mathbf x)\), the probability of a label given the inputs. A generative model estimates \(p(\mathbf x)\), or the joint \(p(\mathbf x,y)=p(\mathbf x\mid y)p(y)\), and is usually fitted by maximum likelihood,
Knowing \(p(\mathbf x)\) gives four uses in finance: a low density marks a suspicious observation (anomaly detection); draws from the model are scenarios for stress tests and risk measures; synthetic samples can be shared without client records; and a latent representation compresses many correlated series into a few factors.



NN_L8_ae_error_dist










