Making successful video advertisements has long been considered a combination of art and business acumen. In this work, we propose a system to assist human designers to produce more effective advertisements with predictable outcomes. We formalize this concept with a dynamic Bayesian network (DBN), where we represent the knowledge base with data collected from large-scale field experiments in a novel setting. Face and eye tracking which continuously measures viewers emotional responses and viewing interest on 169 television advertisements for 2334 participants, along with moment-to-moment branding activities in the advertisements are used to estimate the model. The resulting DBN represents relationships across advertisement content, viewers emotional responses, as well as effectiveness metrics such as ad avoidance, sharing and influence on purchase. Conditioned on the specified requirement on the ad, a human designer can draw high scoring samples from the DBN, which represent the optimized sequences of branding activities and entertainment content.