Henrique Laurino dos Santos

Empirical methodologist in quantitative Marketing
Ph.D. candidate at the Wharton School

Currently, I am pursuing my Ph.D. at the Wharton School, University of Pennsylvania. I expect to go on the academic job market in 2024, and defend my thesis in 2025.


My research focuses on momentum-based marketing policies, whether that be controlling wearout in digital advertising, detecting community formation in adoption networks, or parsing the interplay between new product adoptions and user reviews. I am interested in how dynamic patterns emerge among consumers, between consumers and marketers, and within products themselves. 


Methodologically, I work in problems around Bayesian machine learning and point processes. I am interested in how Bayesian methods can counteract the tendencies of neural networks to overfit and explode, and how machine learning can help us interpolate posteriors faster.


You can reach me by email at hlauri@wharton.upenn.edu

Published works

In this article we use sentence-level embeddings to plot "narrative curves" for a large number of movie scripts. We then relate those to audience popularity and find common shapes of storytelling that are correlated with high scores.

Working papers (available by request)

My job market paper.  I propose a methodology for modelling clickstream data with Hawkes processes. Although Hawkes-type models are uncommon in the Marketing literature, they nest many usual alternatives and have a variety of advantages over them when datapoints are both sparse and clumpy. I present new parametric forms for models of this type, review methods of augmenting them with Bayesian hierarchical structures and nonparametric components, and develop new estimators for a crucial issue in the current digital advertising environment: how to estimate our models when users may not consent to cookies and ad impressions are not individually observable.

Starting from the methodology presented in my job market paper, we solve a variety of optimal dynamic advertising problems. We consider optimal policies for advertisers subject to different legal restrictions and with varying goals and planning horizons. We find that a variety of incentives, both structural and idiosyncratic, may lead to overadvertising practices.

Adding to the long tradition of forecasting how new movies will perform in theaters, we present a novel model that captures the endogenous loop between tickets sold, volume of user reviews posted, and score distribution among those reviews. We combine Bayesian hierarchical models and posterior distributions approximated by deep learning to devise forecast updating methods that achieve great performance with as few as three days of post-release data available.

Research in progress