About
I'm an Assistant Professor at the
I completed my PhD in 2020 at the
I'm currently working on projects related to matching, procurement, online platforms and behavioral operations.
Education
PhD in Operations, Information and Technology
2020 • Stanford University
M.A. in Economics
2020 • Stanford University
M.S. in Operations Management
2014 • Universdad de Chile
B.A. in Industrial Engineering
2014 • Universidad de Chile
Working Papers
Improving Match Rates in Dating Markets through Assortment Optimization
+ Abstract 

We study how online platforms can leverage the behavioral considerations of their users to improve their assortment decisions. Motivated by dating markets, we consider a platform that must select an assortment of potential partners to show to each user in each day. Users can like/not like each potential partner presented to them, and a match is formed if two users (eventually) like each other. The platform's objective is to maximize the expected number of matches in a time horizon. Using data from a dating company, we estimate the main inputs of our model, namely, the like and log in probabilities, and find that the number of matches obtained in the recent past has a negative effect on the like behavior of users. Leveraging this finding, we propose a family of heuristics that decide the assortment to show to each user on each day. We show that these heuristics considerably improve the number of matches generated by the platform in markets simulated based on the real data and further identify the sources of improvement. Finally, we implemented a field experiment that confirms that the improvement achieved by our algorithms also translates into our partner's platform. 

Increased Transparency in Procurement: the Role of PeerEffects + Abstract 

Motivated by recent initiatives to increase transparency in procurement, we study the effects of disclosing information about previous purchases in a setting where an organization delegates its purchasing decisions to its employees. When employees can use their own discretionwhich may be influenced by personal preferences—to select a supplier, the incentives of the employees and the organization may be misaligned. Disclosing information about previous purchasing decisions made by other employees can reduce or exacerbate this misalignment, as peer effects may come into play. To understand the effects of transparency, we introduce a theoretical model that compares employees' actions in a setting where they cannot observe each other's choices, to a setting where they can observe the decision previously made by a peer before making their own. Two behavioral considerations are central to our model: that employees are heterogeneous in their reciprocity towards their employer, and that they experience peer effects in the form of income inequality aversion towards their peer. As a result, our model predicts the existence of negative spillovers as a reciprocal employee is more likely to choose the expensive supplier (which gives him a personal reward) when he observes that a peer did so. A laboratory experiment confirms the existence of negative spillovers and the main behavioral mechanisms described in our model. A surprising result not predicted by our theory, is that employees whose decisions are observed by their peers are less likely to choose the expensive supplier than the employees in the no transparency case. We show that observed employees’ preferences for compliance with the social norm of "appropriate purchasing behavior" explain our data well. 

Improving the Chilean College Admissions System + Abstract 

In this paper we present the design and implementation of a new system to solve the Chilean college admissions problem. We develop an algorithm that obtains all stable allocations when preferences are not strict and when all tied students in the last seat of a program (if any) must be allocated. We use this algorithm to determine which mechanism was used to perform the allocation, and we propose a new method to incorporate the affirmative action that is part of the system and correct the inefficiencies that arise from having doubleassigned students. By unifying the regular admission with the affirmative action, we have improved the allocation of approximately 3% of students every year since 2016. From a theoretical standpoint, we show that some desired properties, such as strategyproofness and monotonicity, cannot be guaranteed under flexible quotas. Nevertheless, we show that the mechanism is strategyproof in the large, and therefore truthful reporting is approximately optimal. 

School Choice in Chile + Abstract 

Centralized school admission mechanisms are an attractive way of improving social welfare and fairness in large educational systems. In this paper we report the design and implementation of the newly established school choice mechanism in Chile, where over 274,000 students applied to more than 6,400 schools. The Chilean system presents unprecedented design challenges that make it unique. On the one hand, it is a simultaneous nationwide system, making it one of the largest school admission problems worldwide. On the other hand, the system runs at all school levels, from PreK to 12th grade, raising at least two issues of outmost importance; namely, the system needs to guarantee their current seat to students applying for a school change, and the system has to favor the assignment of siblings to the same school. As in other systems around the world, we develop a model based on the celebrated Deferred Acceptance algorithm. The algorithm deals not only with the aforementioned issues, but also with further practical features such as softbounds and overlapping types. In this context we analyze new stability definitions, present the results of its implementation and conduct simulations showing the benefits of the innovations of the implemented system. 

Strategic Behavior in the Chilean College Admissions Problem
+ Abstract 

We analyze the application process in the Chilean College Admissions problem. Students can submit up to 10 preferences, but most of the students do not fill their entire application list ("shortlist"). Even though students face no incentives to misreport, we find evidence of strategic behavior as students tend to omit programs if their admission probabilities are too low. To rationalize this behavior, we construct a portfolio problem where students maximize their expected utility of reporting a ROL given their preferences and beliefs over admission probabilities. We adapt the estimation procedure proposed by Agarwal and Somaini (2018) to solve a large portfolio problem. To simplify this task, we show that it is sufficient to compare a ROL with only a subset of ROLs ("oneshot swaps") to ensure its optimality without running into the curse of dimensionality. To better identify the model, we exploit a unique exogenous variation on the admission weights over time. We find that assuming truthtelling leads to biased results. Specifically, when students only include programs if it is strictly profitable to do so, assuming truthtelling underestimates how preferred selective programs are and overstates the value of being unassigned and the degree of preference heterogeneity in the system. In addition, ignoring the constraint on the length of the list can also result in biased estimates, even if the proportion of constrained ROLs is relatively small. Our estimation results strongly suggest that "shortlist" students should not be interpret as truthtellers, even in a seemingly strategyproof environment. 

Dynamic College Admissions and the Determinants of Students’ College Retention
+ Abstract 

We analyze the relevance of incorporating dynamic incentives and eliciting private information about students' preferences to improve their welfare and outcomes in dynamic centralized assignment systems. We show that the most common assignment mechanism, the Deferred Acceptance (DA) algorithm, can result in significant inefficiencies as it fails to elicit cardinal information on students' preferences. We collect novel data about students' preferences, their beliefs on admission chances, and their college outcomes for the Chilean college system. We analyze two main behavioral channels that explain students' dynamic decisions. First, by exploiting discontinuities on admission cutoffs, we show that not being assigned to ones' topreported preference has a positive causal effect on the probability of reapplying to the centralized system and switching one's major/college, suggesting that students switch to more preferred programs due to initial mismatches. Second, we find that a significant fraction of students change their preferences during their college progression, and that these changes are correlated with their grades, suggesting that students may learn about their matchquality. Based on these facts, we build and estimate a structural model of students' college progression in the presence of a centralized admission system, allowing students to learn about their matchquality over time and reapply to the system. We use the estimated model to disentangle how much of students' switching behavior is due to initial mismatches and learning, and we analyze the impact of changing the assignment mechanism and the reapplication rules on the efficiency of the system. Our counterfactual results show that policies that provide score bonuses that elicit information on students' cardinal preferences and leverage dynamic incentives can significantly decrease switchings, dropouts, and increase students' overall welfare. 

Application Mistakes and Information Frictions in College Admissions
+ Abstract 

We analyze the prevalence and relevance of application mistakes in a seemingly strategyproof centralized college admissions system. We use data from Chile and exploit institutional features to identify a common type of application mistake: applying to programs without meeting all requirements (admissibility mistakes). We find that the growth of admissibility mistakes over time is driven primarily by growth on active score requirements. However, this effect fades out over time, suggesting that students might adapt to the new set of requirements but not immediately. To analyze application mistakes that are not observed in the data, we design nationwide surveys and collect information about students’ true preferences, their subjective beliefs about admission probabilities, and their level of knowledge about admission requirements and admissibility mistakes. We find that between 2%  4% of students do not list their true most preferred program, even though they face a strictly positive admission probability, and only a fraction of this skipping behavior can be rationalized by biases on students’ subjective beliefs. In addition, we find a pulltocenter effect on beliefs, i.e., students tend to attenuate the probability of extreme events and underpredict the risk of not being assigned to the system. We use these insights to design and implement a largescale information policy to reduce application mistakes. We find that showing personalized information about admission probabilities has a causal effect on improving students’ outcomes, significantly reducing the risk of not being assigned to the centralized system and the incidence of admissibility mistakes. Our results suggest that information frictions play a significant role in affecting the performance of centralized college admissions systems, even when students do not face clear strategic incentives to misreport their preferences. 

TwoSided Assortment Optimization
+ Abstract 

We study the twosided assortment problem recently introduced by Rios, Saban and Zheng (2021). A platform must choose an assortment of profiles to show each user on each side of the market in each stage. Users can either like/dislike as many profiles as they want, and a match occurs if two users see and like each other. The goal of the platform is to maximize the expected number of matches generated. Given the complexity of the problem, we focus on the one lookahead version of the problem (i.e., two stages), and we provide performance guarantees for different variants of the problem. Specifically, we show that if users cannot see each other simultaneously, there is an approximation guarantee of 1−1/𝑒 obtained using submodular optimization techniques. For the more challenging case where simultaneous shows are allowed, we provide an approximation guarantee of (1−1/𝑒)/24. Finally, using data from our industry partner (a dating app in the US), we numerically show that our proposed heuristic outperforms several relevant benchmarks. 

Work InProgress
Scoring Systems in TwoSided Platforms


We propose a rating system for twosided dating markets where i) men and women ratings lie on the same rating scale and ii) rating of an individual takes into account both the liking behavior and the ratings of the counterparties who are recommended their profile for evaluation. We find four uses of this dimension reduction exercise using this ELOstyle rating system. First, incorporating ratings of counterparties in adjusting one's rating allows the platform to more efficiently use information and is an improvement over simple measures such as percentage likes amongst recommendations the platform has suggested. Second, assortative matching by recommending manwoman pairs closest in rating maximizes the expected number of matches on the platform. Third, for any individual i, using data from all N\i to compute i's estimated rating, and then using maximumlikelihood techniques to use i's liking behavior holding N\i ratings fixed to calculate his behavioral rating, allows detection of behavioral biases such as overconfidence. Lastly, the singledimension rating system interacted with other covariates allows the platform to better identify which dimensions an individual behaves differently relative to their estimated rating: i.e., whether individual i's preference criterion are different for counterparties with same race, religion, hobbies, etc. 

Published
MultiPeriod Forecasting and Scenario Generation with Limited Data + Abstract 

Data for optimization problems often comes from (deterministic) forecasts, but it is naïve to consider a forecast as the only future possibility. A more sophisticated approach uses data to generate alternative future scenarios, each with an attached probability. The basic idea is to estimate the distribution of forecast errors and use that to construct the scenarios. Although sampling from the distribution of errors comes immediately to mind, we propose instead to approximate rather than sample. Benchmark studies show that the method we propose works well. 

Modeling and Estimating Commodity Prices + Abstract 

A new methodology is laid out for the modeling of commodity prices, it departs from the "standard" approach in that it makes a definite distinction between the analysis of the short term and long term regimes. In particular, this allows us to come up with an explicit drift term for the shortterm process whereas the longterm process is primarily driftless due to inherent high volatility of commodity prices excluding an almost negligible mean reversion term. Not unexpectedly, the information used to build the shortterm process relies on more than just historical prices but takes into account additional information about the state of the market. This work is done in the context of copper prices but a similar approach should be applicable to wide variety of commodities although certainly not all since commodities come with very distinct characteristics. In addition, our model also takes into account inflation which leads us to consider a multidimensional system for which one can generate explicit solutions. 

Stochastic Unit Commitment at ISO Scale, an ARPAe Project + Abstract 

We describe a multifaceted effort to demonstrate the practical implementation of stochastic unit commitment at RTO scale. The inputs for the scheduling engine consist of multiple hourly trajectories of load and variable generation availability together with probabilities that accurately describe their likelihood of occurrence given information available on the day ahead. Preliminary computational results indicate that the methods have promise. 

Effect of Including HighSchool Grades Rank in the Admission Process to Chilean Universities + Abstract 

This paper analyses the effect of including high school grade rankings as a new factor in the admission process to Chilean universities. The paper evaluates the impact of different weighting strategies of the high school grade ranking and identifies socioeconomic and gender characteristics of the students who benefited and were harmed by the inclusion of this new factor. Starting with the weightings of the different factors considered in the Admission Process for 2012, we simulate, using a selection algorithm, alternative weightings for high school grade rankings in the 2013 Admission Process. We also evaluate the effect of the actual increase in the weighting of grade ranking in the 2014 admission process. Even though the impact on students' entrance and exit from the selection list is rather small, the introduction of the high school grade ranking into the admission process has an effect on the composition of the students selected, producing greater socioeconomic and gender equality. 

Building a Stochastic Programming Model from Scratch: A Harvesting Management Example + Abstract 

We analyse how to deal with the uncertainty before solving a stochastic optimization problem and we apply it to a forestry management problem. In particular, we start from historical data to build a stochastic process for wood prices and for bounds on its demand. Then, we generate scenario trees considering different numbers of scenarios and different scenariogeneration methods, and we describe a procedure to compare the solutions obtained with each approach. Finally, we show that the scenario tree used to obtain a candidate solution has a considerable impact in our decision model. 

Toward Scalable Stochastic Unit Commitment  Part 1: Load Scenario Generation + Abstract 

Unit commitment decisions made in the dayahead market and during subsequent reliability assessments are critically based on forecasts of load. Traditional, deterministic unit commitment is based on point or expectationbased load forecasts. In contrast, stochastic unit commitment relies on multiple load scenarios, with associated probabilities, that in aggregate capture the range of likely load timeseries. The shift from pointbased to scenariobased forecasting necessitates a shift in forecasting technologies, to provide accurate inputs to stochastic unit commitment. In this paper, we discuss a novel scenario generation methodology for load forecasting in stochastic unit commitment, with application to real data associated with the Independent System Operator of New England (ISONE). The accuracy of the expected load scenario generated using our methodology is consistent with that of point forecasting methods. The resulting sets of realistic scenarios serve as input to rigorously test the scalability of stochastic unit commitment solvers, as described in the companion paper. The scenarios generated by our method are available as an online supplement to this paper, as part of a novel, publicly available largescale stochastic unit commitment benchmark. 

Other Working Papers
Using Stochastic Programming with Spatial Scenarios to Inform Management of Flammable Forest Landscapes
+ Abstract 

Fire has a significant impact on timber production on many forest landscapes. In this article we describe how we used the method of generating a discrete set of fire scars described in Kulhmann et al. (2015) to generate a set of fire scar scenarios which we incorporated in a two stage stochastic linear programming (SLP) timber harvest scheduling model. We solve the resulting SLP model to produce timber harvesting strategies which we evaluate by assessing how well they perform in a simulated forest management planning environment in which fire burns a portion of the forest each period. We explore how the quality of the SLP solution varies as the number of fire scar scenarios varies and the sensitivity of the solutions to some of the parameters of the fire simulator. We then compare the performance of the SLP solutions with those produced when fire is ignored and when a traditional burn probability model is used to estimate the probability that each stand will burn. We found that the SLP solutions produced only slightly better expected returns but significantly better performance with respect to risk measures. 

College Admissions Problem with Ties and Flexible Quotas
+ Abstract 

We study an extension of the classical college admission problem where applicants have strict preferences but careers may include ties in their preference lists. We present an algorithm which enables us to find stable assignments without breaking ties rules, but considering flexible quotas. We investigate the properties of this algorithmstability, optimalityand we show that the resulting algorithm is neither monotone nor strategyproof. The mechanism is used to solve real instances of the Chilean college admission problem. Among our results, we show that the welfare of students is increased if flexible quotas and a studentoptimal assignment are combined. Finally, we argue why such assignment may be desirable in the Chilean context. 
