Senior Lecturer and Computing Facilities in Charge, Department of Economics, The Chinese University of Hong Kong
Program Committee Member, Undergraduate Programme in Data Science and Public Policy, The Chinese University of Hong Kong
Associate Director, M.Sc. Programme in Economics, The Chinese University of Hong Kong
University of California, Berkeley Ph.D. in Economics 2011
Cornell University B.A. cum laude and Distinction in All Subjects 2005
“Predicting Auction Price of Vehicle License Plate with Deep Residual Learning” PAKDD 2019: Trends and Applications in Knowledge Discovery and Data Mining, pp. 179-188.
Due to superstition, license plates with desirable combinations of characters are highly sought after in China, fetching prices that can reach into the millions in government-held auctions. Despite the high stakes involved, there has been essentially no attempt to provide price estimates for license plates. We present an end-to-end neural network model that simultaneously predict the auction price, gives the distribution of prices and produces latent feature vectors. While both types of neural network architectures we consider outperform simpler machine learning methods, convolutional networks outperform recurrent networks for comparable training time or model complexity. The resulting model powers our online price estimator and search engine.
“Predicting Auction Price of Vehicle License Plate with Deep Recurrent Neural Network” Expert Systems with Applications.
In Chinese societies, superstition is of paramount importance, and vehicle license plates with desirable numbers can fetch very high prices in auctions. Unlike other valuable items, license plates are not allocated an estimated price before auction. I propose that the task of predicting plate prices can be viewed as a natural language processing (NLP) task, as the value depends on the meaning of each individual character on the plate and its semantics. I construct a deep recurrent neural network (RNN) to predict the prices of vehicle license plates in Hong Kong, based on the characters on a plate. I demonstrate the importance of having a deep network and of retraining. Evaluated on 13 years of historical auction prices, the deep RNN's predictions can explain over 80 percent of price variations, outperforming previous models by a significant margin. I also demonstrate how the model can be extended to become a search engine for plates and to provide estimates of the expected price distribution.
Arxiv Preprint: 1701.08711
Recent advance in statistical techniques and computational power have brought new ways to model complex data. At the center of these breakthroughs are artificial neural networks (ANN), biologically-inspired models that have been shown to be highly effective in modelling language, speech and image data. In this paper, I use simulations to study whether ANN can be used to estimate unobserved effects across a range of different data types.
“Challenges in Changing Social Norms: Evidence from Interventions Targeting Child Marriage in Ethiopia” with Eva Vivalt. Under review.
We study a set of interventions in Ethiopia geared towards eliminating child marriage. This is the first impact evaluation of a set of interventions specifically geared towards reducing child marriage, rather than obtaining a reduction of child marriage as a side effect. The interventions provide economic incentives and information about the potential harms of early marriage. Changing social norms is often thought of as very difficult, and it might be especially hard to change the typical age at first marriage given that marriage involves a matching problem. Regardless, we see clear evidence that both interventions reduce early marriage by about 8 percentage points. We also observe some positive spillover effects: the program appears to have increased the intra-household decision-making power of women. However, we find suggestive evidence that acceptance of violence against women may have increased and some evidence of increased polarization in beliefs about child marriage.
This study revisits the question of children's decision-making quality in light of the methodological shortcomings in early studies. Utilizing improved techniques and a considerably larger subject pool, I estimated that young children are significantly less consistent than previously reported. I find significant improvement in consistency between junior and senior elementary school students, with moderate evidence that the effect is mostly due to aging. I also confirm students' mathematics performance is significantly correlated with consistency, and that students' age, education and mathematics performance are all significantly correlated with their risk preference.
“ Addiction, self-control, and habit in online game playing: evidence from a field experiment ” with Dan Acland. Journal of the Economic Science Association, July 2018, Vol. 4, Issue 1, pp. 46-62.
We examine the utilization of commitment devices on an online word game that has few substitutes. The experimental environment allows us to accurately measure time spent on the game based on the activities of over 55,000 players. we find that approximately one-quarter of treatment-group players make use of one or both of the devices at least twice. In any given week, approximately 15% of active players make use of the devices, with the median user implementing one or other of the devices in 75% or more of game sessions. The treatment effect on session length was a reduction of approximately 3% for the treatment group as a whole, translating into as much as a 13% reduction for the group of users specifically. Overall the average reduction in games per week was 6.5% for the treatment group as a whole, and as much as a 28% reduction for users specifically.
The amount of time youth spend on game-playing is significant, with a recent representative survey pinning the number at 0.84 hours per day. To investigate the extent to which gameplaying can be attributed to self-control problems, I implemented a field experiment on a type of widely played multiplayer online games. I monitored the amount of time 105 undergraduates spent playing online games for a period of 3.5 months. Students assigned to the treatment group were additionally given software that they could use to limit their duration of play.
I found that the demand for commitment appears limited—while 79 percent of the treatment subjects used the software voluntarily in the first four weeks, the fraction eventually dropped to around 5 percent. At the end of the experiment, 10.4 percent of treatment subjects had a positive willingness-to-pay for the software. There is suggestive evidence that usage of the commitment device reduces duration of play but not frequency of play—subjects in the treatment group played an estimated 66.4 percent less than those in the control group, as measured by total hours played. The difference is driven by a reduction among heavy players, and persists even after most subjects stopped using the devices. Lastly, I find that players on average overestimated how long they would play.