Matheus Silva

I am a Ph.D. student in Economics at NYU. I am interested in different subfields of econometrics but I am particularly focused on the quantification of uncertainty around economic forecasts and the use of machine learning to construct such predictions.


Email: matheus at nyu dot edu

Some recent, non-academic projects

I use lap-by-lap data to calculate the probability of different drivers and constructors to win a Formula One season. I propose a fixed-effects functional regression to exploit the functional nature of lap time data, controlling for driver, constructor, circuit, and other factors. I combine this approach with a resampling strategy to calculate different podium probabilities and find that, all else equal, Lewis Hamilton is the best driver and Mercedes is the best constructor. I also conduct the simulation of a hypothetical “all star" season where the winners of the 1996-2022 seasons race against each other and find that Lewis Hamilton, Max Verstappen, and Nico Rosberg are the most likely to finish in the first three positions when controlling for other factors.


We combine data from traditional election polls with Google Trends data to infer the percent of the population that would vote for each presidential candidate in the Brazilian election in a daily frequency. We do so by treating the polling average and trends data as signals of the vote share each candidate will receive on the election day. Our results point toward a significantly more competitive electoral race than traditional polls indicate.

Replications

Michael P. Clements' and David F. Hendry's  "Forecasting economic time series"

Research papers

Forecasters usually report point predictions; however, understanding the randomness around such values is of practical importance. An example: a central bank predicts 2% inflation next quarter but is also interested in an interval (0%-4%, for example) that will contain the future realization of this series with a set probability. I show how to construct intervals as such, and I prove their asymptotic validity. I propose a model free method that encompasses, but not limited to, any off-the-shelf machine-learning method including high-dimensional ones. The method is based on a subsampling estimation strategy, consisting of analysing smaller cuts of the original time series. I prove the prediction intervals constructed with the subsampling method remain valid even when the data exhibits nonstationarities of many kinds –- such as time-varying parameters, structural breaks, unit roots, and transitions between steady-states. In addition to this theoretical work, I provide simulation studies to show the numerical performance of this method. I also apply the method to a demand dataset and to the forecast of inflation in a high-dimensional setup. The subsampling procedure extends to allow for comparisons of predictive accuracy between different models.

Estimating market power is an essential, if not complicated, task for market designers. This is rendered even harder when there are few sellers and few buyers in a given market. This research aims to allow for the distinction of market power from both the demand and supply sides simultaneously. I propose a bootstrap-based procedure to estimate multi-unit, two-sided auctions in which all agents can submit elastic price schedules. I use this methodology to estimate the private values of all agents in the Italian electricity market. The results suggest that buyers and sellers both have market power. An earlier version of this paper was awarded a student research prize and this paper is being expanded with Quang Vuong.

Empirical evidence shows that deep learning performs well in situations where the curse of dimensionality is computationally prohibitive. This situation is common when solving finite-time, dynamic discrete choice models because they feature infinite dimension state-space, and also because of the nonstationary nature of these problems. I show how to approximate policy functions of those models with deep neural networks. These methods reduce the complexity of the problem, simplifying the computation process when closed-form solutions are not available.

Work in progress

Teaching