Cedric Ehouarne

Cedric Ehouarne, Ph.D.  
Managing Director, Senior Quantitative Finance Manager
Bank of America | Model Risk Management 

Manhattan, New York City, NY 

LinkedIn

I am a macroeconomist with expertise in credit risk and research interests in consumer finance and computational economics. I manage a team of quants in Model Risk Management at Bank of America. Over the years, I led various teams to review, challenge, and improve key models used for credit underwriting and credit life cycle management, loss forecasting (CCAR/CECL), finance / revenue forecasting (PPNR), Climate Risk, and Risk Appetite in the retail space, notably consumer card and small business, mortgage, overdraft, and consumer vehicle lending. The material on this website does not reflect the views of the bank and mostly reflects my work at Carnegie Mellon University during my Ph.D. studies. 

I graduated from Carnegie Mellon University in May 2016 with a Ph.D. in Economics and received the Alexander Henderson Award for Excellence in Economic Theory. My thesis studied the U.S. business cycle at the cross-sectional level to shed new light on classical puzzles in Macro-Finance. My dissertation committee members were Lars-Alexander Kuehn (chair), Sevin Yeltekin, Ariel Zetlin-Jones, and Daniele Coen-Pirani.  

Interests 

Macroeconomics: business cycle, credit risk, unemployment, housing market, asset bubbles
Consumer Finance: household leverage, optimal default, portfolio decisions, incomplete markets
Computational Economics: heterogeneous-agent models (households, firms), optimization problems

Research

“Cluster Search Optimization” (toy project) 

This article proposes a stochastic-search, population-based algorithm for solving large-scale optimization problems where the objective function is both highly-non linear and computationally costly to evaluate. Common examples are encountered in economics and finance for solving decision problems with multiple controls or estimating structural models by a simulated method of moments. The key novelty is that the algorithm optimally trades off the exploration for multiple local optima with the convergence toward a precise solution by conducting a search on endogenous clusters. To do so, the algorithm operates in two phases. The first phase maximizes the number of disjoint search areas or clusters in order to cover as many local optima as possible given an initial sample of random draws. Each cluster is formed of multiple ball-shaped search neighborhoods. The second phase shrinks the size of each cluster by reducing the radius of their associated neighborhoods. The size reduction is chosen such that the search area is spatially bounded by candidate points previously found sub-optimal. 

This figure illustrates the mechanism of the cluster search algorithm and compares it with some alternatives. Each algorithm draws 40 points per iteration to locate the global maximum of a two-dimensional inverse paraboloid with several local optima. This function reaches its maximum of 100 at the coordinates (16.6805, 0.3652). The current maximum found by each algorithm is marked with a yellow circle. 

In this example, the cluster search algorithm identifies up to six disjoint areas to explore. After a few iterations, the number of clusters gradually decreases and their radius collapses. During this process, the density of points in the active search area rises which improves the convergence toward the exact solution. 

 In contrast, the genetic algorithm converges slowly to the global optimum because the combination of two local optima often leads to a sub-optimal candidate. On the other hand, the particle swarm is attracted to local optima at the initial stage, which leads to an inefficient number of function evaluations. Lastly, the Monte Carlo method robustly finds the global maximum but is inefficient at converging toward a precise value because no information is shared between the different candidates. 

The cluster search algorithm can be interpreted as a metaheuristic that mimics the expansion phase and decay phase of bacterial growth in a petri dish. The clustering component or expansion phase ensures that disjoint search neighborhoods are evenly spread out around the local optima while the shrinking component or decay phase increases the density of points in the search area to obtain a precise solution. This algorithm is promising for optimization problems where the objective function is noisy and difficult to evaluate.