Thiago de Oliveira Souza

Ph.D. Candidate in Finance and Economics

             Queen Mary, University of London — School of Economics and Finance

             Université libre de Bruxelles, ULB — Solvay Brussels School of Economics and              Management / ECARES


I will be available for interviews at the ASSA meetings in January 6-8, 2012 in Chicago, IL and in the RES meetings in January 21-22, 2012 in London - UK.



Job market paper


Strategic asset allocation with heterogeneous beliefs

Abstract: In this paper, I show how the presence of agents with heterogeneous beliefs generates the price trends observed in the financial markets. I develop an asset pricing model in which agents have long horizon objectives, based on a stream of consumption. Each agent chooses a forecasting model and maximises a recursive utility function. The choice of the forecasting model in each period determines the agent type. However their types change over time according to the relative performance of the forecasting models. This happens because agents have an incentive to adopt the forecasting model with the best performance in the previous period to coordinate with the market. I estimate the asset pricing model using data on the international stock markets. The exercise shows that especially for very risk averse individuals, the accounting for the intertemporal hedging demand is crucial.







Research papers (in progress)


Regularized minimum-variance portfolios using asset group information

(co-authored with Marcelo Fernandes and Guilherme Rocha)

Abstract: Regularized minimum-variance portfolios obtained by restricting the norm of the portfolio-weights vector to be smaller than a certain value often have better out-of-sample performance than unrestricted ones. The portfolios' performances are further increased by adding an algorithm that takes into account the group structure of the assets in the optimisation procedure. We present and compare two alternatives to include the group structure based on a framework originally applied to improve the finite sample properties of the coefficients obtained by ordinary least squares. We compare empirically the out-of-sample performance of this new class of portfolios with their counterparts that do not use the group information. The Sharpe ratios are statistically larger than the previous ones depending on the choice of the threshold and which norm to restrict, while there is no significant difference otherwise. The new portfolios often outperform the previous ones in terms of variance.


Forecasting investment-grade credit-spreads – A regularized approach

Abstract: It is common for banks to have liabilities attached to the Treasury's rate and assets attached to a corporate rate. A change in the difference between these rates (i.e., a change in the credit-spread) impacts the banks' balance sheet. In order to forecast this risk, I propose the use of (very) short estimation windows using the lasso estimation. The lasso shrinks some of the estimated coefficients to zero, improving their finite sample performance also allowing the use of smaller estimation windows. I compare the out-of-sample performance of several credit-spread forecasting models for each investment-grade credit-rating in the period between 2000 and 2011. Considering the 6 and 12 months forecasts of AAA-rated credit-spreads, the historical average outperforms, in terms of mean absolute prediction error, the Martingale and several other forecasting models based on the shape (level, slope and curvature) of the risky and risk-free yield curves, and based also on the spot, forward and average past yields. Considering all other credit-ratings, the forecasts given by the lasso tend to outperform those based on long estimation windows.