Enhanced earnings revision strategy: a Bayesian Machine Learning approach

This thesis investigates potential enhancements of a generic analyst earnings revision strategy that consists in buying stocks experiencing a high number of upward analyst earnings revisions and selling the stocks experiencing a high number of downward analyst earnings revisions. The strategy proved to be profitable in the past leading to positive out-of-sample abnormal returns, even when controlling for the common financial factors. We first show that the performance of the consensus revision strategy can be enhanced in the first 20 years of the sample by assigning different weights to the revisions based on their informativeness captured by some analyst/revision characteristics. These results cannot be explained by market, size, value, profitability, investment and momentum exposures and are robust across regions. After 2011, the enhanced strategy stopped adding value to the equally weighted one, probably due to the diffusion of popular enhanced revision models in the financial industry. We then construct a novel Bayesian machine-learning algorithm to relate the analyst revisions with the future stock price changes in an attempt to improve the strategy performance. We show that this methodology leads to a positive and significant alpha in the US, but the signal captured by the algorithm is weak compared to the noise, mainly due to the high dimensionality and sparsity of the input matrix. We confirm this finding by looking at the results of the model outside the US universe where no significant improvement of the equally weighted revision strategy is found.