Seminarios

This "data sparsity" makes the utilization of current modern estimation and prediction techniques, such as PCA and Kalmann Filter models, difficult to implement as those rely heavily on the existence of abundant market data. This is especially important in emerging markets such as LATAM countries, in which the implementation of traditional methods typically presents large margin of errors and it is not as reliable as in more liquid markets.

In this paper, we propose the use of Non Negative Matrix Factorization (NNMF) techniques to obtain latent factors that represent most of the information content of the Yield curve. In contrast with similar traditional methods such as PCA, NNMF is able to estimate these factors with high precision even when dealing with highly sparse datasets. Moreover, as the term structure is represented as latent Factors that are better estimated using NNMF, it is possible to increase the forecasting accuracy of such factors using state of the art neural network architectures, such as, RNN and LTSM networks.