Are financial ratios relevant for trading credit risk? Evidence from the CDS market

Abstract

We propose a combination of LASSO with panel-consistent estimation methods to investigate whether financial ratios are used in the decision-making process of CDS traders. Our results indicate that financial statement information does play a role in all the trading horizons surrounding the announcement date and the corresponding styles. These include pro-active analysts trying to predict quarterly results, news traders reacting to unanticipated information and value traders who fine-tune their estimates and act accordingly at a later stage. Our findings also suggest that CDS traders respond asymmetrically to financial ratio updates of different sign and intensity.

Appendices

Appendix 1: Other LASSO applications

LASSO is especially helpful in cases where the number of potential predictors is large relative to the number of observations in a dataset. By adapting the magnitude of the penalty, by means of calibrating the tuning constant $\lambda$, the number of non-zero regressors is determined. Although the ridge regression follows a similar rationale in contracting parameter estimates by minimizing $\lambda \sum \nolimits _{j=1}^p b_j^2$, it is not as efficient in excluding many regressors. This flaw of ridge regression stems from the fact that for smaller values of the coefficients ($b_j )$, $b_j^2$ is considerably smaller than $\left| {b_j } \right|$.

Not surprisingly, LASSO applications abound in the area of biogenetics, since the number of potential genetic predictors that are related with the occurrence of a disease far exceeds the number of observations in the samples examined. Wu et al. (2009) assess the performance of LASSO logistic regression in selecting among a large number of single nucleotide polymorphisms (SNPs) for a disease gene mapping. By testing their model both on simulated and on real data, they not only find the same determinants with previous studies for coeliac disease but also unveil the potential interactions among various SNPs. Other studies from the field of biogenetics that use LASSO include Shi et al. (2008), who aim in identifying significant patters of various risk factors among a wide range of potential candidates, and Park and Hastie (2008), who examine the interaction of various genes in the occurrence of a disease.

LASSO has been also used in economic time series forecasting. Bai and Ng (2008) utilize LASSO as a soft thresholding means to identify the top-ranked predictors, that is, the “targeted predictors” to be used in the formation of the factors under the principal component analysis. They assess sundry methods in predicting inflation across various horizons and over several samples. The selection of inflation over other macroeconomic variables is due to its heightened importance for decision making both by private investors and by government bodies. On top of that, forecasting inflation series constitutes not an easy task. The use of LASSO enables Bai and Ng (2008) to avoid uninformative predictors, while at the same time refrain from selecting highly correlated ones. Their results indicate that the LARS-LASSO EN estimator outperforms the remaining methods in the selection of the “best” predictors for forecasting. Furthermore, Wang and Zhu (2010a, b) also utilize a two-step kernel learning methodology in financial forecasting. Kernel learning is attained via the use of L 1-norm regularization approach. Another application to finance is found in Lobo et al. (2007), who use the L 1 norm in portfolio optimization. Particularly, they address cases where non-linear fixed transaction costs are present, and hence convex optimization methods cannot be used to directly solve the problem.

In industry and manufacturing, LASSO has been used by Zou et al. (2012) to establish a new approach for monitoring general multivariate linear profiles, and test their approach by using data from a logistics service. Essentially, they use observed profile data and use their suggested control chart so as to identify the shift direction. Furthermore, Kim et al. (2014) develop an interval based classifier that can be employed in diagnosing problematic conditions in manufacturing, thus, preempting the production of defective products.

Applications of LASSO also abound in various fields, other than the ones mentioned above. For example, Verbesselta et al. (2009) utilize LASSO to identify the best measures whose changes are able to predict tree mortality, by examining images of the respective areas taken by a satellite. Similarly, Wang and Zhu (2010a, b) utilize the L 1-norm penalty in image denoising. This approach succeeds in removing only the noisy pixels, whereas at the same time the non-noisy pixels are kept intact. While, Kamarianakis et al. (2012) employ LASSO in real time forecasting of road traffic. In particular, they use LASSO for simultaneous model selection and coefficient calibration for each location.

Last but not least, LASSO has been used in data cleaning. Au et al. (2010) propose a generalized irregularity enlightenment structure so as to capture numerous irregularities in large datasets. In doing so, they use LASSO so as to opt for the significant characteristics occurring in large datasets. All in all, since LASSO succeeds in correctly identifying the most relevant predictors for explaining the dependent variable, its applications can be extended in almost all research fields for which a relation between a dependent variable and an extended series of regressors has to be investigated.

Appendix 2: Trace plots of LASSO coefficients

The trace plots (1–6) presented below illustrate how the number of the regressor coefficients shrinking to zero increases, as lambda increases in value itself. These graphs correspond to the model specifications under Hypotheses 1, 2a and 2b. The dashed vertical green line (the second one on the left side of the plot) reflects the number of non-zero regressor, when the Mean Square Error is minimized. In effect it denotes the size of the optimal model picked by our LASSO application.

• 1. Trace plot on coefficients fit by LASSO on residuals$_{q }$(unexplained part of $\Delta \hbox {CDS}_{q})$

  

• 2. Trace plot on coefficients fit by LASSO on residuals$_{30-1 }$(unexplained part of $\Delta \hbox {CDS}_{30-1})$

  

• 3. Trace Plot on coefficients fit by LASSO on residuals$_{1-1 }$(unexplained part of $\Delta \hbox {CDS}_{1-1})$

  

• 4. Trace plot on coefficients fit by LASSO on residuals$_{1-7}$ (unexplained part of $\Delta \hbox {CDS}_{1-7})$

  

• 5. Trace plot on coefficients fit by LASSO on residuals$_{7-14 }$(unexplained part of $\Delta \hbox {CDS}_{7-14})$

  

• 6. Trace plot on coefficients fit by LASSO on residuals$_{14-30 }$(unexplained part of $\Delta \hbox {CDS}_{14-30})$