Back to search

FINANSMARK-Finansmarkedet

Statistical modelling and inference for (high-dimensional) financial data

Alternative title: Statistisk modellering og inferens for (høy-dimensjonale) finansdata

Awarded: NOK 0.60 mill.

Project Manager:

Project Number:

309218

Project Period:

2020 - 2023

Funding received from:

Subject Fields:

During the last 10 years, after the economic crisis in 2008, a wide class of models has been introduced to model financial data features, aiming at forecasting the risk of defaults, future returns, etc, and at minimizing the forecasting error. This is typically done by introducing models able to capture the temporal dynamics of the processes generating the data. Moreover, the huge amount of data currently available, on the one side, is a rich source of information, but, on the side, require ad-hoc methodologies able to address the curse of dimensionality and the computational burden involved in any "big-data" modelling. This project has introduced a class of models, into the wide framework of models known as hidden or latent Markov models, to better characterize financial data in terms of crisis vs. non-crisis periods, and to capture possible sources of contagion into specific financial sectors. Of course, this is done by taking into account macro-economic indicators that may be used to increase the accuracy of the predictions/estimates, as they are often used of proxies of the economic health/status. As the data under analysis are massive, proper models to reduce the dimension of the problem as well as computationally feasible algorithms are developed. This ensures the application of the proposed novel methods by non-experts and their usefulness in a wide range of financial application, where "big-data" are not easy to deal with. The project group has had several articles published in academic journals, particularly a recent work in the "Journal of the Royal Statistical Society", entitled "Modelling clusters of corporate defaults: Regime-switching models significantly reduce the contagion source". In this article, a regime model is combined with a count time series model, and we show that the process of corporate defaults is more dynamic than previously believed. Moreover, the contagion effect, that current defaults affect the probability of other firms defaulting in the future, is reduced compared to models without regime-switching, and is only present in one regime. Furthermore, the project has publised several articles related to the estimation of Hidden Markov models, and related models.

Both sub-projects address important issues within the financial service industry. For sub-project a) a possible output from the study of the clustering of corporate defaults could be that more sophisticated class of models may be needed to model bankruptcies. Adjusting current credit risk models to comply with these specifications would not only be relevant for internal risk assessment, but also for external supervision of financial institutions This is shown to be the case in the research done in the project. For sub-project b), the introduction of the new risk measures can mprove the estimation of tail risks. This would be of benefit for many financial institutions, including banks and insurances.

One often encounters non-linear dependencies and non-Gaussianity in the joint distribution of stochastic variables in finance. This typically arises as easily recognized features of asset returns observed in financial markets. In this project we will extend both the theory and applications of non-linear stochastic models in two problem areas of finance, which are the a) corporate default modelling by non-linear integer-valued time series with regimes and b) new risk measures in high-dimensional multivariate settings. In sub-project a) we will further develop the theory and applications regarding non-linear integer time series models by allowing for time-varying parameters in these models. The motivation for such an extension is the modelling and forecasting corporate defaults. A stylized fact of these defaults is their tendency to cluster in time. This default clustering phenomenon has been explored in the financial literature. We will contribute to this field by proposing a new class of non-linear integer-valued time series models with regimes. In this way, we will achieve a more realistic modelling of corporate default counts and will be able to provide improved insight into their behaviour. In sub-project b), our main interest is to introduce and estimate the recently proposed risk measures conditional value at risk (CoVaR) and conditional expected shortfall (CoES) for the class of parsimonious contaminated hidden semi-Markov models. This approach allows for non-linear dependence between asset returns.

Funding scheme:

FINANSMARK-Finansmarkedet