Resumo

Crises can be defined as extreme manifestations of the interactions between the financial sector and the real economy. Their origins can be domestic or external, coming from public or private sectors. Financial crises severely affect economic activity and can trigger periods of recession, spreading to various regions or sectors in a process called contagion. Specific sectors can be major crisis propagators. This was the case of the technology sector during the dot-com bubble, and the financial sectors, such as banking and insurance, during the subprime crisis.

Are banking and insurance sectors the main crises propagators? We aim to model sectorial interdependence of US economy through the Dynamic Bayesian Networks technique. By treating dependence in a probabilistic -not merely correlational- way, we follow the evolution of sectorial dependency structure dynamically over time. We model a great variety of sectors that represent the US economy, while previous literature either studied contagion among countries or a small group of sectors. As a longitudinal study, we can analyze several crises: from dot-com bubble burst to current Covid-19 pandemic.

Some authors define contagion as the spread of market shocks, with mostly negative consequences, observed through co-movements in exchange rates, stock prices, increases in sovereign risks and in capital flows. This type of shock transmission can trigger contagion effect. There are several studies with different methods for contagion analysis. Possibly, the most common way to evaluate the contagion effect is to analyze stock exchange indices returns. Our innovation is to model the dependence structure directly using Dynamic Bayesian Networks, through a multivariate time series system.

Bayesian Networks are sophisticated machine learning modeling technique that represent the probabilistic relationships among many variables and allow making probabilistic inference. It provides an innovation to contagion studies as Bayesian Inference assumes that a parameter (correlation between two sectors) is a random variable. When performing hypothesis testing, classical inference procedures tend to use p-value as its main significance measure, which is a problem because p-value is not a probability measure. This issue is addressed by using a different significance measure: the q-value.

The results show evidence that banking sector presents itself as a frequent crises propagator towards insurance, but not the other way around, in accordance with literature. The oil&gas and real estate sectors predominate as the main propagators throughout the period, which had not been addressed yet in other studies. Finally, this study was a pioneer in modeling the current configuration of the sectorial network during the Covid-19 pandemic, which proved to be identical to subprime’s network configuration. Both were the greatest financial crises during the analyzed period.

The use of a sophisticated machine learning modeling technique on sectorial contagion has brought satisfactory results. Most of the relationships captured by the Dynamic Bayesian Networks finds support on the literature, showing that future economic analyses can be enriched by using this instrument to capture sectorial dependence relationships. Furthermore, this technique allows us to follow the dependence structure evolution over time, not only by verifying the relations’ appearance or vanishing, but also observing the changes in correlation magnitudes when comparing periods.

Carvalho, J. V. F., & Chiann, C. (2013). Redes bayesianas: Um método para avaliação de interdependência e contágio em séries temporais multivariadas. Revista Brasileira de Economia, 67(2), 227–243. Collet, J., & Ielpo, F. (2018). Sector spillovers in credit markets. Journal of Banking and Finance, 94, 267–278. Kaserer, C., & Klein, C. (2019). Systemic Risk in Financial Markets: How Systemically Important Are Insurers? Journal of Risk and Insurance, 86(3). Pino, G., & Sharma, S. C. (2019). On the Contagion Effect in the US Banking Sector. Journal of Money, Credit and Banking, 51(1), 261–280.