Statistical methods for scale-invariant and multifractal stochastic processes.

Abstract

This thesis focuses on stochastic modeling, and statistical methods, in finance and in climate science. Two financial markets, short-term interest rates and electricity prices, are analyzed. We find that the evidence of mean reversion in short-term interest rates is week, while the “log-returns” of electricity prices have significant anti-correlations. More importantly, empirical analyses confirm the multifractal nature of these financial markets, and we propose multifractal models that incorporate the specific conditional mean reversion and level dependence. A second topic in the thesis is the analysis of regional (5◦ × 5◦ and 2◦ × 2◦ latitude- longitude) globally gridded surface temperature series for the time period 1900-2014, with respect to a linear trend and long-range dependence. We find statistically significant trends in most regions. However, we also demonstrate that the existence of a second scaling regime on decadal time scales will have an impact on trend detection. The last main result is an approximative maximum likelihood (ML) method for the log- normal multifractal random walk. It is shown that the ML method has applications beyond parameter estimation, and can for instance be used to compute various risk measures in financial markets.