Masashi Egi

Automatically Tuned Sparse Eigensolvers

Automatic tuning methods for matrix computations have been investigated since the 1990s. These investigations have revealed that sparse matrix solvers require higher matrix software semantics tuning and more adaptive run-time parameter tuning. The goal of this chapter is to propose a new method of run-time automatic tuning for sparse matrix eigensolvers. In the proposed method, a statistical criterion that captures fluctuations of convergence is introduced for run-time automatic tuning. The results of numerical experiments demonstrate that, in the most effective cases, the proposed method achieves approximately five times faster performance. These results imply that automatic tuning with higher semantics is critical for high performance.

Random Matrix Theory Applied to Portfolio Optimization in Japanese Stock Market

We examined the effectiveness of random matrix theory applied to portfolio optimization using Japanese stock market data. We carried out 48 back tests for different historical periods and confirmed that it was possible to drastically improve the accuracy of portfolio risk evaluation using random matrix theory.

Method of Analyzing Weather Derivatives Based on Long-range Weather Forecasts

We examined the effectiveness of long-range weather forecasts applied to analyze weather derivatives. We carried out 651 back tests for different historical periods and confirmed that the accuracy of evaluating the risk of weather derivatives could be drastically improved by using long-range weather forecasts.