Information - Theoretic Mask Analysis of Rainfall Time Series Data

This study explores an information-theoretic/log-linear approach to multivariate time series analysis. The method is applied to daily rainfall data (4 sites, 9 years), originally quantitative but here treated as dichotomous. The analysis ascertains which lagged variables are most predictive of future rainfall and how season can be optimally de ned as an auxiliary predicting parameter. Call the rainfall variables at the four sites A...D, and collectively, Z, the lagged site variables at t-1, E...H, at t-2, I...L, etc., and the seasonal parameter, S. The best model, reducing the Shannon uncertainty, u(Z), by 22%, is HGFSJK Z, where the independent variables, H through K, are given in the order of their predictive power and S is dichotomous with unequal winter and summer lengths.