Kernel Analog Forecasting of Intraseasonal Oscillations

Wednesday, 20 April 2016: 11:00 AM
Miramar 1 & 2 (The Condado Hilton Plaza)
Dimitrios Giannakis, New York University, New York, NY; and R. Alexander, E. Szekely, and Z. Zhao

We discuss nonparametric forecasting of the MJO and the boreal summer intraseasonal oscillation (BSISO) in brightness temperature data. First, we extract eigenmodes representing the dominant ISOs from the CLAUS archive using the nonlinear Laplacian spectral analysis (NLSA) algorithm. A key advantage of NLSA is that it requires no preprocessing such as bandpass filtering or seasonal partitioning of the input data, enabling simultaneous recovery of the dominant ISOs and other patterns influenced by or influencing ISOs. In particular, the MJO and BSISO naturally emerge as distinct families of modes exhibiting strong seasonality and eastward- and poleward-propagating patterns, respectively. We then build nonparametric forecast models for these modes using a recently developed kernel analog forecasting technique. This method modifies the traditional analog forecasting approach to create weighted ensemble analog forecasts with weights computed from local similarity kernels. We use this approach to perform hindcasts of the NLSA-based MJO and BSISO modes, over the period July 2006 to February 2009. These hindcasts have statistically significant skill (based on a 0.6 correlation coefficient threshold) for 40-50 day leads.

- Indicates paper has been withdrawn from meeting
- Submission entered in competition