Prediction of northern summer low‐frequency circulation using a high‐order vector auto‐regressive model

The atmospheric extra-tropical flow is characterized as chaotic motions that are sensitive to initial conditions and thus is merely predictable by operational weather forecast models 2 weeks in advance given the current observational and modelling accuracy (e.g., Lorenz 1969; Leith 1971; Tribbia and Baumhefner 2004). On the other hand, the seasonal and inter-annual variations in the extra-tropical flow are subject to ocean boundary conditions and captured by climate models reasonably well (Kumar et al. 1996; Shukla 1998; Shukla et al. 2000; Goddard et al. 2001; Smith et al. 2012). The predictability in the intermediate range, between approximately 10 days and up to a month, is challenging for both weather forecast and climate models, due to the lack of significant influence from either the initial or the boundary conditions. However, low frequency (>10 days or longer than synoptic scales but shorter than seasonal scales) variability tends to dominate the total variability of the sub-seasonal tropospheric circulation (Blackmon 1976) for Northern Hemisphere (NH) winter and summer. Blackmon (1976) showed further that the structure of the sub-seasonal low-frequency fluctuations are dominated by large scale planetary waves, as compared to the high frequency (<10 days) fluctuations, which are dominated by synoptic waves. There has been much less work done on the summer low-frequency variability and its predictability, although there is increased threat of persistent summer circulation anomalies associated with heat waves, floods and droughts throughout the world (e.g., Beniston 2004; Dole et al. 2011; Barriopedro et al. 2011; Coumou and Rahmstorf 2012). Similar to northern winter, the magnitude of the northern summer low frequency variability is generally much larger than that of the high frequency component in both the lower Abstract A data-driven, high-order vector auto-regressive (VAR) model is evaluated for predicting the Northern Hemisphere summer time (May through September) low frequency (>10 days or so) variability. The VAR model is suitable for linear stationary time series, similar to the commonly used linear inverse model (LIM), with additional temporal information incorporated to improve forecast skill. The intraseasonal forecast skill of the 250/750 hPa streamfunction is investigated using observational data since 1979, which shows significant improvements in highorder VAR models than the first-order model LIM. Furthermore, the tropical diabatic heating is found to significantly improve the forecast skill of the atmospheric low frequency circulation when included in the VAR model. The forecast skill of 250 hPa streamfunction at Arabian Peninsula is particularly enhanced for up to 5 weeks lead-time through circumglobal wave propagation associated with the persistent tropical eastern Pacific and equatorial Atlantic heating anomalies and the intraseasonal evolution of the tropical Indian Ocean and western Pacific heating anomalies.