“…Since our all-season LIM does not perform equally well at all amplitudes and phases of the MJO, it is possible that an MJO behavior-, amplitude-or phase-dependent LIM, resulting in a piecewise-stationary linear model, could yield large increases in MJO forecast skill, provided any hindcast stitching procedures do not produce amplifying errors and that the observed model skill deficiencies are not merely manifestations of the intrinsic predictability of the tropical atmosphere itself. Lastly, it is well known that linear stochastic models with Gaussian white noise produce Gaussian forecasts, whereas high-frequency climate statistics are markedly non-Gaussian (Perron andSura 2013, Cavanaugh andShen 2014). An alternative mode reduction strategy, strategic choice of additional independent observable inputs containing MJO precursor and dynamics information, and/or augmentation of the system with correlated additive and multiplicative noise, suggested by Sura et al (2005), may improve forecasts by accounting for more aspects of variability, while maintaining the simplicity of a linear stochastic framework.…”