Further evidence on the predictability of long-term solar activity: Tests using new data reconstructions

It has been understood for some time that radionuclides can be used to reconstruct long-term solar activity. Recently, new reconstructions have become available, spanning the interval from 6755 BC to the late nineteenth century. This can be spliced together with actual sunspot data to create a continuous series for sunspots at a decadal resolution. While the timing of minima and maxima are comparable to prior estimates, the new reconstruction implies higher overall solar activity and less trending. Forecasting experiments are run over horizons of 10–80 years using regressions, combinations of regressions and neural networks, and support vector machines. Despite the differences between the new and prior reconstructions, the findings are essentially the same as in earlier studies. Long-term solar variations do not exhibit regular periodicities. While cycles can be identified using spectral analysis, these are not confirmed in the time domain. The autocorrelation function and regressions on lags show no evidence of cycles at lower frequencies. In the experiments run here, the models generally predict well up to 30 years. At 40 years, forecast accuracy begins to deteriorate. At horizons of 50–80 years, the models either converge to the mean of the data or replicate recent fluctuations with a lag. Based on the results, solar activity is essentially stochastic and cannot be predicted by time series or artificial intelligence models beyond horizons of a few decades.