Dramatic historical turning-points reduce social scientists' ability to use past patterns to predict subsequent events. We trace this ability through the study of more than 10,000 Reuters news stories on Israeli-Palestinian interactions from the period 1979 to 2005, which have been systematically coded to provide a rough summary of positive/negative interactions on a daily basis. Using rolling windows, we show that the accuracy of out-of-sample forecasting fluctuates tremendously during this period, and that the errors of these forecasts are particularly prominent just after major events such as the first and second intifadas. Over the months following these dramatic moments, as new routines of interaction develop, forecasting errors decline. These findings are produced consistently regardless of which time-series methods is used: vector autoregression (VAR), autoregressive integrated moving average (ARIMA), and seemingly unrelated regressions (SUREG). This analysis is the first of its kind -- the field of forecasting studies, including other analyses of Israeli-Palestinian interactions, focuses on overall model fit, rather than variation in forecasting errors. We propose that this new approach speaks to the limits of social scientific prediction at moments when existing patterns of behavior have changed so dramatically that people cannot know what comes next. Moments of routine behavior may be quite well predicted by previous behavior, but moments of non-routine behavior may lie beyond the powers of prediction.