Interannual variations in catchment streamflow represent an integrated response to anomalies in regional moisture transport and atmospheric circulations and are ultimately linked to large-scale climate oscillations. This study investigates the relationship between Taiwan’s long-term summertime (July to September, JAS) streamflow and manifold teleconnection patterns. Lagged correlation analysis is conducted to calculate how JAS streamflow data derived at 28 upstream and 13 downstream gauges in Taiwan correlate with 14 teleconnection indices in the current or preceding seasons. Of the many indices, the West-Pacific and Pacific-Japan (PJ) patterns, both of which play a critical role in determining cyclonic activity in the western North Pacific basin, exhibit the highest concurrent correlations (most significant <i>r</i> = 0.50) with the JAS flows in Taiwan. Alternately, the Quasi-Biennial Oscillation averaged over the period from the previous October to June of the current year is significantly correlated with the JAS flows (most significant <i>r</i> = −0.66), indicating some forecasting utility. By further examining the correlation results using a 20-year moving window, peculiar temporal variations and possible climate regime shifts (CRSs) can be revealed. To identify suspicious, abrupt changes in the correlation, a CRS test is employed. The late 1970s and 1990s are identified as two significant change points, and during the intermediate period, a marked in-phase relationship (<i>r</i> > 0.8) between Taiwan's streamflow and the PJ index is observed. Linear regression models that incorporate the climate indices into streamflow prediction are found to provide reasonable prediction skill in general, and the models are then used to illustrate the dramatic variations in prediction skill from the pre- to post-regime shift epoch. It is verified that the two shifts are in concordance with the alteration of large-scale circulations in the Pacific basin. The changes in pattern correlation and composite maps before and after the change point are discussed, and our results suggest that empirical forecasting techniques should take into account the effect of CRSs on predictor screening.