A graph based approach is introduced that dynamically identifies teleconnections in climate data. A network is generated by correlating anomaly time series of all pairs of locations. The approach identifies regions that are connected by high positive correlation and that share a neighborhood of negative correlations, thus forming two ends of a dipole.

This shared nearest neighbor graph approach is applied to variables such as sea-level pressure, geopotential height at different levels and surface temperature. The resulting data driven climate indices generated from this algorithm can vary over time, but still correlate well with the static indices derived by the fixed locations proposed by e.g. the Climate Prediction Center. The dynamic indices outperform the static indices in terms of capturing the impact on global temperature and precipitation pattern.

Another salient point of this approach is that it can generate a single snapshot picture of all the global dipole connections in a given dataset. This allows an intercomparison of different data sets, e.g. climate model simulations. Given the importance of teleconnections and the importance of model simulations in understanding the impact of climate and climate change, this methodology has the potential to provide valuable insights into future climate projections. This research has been funded by the NSF project 1029711.