The Balanced Subgraph problem (edge deletion variant) asks for a 2-coloring of a graph that minimizes the inconsistencies with given edge labels. It has applications in social networks, systems biology, and integrated circuit design. We present an exact algorithm for Balanced Subgraph based on a combination of data reduction rules and a fixed-parameter algorithm. The data reduction is based on finding small separators and a novel gadget construction scheme. The fixedparameter algorithm is based on iterative compression with a very effective heuristic speedup. Our implementation can solve biological realworld instances exactly for which previously only approximations [DasGupta et al., WEA 2006] were known.