We introduce a machine learning approach to detect value-relevant foreign information by modeling stock-specific, time-varying relationships between foreign signals and stock returns. A long-short portfolio exploiting foreign signals generates 12% annual abnormal returns. Return predictability is more pronounced among domestic firms, those with low foreign institutional ownership, and during periods of low media coverage and high model agreement. Notably, performance concentrates on the long side, enabling cost-effective long-only implementations. Signal importance analysis reveals our algorithms detect valuable signals by tracking key international trading partners, monitoring shifts in monetary policy and political stability, and leveraging information particularly from under-covered emerging markets.
