Financial time series forecasting is challenging due to limited sample size, correlated samples, low signal strengths, among others. Additional information with knowledge graphs (KGs) can allow for improved prediction and decision making. In this work, we explore a framework GregNets for jointly learning forecasting models and correlations structures that exploit graph connectivity from KGs. We propose novel regularizers based on KG relations to guide estimation of correlation structure. We develop a pseudo-likelihood layer that can learn the error residual structure for any multivariate time-series forecasting architecture in deep learning APIs (e.g. Tensorflow). We evaluate our modeling and algorithmic proposals in small sample regimes in real-world financial markets with two types of KGs. Our empirical results demonstrate sparser connectivity structures, runtime improvements and high-quality predictions.