We propose a causal-graph forecasting framework for early cyberattack prediction that integrates causal network modeling with dynamic bipartite graph–based time series forecasting. Our method captures both the cause–effect structure of network traffic and the co-evolving, non-stationary dynamics of attack-related signals. Network entities are represented as nodes, while edges encode directional causal influences inferred from traffic flows. To model evolving dependencies, we approximate each traffic series over contiguous intervals with parametric functions and construct temporal bipartite graphs that link these models to time series, revealing shared latent patterns and context-driven relationships. This hybrid representation enables adaptive model selection and accurate forecasting of malicious behavior before it escalates. Experiments on benchmark of network traces show that our approach delivers early, interpretable, and highly accurate cyberattack predictions.