EventConnector uses temporal graphs to improve event forecasting by 6.87%

A team of researchers (Zijie Lei, Haofei Yu, Ge Liu, Jiaxuan You) has published EventConnector, a new approach to understanding relationships between real-world events by modeling them as a temporal graph. Current methods rely on semantic similarity or global time-series alignment, missing the transient, directional dependencies that often govern event correlations. EventConnector constructs a graph capturing localized co-fluctuations and lead-lag patterns directly from event time-series data, enabling more accurate retrieval of related events for downstream tasks like forecasting.

Building on the graph, the team introduces EC-Fusion, an adaptive retrieval mechanism that combines EventConnector's graph-based scores with a complementary Granger-causal signal using a graph-quality-aware mixing weight. Tested on two real-world prediction market benchmarks (Polymarket and Kalshi) across nine forecasting architectures and three random seeds, EC-Fusion was the best non-oracle retrieval method in 17 out of 18 model–dataset combinations. It achieved a statistically significant 6.87% average RMSE reduction (up to 10.86%) over the strongest comparable baseline (p < 0.01 after Holm–Bonferroni correction). These results show that temporally grounded graph modeling, augmented with causal signal fusion, can capture latent event relationships far beyond what semantic similarity or traditional alignment techniques offer.