A spatio-temporal graph neural network model for predicting economic impact of sports events and analyzing regional synergy effects

Major sporting events reshape economic activity in host cities and their surroundings through channels that input-output, computable general equilibrium, and spatial econometric tools capture only in part, because each of these tools fixes multipliers or adjacency rules before it sees the data. This study develops ST-GAT, a spatio-temporal graph attention network that forecasts a monthly regional activity index around event cycles and learns inter-regional coupling from data, with gravity-based priors entering as side information rather than as fixed weights. The forecasting target combines card-transaction consumption with night-time light intensity at the prefecture-month level; the model therefore predicts event-period activity and does not estimate welfare or causal effects. Its architecture couples attention-weighted graph convolution with multi-scale gated temporal convolution and a synergy-regularized objective, and a downstream routine turns the learned attention into a directional synergy matrix. We evaluate the framework on eight Chinese host environments, which involve fourteen host or co-host prefectures among 96 prefectural nodes observed monthly from 2014 to 2024. Averaged over five random seeds, ST-GAT lowers RMSE by approximately 25% relative to STGCN at a six-month horizon. The margin narrows against DCRNN, ASTGCN, and Graph WaveNet: the Graph WaveNet comparison ( p = 0.032) survives a Holm adjustment across seven baselines but not a Bonferroni bound, so we treat it as provisional. The advantage widens at longer horizons and persists when the pandemic window is excluded. Three exploratory patterns follow, each a property of the fitted model rather than an estimated effect. Legacy half-life rises monotonically with pre-event fixed-asset investment across the eight environments, although part of this association may be mechanical because investment is also a model input. Suzhou, Foshan, and Langfang act as predictive hubs in the attention network without hosting an event. Perturbing policy-alignment information moves the synergy output roughly twice as far as perturbing transport accessibility, a contrast that motivates, but cannot substitute for, a causal test. We therefore present ST-GAT as a forecasting and hypothesis-screening tool that indicates which co-hosting, industrial, and infrastructure relationships deserve dedicated causal evaluation.