GRID

GRID is built on continuous GFlowNets and introduces two key designs for high-dimensional, sequential macro-financial data: a mixture-policy architecture and plausibility constraints.

The mixture-policy architecture assigns a different distribution family to each variable to improve modeling of complex financial data.

The plausibility constraints ensure that generated trajectories are realistic and feasible in a financial sense, avoiding unrealistic extreme scenarios.

Experiments on three real-world crisis datasets show that GRID outperforms RL and MCMC baselines in scenario diversity, realism, and tail-risk coverage.

GFlowNets sample proportionally rather than maximizing reward, so even when the out-of-sample accuracy of the reward proxy is low, the model maintains broad exploration and is robust to an imperfect reward proxy.

The study conducted experiments on three real-world crises, and ablation experiments explored the effects of the oracle architecture and the choice of distribution family on the behavior of continuous GFlowNets.

How to further optimize the mixture-policy architecture and plausibility constraints when the reward proxy is imperfect remains an open question.

The experiments cover only three crisis datasets, and broader market conditions and asset classes remain to be validated.