Beyond the Row: Sequence-Aware Diffusion for Rare-Event Financial Trajectories

Rare financial events—defaults, fraud—are properties of a trajectory, not a row, yet financial-diffusion oversampling has so far been built almost entirely on row- independent generators that model pθ (xi,t | yi) per event. We instantiate the sequence-aware alternative, pθ (Xi | yi), as a class-conditional diffusion model with an interchangeable sequence-mixing block (bidirectional attention or a selec- tive state-space scan), and evaluate it against row-independent diffusion, classical oversampling, and no augmentation on three financial-trajectory datasets with as few as 10–20 real positive entities. Against the default configuration of the row- independent baseline, sequence-aware diffusion reduces temporal fidelity error on two of three datasets—0.159 to 0.108 on a Czech loan panel (32%) and 0.126 to 0.105 on a Taiwanese credit panel (17%)—and its fidelity does not degrade as more positive labels become available. This advantage is specific to that con- figuration: a capacity- and guidance-matched baseline closes it, and at a doubled training budget the row-independent baseline is more faithful on all three datasets. Fidelity differences, in either direction, do not reliably convert into downstream utility: no augmentation method reliably beats training on the untouched real data, and across 225 (method, scarcity, seed) cells the Spearman correlation between temporal fidelity and PR-AUC improvement is −0.044 (95% CI [−0.29, 0.21]) under an order-blind XGBoost evaluator and −0.041 ([−0.22, 0.15]) under an order-aware GRU, ruling out moderate associations. We treat this null result as the paper’s central finding: improving the standard fidelity metric for rare finan- cial trajectories provides no measurable guarantee of improving the decision it is meant to support. Financial inclusion motivates, but is not directly validated by, this study.