Rare-event prediction in multivariate time series (MVTS) is difficult because target events are scarce, their evidence is temporally structured, and the majority class is often large and heterogeneous. Solar flare forecasting provides a demanding case. Operationally important M- and X-class flares are rare, active regions evolve over time, and overlapping observation windows can create ambiguous negative examples. Using the Space Weather Analytics for Solar Flares (SWAN-SF) benchmark as the primary empirical setting, this dissertation studies how training evidence can be curated, expanded, aligned with prediction, and made inspectable. Four connected studies examine these dimensions. First, Isolation Forest identifies candidate anomalous N-class samples whose associations with flare activity are consistent with heterogeneity within the negative class. Changing their treatment substantially affects a controlled N–X task, whereas effects in the broader NBC–MX task depend more on the temporal partition. Second, Mean Gaussian Noise (MGN) generates minority samples around a classlevel temporal prototype. Relative to controlled alternatives, MGN produces a favorable balance between the true skill statistic (TSS) and the updated Heidke skill score (HSS2), with higher HSS2 and normalized distance-to-perfect summary but lower TSS. Third, UniTSGAN couples minority generation with discriminative learning. Its integrated evaluation yields a favorable observed aggregate TSS–HSS2 balance on SWAN-SF, while detached evaluations and additional datasets show that augmentation utility varies with the downstream classifier, temporal partition, and dataset. Fourth, boundary-focused counterfactual augmentation applies localized temporal edits to teacher-identified hard negatives and assesses them with a separately trained downstream classifier. It yields a higher aggregate mean HSS2 and normalized distance-to-perfect summary than the listed comparison settings, together with a lower mean TSS than most alternatives. Saved factual–counterfactual pairs enable direct inspection of edit locations and temporal behavior, while motivating further work on interval diversity and within-interval coherence. Across the evaluated settings, the results highlight four aspects of training-evidence construction: class definition, minority representation, alignment between generation and prediction, and traceability to observed trajectories. The findings are grounded primarily in SWAN-SF, and broader transfer to other rare-event MVTS domains remains an important direction for future investigation. Establishing whether prediction-useful synthetic trajectories are also physically plausible will require physics-informed constraints and domain validation.