In illegal online transaction forecasting, models face significant challenges due to temporal non-stationarity, long-range dependencies, and the frequent misclassification of event data as anomalies, which collectively undermine prediction accuracy and robustness. To address these issues, we propose DS-ADV, a novel temporal forecasting model that integrates adaptive slicing and dual-branch adversarial enhancement. DS-ADV consists of two core components: an Adaptive Slice-Length Resizing Model (ASRM), which dynamically adjusts temporal granularity based on local variance to capture feature drift in non-stationary series, and a Robust Dual-branch Temporal Forecasting Model (RDFM), which combines a discrepancy-aware forecasting branch with a dynamic statistical feature forecasting branch. The ASRM performs dynamic slicing based on the mean and standard deviation within a sliding window to tackle the inherent non-stationarity of the data. The RDFM employs a hybrid TCN-Transformer architecture and seasonal-trend decomposition to effectively enhance robustness against anomalous data, employing adversarial domain alignment to minimize the distribution shift between event and base data. Extensive experiments on real-world transaction data and public benchmarks show that DS-ADV significantly outperforms baseline methods, reducing MSE and MAE by up to 8.7% and 6.9%, respectively, under various forecasting horizons.