Impact-driven event embeddings for context-aware forecasting and anomaly detection in financial time series

Financial transaction time series are strongly shaped by recurring external events such as holidays, promotional campaigns, and settlement cycles. In practice, however, anomaly detection systems often evaluate these series without explicitly modeling event context, leading to excessive false positives during predictable event-driven fluctuations. We propose a context-aware framework based on impact-driven event embeddings, where events are represented by the residual behavioral patterns they induce in transaction series rather than by labels alone. To study this setting under controlled conditions, we construct a synthetic benchmark calibrated to empirical signatures from production financial transaction data, spanning heterogeneous financial accounts, realistic event calendars, and injected anomalies. The framework estimates non-event baselines from observed series, learns compact event embeddings from event-centered residual responses, and uses them as future-known covariates in a probabilistic forecasting model. Anomalies are then defined as deviations from the resulting event-conditioned forecast distribution. In downstream experiments, the proposed approach outperforms both no-event and binary-event baselines for event-sensitive account profiles. Lag-bucketed embeddings are most effective for delayed-response cash accounts, while same-day embeddings are strongest for receivable and revenue accounts, where event effects are concentrated nearer the event itself. These results support the view that anomaly detection in financial time series is better framed as a context-aware forecasting problem in which learned event representations improve calibration and reduce false positives during known events.