Time-Decay Function

The most common parametric form, expressed as w(t) = e^(-λt), where λ (lambda) controls the decay rate. A larger λ causes weights to diminish more aggressively, creating a shorter 'memory' horizon. This formulation ensures that the influence of a transaction halves over a consistent time period, known as the half-life. In fraud systems, the half-life is a tunable hyperparameter calibrated to match the velocity of legitimate user behavior changes versus adversarial adaptation speeds.

A polynomial alternative defined as w(t) = 1 / (1 + k*t). Unlike exponential decay, this function exhibits a heavier tail, meaning very old observations retain non-negligible influence for longer. This is useful when modeling long-term baseline behaviors that should not be entirely forgotten, such as a user's stable annual spending patterns. The parameter k adjusts the steepness of the initial drop-off without fully extinguishing historical signal.

Practical implementations often combine a decay function with a hard temporal window to bound computational cost. Weights for events older than a cutoff threshold are explicitly set to zero. This creates a finite receptive field for sequence models. Common strategies include:

• Sliding windows: Fixed-length lookback (e.g., 90 days)
• Landmark windows: All data since a significant event (e.g., password reset)
• Fading windows: Weights decay to zero at the boundary, ensuring smooth transitions

Time-decay functions transform raw event streams into weighted aggregate features for downstream classifiers. Instead of simple counts, features become decay-weighted sums. For example, a decay-weighted transaction count gives full weight to today's transaction, half weight to yesterday's, and near-zero weight to last month's. This captures recency-sensitive velocity—a sudden burst of activity today will dominate the feature value, triggering anomaly scores even if the raw monthly total appears normal.

Modern architectures like the Temporal Fusion Transformer and Mamba incorporate learned or structural time-decay directly into their attention or state-space mechanisms. Rather than a fixed exponential function, these models learn input-dependent decay rates. A transaction from a known compromised merchant might be assigned a slower decay, retaining its relevance longer, while a routine login fades quickly. This adaptive temporal weighting allows the model to dynamically modulate its memory horizon based on event significance.

The decay rate λ is a direct lever for managing temporal bias-variance tradeoff. A fast decay (high λ) creates a model highly responsive to new fraud patterns but prone to noise and false positives. A slow decay (low λ) provides stability but risks concept drift blindness, where the model fails to adapt to shifting criminal tactics. Optimal calibration often involves monitoring the half-life of predictive power—the period over which historical features remain statistically significant for predicting current fraud labels.

A foundational forecasting method that applies decreasing weights to past observations, directly implementing a time-decay principle. The ETS framework models error, trend, and seasonality components to establish a dynamic baseline for expected transaction behavior. The smoothing parameter α controls the decay rate, with values closer to 1 giving more weight to recent observations and values closer to 0 distributing weight more evenly across the series.

A derived feature measuring the rate of transactions over a specific time window, such as login attempts or payment transfers per minute. Time-decay functions are critical here because a simple count over a fixed window creates artificial discontinuities at window boundaries. A decay-weighted velocity applies exponential or inverse weighting to events within the window, producing a smoother, more responsive signal that is a critical real-time indicator of automated bot attacks or account takeover.

A self-exciting temporal point process where the occurrence of an event increases the probability of future events in the near term. The excitation function typically includes an exponential decay kernel of the form αe^(-βt), where α controls the magnitude of self-excitation and β governs the decay rate. This directly models bursty behavior in financial transactions or high-frequency trading, where a rapid sequence of small transactions may precede a large fraudulent one.

The phenomenon in streaming data where the underlying statistical relationship between input features and the target variable changes over time. Time-decay functions serve as a mitigation strategy by continuously down-weighting obsolete patterns. In adaptive learning contexts, a decay factor on historical loss contributions ensures the model prioritizes recent data distributions, preventing it from being anchored to outdated fraud patterns that adversaries have already abandoned.

A scalar value quantifying the abnormality of an entire sequence of events, derived from the reconstruction error of an autoencoder or the negative log-likelihood of a sequence model. Time-decay weighting is often applied when aggregating per-event anomaly scores into a session-level score, ensuring that suspicious activity at the end of a session contributes more to the final risk assessment than ambiguous signals at the beginning. This temporal prioritization reduces false positives from stale signals.

A sequential analysis technique that accumulates deviations from a target mean over time, triggering an alert when the cumulative sum exceeds a threshold. While standard CUSUM uses uniform weighting, decay-enhanced CUSUM variants apply a forgetting factor to the accumulated sum, making the statistic more sensitive to recent shifts and preventing saturation from historical noise. This makes it highly sensitive to small, sustained shifts in transaction behavior that emerge gradually.