Temporal Fusion Transformer (TFT)

TFT employs Variable Selection Networks at each time step to automatically identify and suppress irrelevant input features. Unlike standard models that treat all inputs equally, this gating mechanism performs instance-wise variable selection, allowing the model to focus on the most salient drivers of lead time—such as carrier performance or port congestion—while ignoring noise. This is critical for supply chain data where hundreds of potential covariates exist but only a few are predictive at any given moment.

The Gated Residual Network is the core building block of TFT, replacing standard feed-forward layers. It applies Gated Linear Units (GLUs) and residual connections to control the flow of information through the network. Key benefits include:

• Non-linear processing: Captures complex relationships in lead time dynamics
• Adaptive depth: The gating mechanism can skip unnecessary layers, preventing overfitting on small datasets
• Dropout-based regularization: Ensures robust performance on noisy logistics data

TFT uniquely handles static metadata—features that do not vary over time—through dedicated encoder networks. In supply chain contexts, this includes:

• Supplier category and geographic location
• Product characteristics (hazardous, temperature-sensitive)
• Historical reliability scores These static encodings are integrated into the temporal processing pipeline, conditioning the model's understanding of dynamic patterns. For example, a port strike impacts a single-sourced supplier differently than a multi-sourced one, and TFT captures this interaction explicitly.

TFT incorporates a modified multi-head attention mechanism to learn long-term temporal dependencies without the vanishing gradient problems that plague recurrent architectures. The attention layer:

• Identifies which historical time steps are most relevant to the current forecast horizon
• Captures seasonal patterns and cyclical lead time behaviors across extended lookback windows
• Uses a self-attention design modified to respect the causal ordering of time, ensuring predictions only depend on past observations This allows the model to detect, for instance, that a supplier delay 6 months ago is predictive of a current bottleneck due to recurring quality issues.

TFT generates multi-horizon quantile forecasts (e.g., 10th, 50th, 90th percentiles) rather than single point estimates. This provides planners with a full predictive distribution, enabling risk-based decision-making. The architecture achieves this through:

• Quantile loss functions during training
• Simultaneous prediction across all horizons, avoiding error accumulation from recursive forecasting
• Explicit modeling of asymmetric uncertainty, crucial for lead times where late deliveries are more costly than early ones Planners can use the 90th percentile as a conservative buffer for safety stock calculations while the 50th percentile informs expected arrival dates.

TFT explicitly distinguishes between known future inputs and unknown future inputs through separate processing paths. Known inputs—such as planned holidays, scheduled port closures, or contracted carrier capacity—are fed directly into the decoder. Unknown inputs, like actual demand spikes or weather disruptions, are only available historically. This architectural separation prevents the model from inadvertently leaking future information during training and ensures realistic forecasting that respects the information available at prediction time.