DeepAR

DeepAR trains a single global model on hundreds or thousands of related time series simultaneously, rather than fitting separate models per SKU or location. This approach enables the model to learn shared seasonal patterns, trend behaviors, and demand correlations across the entire product hierarchy. By pooling information from high-volume items, the model improves forecast accuracy for low-volume or intermittent-demand products that lack sufficient individual history. The architecture uses an encoder-decoder LSTM that conditions predictions on both historical values and known future covariates such as planned promotions, pricing changes, or calendar events.

Instead of predicting a single demand number, DeepAR outputs the parameters of a chosen probability distribution at each forecast time step. Common distributions include:

• Gaussian for real-valued data with symmetric uncertainty
• Negative Binomial for overdispersed count data common in retail demand
• Beta for constrained values like fill rates or service levels
• Student's t for heavy-tailed distributions with outlier demand spikes

The model learns to predict distribution parameters directly, enabling the generation of quantile forecasts at any service level—such as the 10th, 50th, and 90th percentiles—without retraining. This allows inventory planners to set safety stock based on explicit prediction intervals rather than heuristic buffers.

DeepAR addresses the cold-start problem by leveraging embeddings learned from similar items during global training. When a new SKU enters the catalog with zero or minimal sales history, the model uses item-level categorical features—such as product category, brand, price tier, or warehouse location—to generate an initial forecast based on the behavior of comparable products. As the first sales observations arrive, the model dynamically updates its predictions through the autoregressive LSTM structure. This capability is critical for fashion retail, electronics launches, and pharmaceutical rollouts where new product introductions are frequent and historical data is absent.

DeepAR natively ingests two categories of external features that shape demand predictions:

• Static covariates: Time-invariant attributes like product category, store size, or geographic region that define the identity of each time series
• Dynamic covariates: Time-varying inputs known in advance across the forecast horizon, such as planned promotions, holiday calendars, price changes, or marketing spend

The model learns complex interactions between these covariates and historical demand patterns. For example, it can capture how a 20% price discount during a holiday weekend affects demand differently for premium versus budget product categories. This context-aware conditioning makes DeepAR particularly effective for promotional demand sensing and scenario planning.

DeepAR generates probabilistic forecasts through ancestral sampling: at each time step, the model draws a value from the predicted distribution and feeds it back as input for the next step. This autoregressive process is repeated hundreds or thousands of times to build an empirical distribution of possible future trajectories. The resulting sample paths capture the compounding uncertainty that grows with the forecast horizon—a phenomenon that point-forecast models ignore. Supply chain planners can use these trajectories to run Monte Carlo simulations for inventory optimization, calculating metrics like expected stockout probability or fill rate under different reorder policies.

DeepAR is trained by maximizing the log-likelihood of observed values under the predicted distribution, which is a strictly proper scoring rule that encourages well-calibrated probabilistic outputs. This contrasts with models trained on point-error metrics like MAE or RMSE, which ignore uncertainty quantification. For practitioners who prefer direct quantile estimation, the architecture can alternatively be trained using pinball loss, an asymmetric loss function that penalizes over-prediction and under-prediction differently depending on the target quantile. The choice between likelihood and pinball loss depends on whether the downstream application requires full distributional information or specific quantile thresholds for safety stock calculation.