Serve a forecast path with an explicit objective

A direct method predicts a particular horizon from the origin’s information; separate direct models can specialize but duplicate work. A joint multi-output method predicts several horizons together, potentially sharing an encoder. A recursive method repeatedly applies a one-step transition, consuming its own generated outcomes after the first step. Training with observed previous outcomes and serving with generated ones creates a distribution shift that must be evaluated through complete forecast rollouts rather than teacher-forced one-step losses alone.

DeepAR provides an autoregressive probabilistic construction, while Temporal Fusion Transformers illustrate direct multi-horizon quantile forecasting with distinct input roles. These are examples of design choices, not guarantees that one architecture wins on every workload. For any model, separate static metadata, observations available only through the issue origin, and genuinely future-known covariates. Weather observations from tomorrow are not interchangeable with today’s weather forecast for tomorrow. Preserve the vintage of the latter when it is used as a predictor.

Squared-error minimization targets a conditional mean under the usual finite-second-moment population interpretation. Absolute error targets a conditional median, and pinball loss targets a specified conditional quantile. If under-reserving capacity is more costly than over-reserving, a higher quantile may be useful, but its probability level should follow a stated decision cost model. Do not label a median as expected demand or add medians across services as though they were additive expectations.

Separate per-horizon weighting from per-example weighting. Long horizons may have fewer mature labels, and a global normalization over active entries can then give short horizons more effective influence. If the objective is an equal average of horizon-specific risks, compute those risks with their own denominators before combining them. Quantile models also need checks for crossing outputs and calibration by horizon. Reordering quantiles may repair ordering but changes the deployed prediction rule and must itself be evaluated.

Return issue timestamp, ordered target intervals, horizon convention, units, model and feature versions, and output type. If a horizon is unsupported, reject it or expose a documented interpolation policy; silently mapping twelve hours to the nearest trained horizon changes the question. For probabilistic samples, preserve path identity across horizons so a downstream planner can evaluate aggregate demand and temporal peaks under the model’s joint dependence assumptions.

Do not treat independent marginal quantiles as a joint scenario. Summing a ninety-fifth percentile at each hour does not generally produce the ninety-fifth percentile of total demand. A planner needs either a valid joint construction or an explicit conservative approximation. Exercise serving with partial horizon masks, missing future-known features, and artifact changes while a request is in flight. The response should be internally version-consistent even when a new model is promoted halfway through computation.