CONCEPT Cited by 1 source
Probabilistic demand forecast¶
Definition¶
A probabilistic demand forecast is a forecast whose output is not a single point estimate but a distribution over possible future demand values — typically parameterised by quantiles, prediction intervals, or a full density function. Required upstream input for any optimisation under uncertainty formulation of inventory / replenishment / capacity problems, because the optimiser needs to reason about the shape of the risk, not just the mean.
Why it matters for replenishment¶
Point forecasts tell you "the expected demand next week is 47
units"; they don't tell you whether the realistic range is
[40, 55] (tight; small safety stock suffices) or [0, 200]
(wide; order more to avoid stockouts, accept some overstock
risk). A stockout vs overstock decision is fundamentally
asymmetric — a probabilistic forecast exposes that asymmetry
to the optimiser, which can then explicitly weigh
C_lost_sales against C_overstock via
Monte
Carlo simulation.
Producing probabilistic forecasts¶
Mechanisms observed on the wiki:
- Conformal inference wrappers around point-forecast models (e.g. Nixtla MLForecast around LightGBM point forecasts, producing calibrated prediction intervals).
- Native probabilistic models (quantile regression, TFT, DeepAR) — mentioned as Zalando alternatives that lost to LightGBM + MLForecast on prototyping speed + ecosystem.
- Parametric assumption + empirical residuals.
Canonical instance (Zalando ZEOS, 2025-06-29)¶
- Horizon: 12 weeks.
- Granularity:
(article_id, merchant_id, week)— see concepts/skus-as-time-series-unit. - Catalogue: 5 million SKUs.
- Refresh: weekly.
- Producer: systems/zeos-demand-forecaster using Nixtla MLForecast with conformal inference on top of LightGBM.
- Consumer: systems/zeos-replenishment-recommender runs Monte Carlo simulation over this distribution to score candidate replenishment decisions.
Seen in¶
- sources/2025-06-29-zalando-building-a-dynamic-inventory-optimisation-system-a-deep-dive — canonical disclosure as the upstream input to gradient-free black-box inventory optimisation.
Related¶
- concepts/monte-carlo-simulation-under-uncertainty — the consumer pattern that needs probabilistic forecasts as input.
- concepts/gradient-free-black-box-optimisation — the decision-making layer fed by Monte Carlo samples from this distribution.
- concepts/skus-as-time-series-unit — granularity choice.
- systems/mlforecast-nixtla · systems/lightgbm · systems/zeos-demand-forecaster