AI Forecasting & Supply Twins

Learned forecasting models and planning software set stock and routes at big retailers; Amazon claims 10-20% better forecasts, but supply-chain twins are early.

Open in the interactive tree →

Forecasting models learn from past sales plus weather, holidays and region, and optimisation software turns the forecast into orders, stock placement and delivery routes. A supply-chain digital twin is a live model of a whole network, used to replay disruptions before they happen. The first part is in use at large firms such as Amazon and UPS; the second needs data from many firms that rarely share it.

As of October 2026

Amazon announced a demand-forecasting foundation model on 11 June 2025, now used in the US, Canada, Mexico and Brazil, and claims 10% better long-term national forecasts for deal events and 20% better regional forecasts for millions of popular items (company claims; no independent audit found). UPS said in 2016 that its routing software ORION should cut 100 million miles and 10 million gallons of fuel a year (company expectation). In the M5 competition on 42,840 Walmart sales series, the organisers saw promise in machine learning and cross-learning, while simple local statistical methods can still compete on fine-grained data. A 2024 barrier study says digital-twin adoption in agro-food chains is limited, citing weak infrastructure, immature technology and high cost.

Open steps

Test forecasts on shocks, not calm years High AI leverageReplay 2020-2022 with models that saw only earlier data and score them against simple baselines; vendor gains such as Amazon's 10-20% have no independent audit.
Plan routes and sailing speeds High AI leverageOptimise routes and ship arrival times: UPS expects 100 million fewer miles a year from ORION; an IMO-backed model gives 14% less fuel per containership voyage.
Build a shared twin of one chain Medium AI leverageJoin data from several firms into one model that can replay a disruption; a 2024 barrier study names weak infrastructure, immature technology and high cost.

Where AI could help

High AI leverage. Demand forecasting and routing are pattern-and-optimisation problems where learned models already do well in tests; shared data and rare shocks limit the gain.

Forecast demand per product and region from sales history, weather and holidays
Place stock closer to likely buyers to shorten hauls and cut empty miles
Replay a disruption on a digital copy of the network before it happens
Propose alternative suppliers and routes when a link fails

Shown so far

Amazon says its demand-forecasting foundation model improved long-term national forecasts for deal events by 10% and regional forecasts for millions of popular items by 20% (company claim, June 2025). source
Chronos, a language-model-style forecaster pretrained on many time series, had comparable and occasionally superior zero-shot results to task-trained methods on new datasets in a 42-dataset benchmark (2024 paper). source
The M5 organisers (42,840 Walmart sales series) found cross-learning and machine learning promising, yet simple local statistical methods can stay competitive on high-granularity data (International Journal of Forecasting, 2021). source
UPS expects its ORION routing software to cut 100 million miles driven and 10 million gallons of fuel each year (company expectation, 2016). source

AI Forecasting & Supply Twins

As of October 2026

Open steps

Where AI could help

Shown so far

Prerequisites

Sources

More in Transport & Mobility · Present