Otto

Objective

Predict what customers will order before they do, to order the merchandise from third-party brands ahead of time, cut surplus stock and returns, and deliver faster from stock already in the right place.

The deployment

Otto is one of Germany's largest e-commerce retailers, with more than 18 million items in its catalog. A large share of its offer comes from third-party brands that Otto does not stock itself, which stretched lead times and multiplied separately shipped parcels. Around 2016-2017, Otto put into production a forecasting engine built on the technology of Blue Yonder, a German vendor in which it holds a stake: a deep-learning algorithm originally designed for particle-physics experiments at CERN, which analyses around 3 billion past transactions and 200 variables (past sales, searches on Otto's site, weather) to predict what customers will buy a week before they order. The system proved reliable enough that Otto lets it buy around 200,000 items a month from third-party brands on its own, with no human intervention. The engine has changed since: Otto has run an in-house forecasting solution since 2019, and the Google Cloud case study documents the current engine, a TiDE model trained on Vertex AI and deployed on GKE. Otto today reports 35% of its product ranges being automatically reordered, on the basis of around 30 billion individual forecasts a month.

Results Proof B

The Economist (April 2017) documents the case by name, names Blue Yonder, and carries the figures word for word: 90% accuracy on 30-day sales, surplus stock down by a fifth, more than 2 million returns avoided a year, 200,000 items bought on its own each month. Two quantified vendor case studies corroborate (Blue Yonder, Google Cloud), and Otto's official site confirms the current automated scope. Important caveat: the Blue Yonder case study defines its own 90% differently (the share of items ordered from partners that sell out within 30 days, per Mathias Stuben) and contains an unrelated 90% (the online shop's share of annual sales); only The Economist's 90%, which does measure forecast accuracy, is used here. No A level, as there are no financial results.

1. 1
   Signal collection Otto systems / data team

   Sales history, customer behaviour, seasonal trends, public holidays, planned marketing activities, and weather are aggregated. The 2017 deployment drew on around 3 billion past transactions and 200 variables.

2. 2
   Demand forecasting AI model

   The model predicts what will sell in the short term. The Economist reports 90% accuracy on 30-day sales for the 2017 Blue Yonder system; the Google Cloud case study credits the TiDE model with up to 30% more accuracy.

3. 3
   Autonomous ordering AI model / purchasing system

   Across the automated scope, the system orders merchandise from third-party brands on its own, with no human intervention: around 200,000 items a month in 2017.

4. 4
   Partial automated scope Otto purchasing

   Automation does not cover the whole catalog: Otto reports 35% of product ranges already being reordered automatically. Public sources do not describe the rule that decides the scope.