EURO 2024 Copenhagen
Abstract Submission

EURO-Online login

2555. Foundation Models for Time Series Forecasting

Invited abstract in session WC-57: Forecasting for the middle mile, stream Optimization at Amazon.

Wednesday, 12:30-14:00
Room: S06 (building: 101)

Authors (first author is the speaker)

1. Chris George
ATS, Amazon
2. Claudio Coppola
Amazon
3. Jason Omedes
Amazon
4. Juba Nait Saada
Amazon
5. Sergio Alvarez Balanya
Amazon

Abstract

Maintaining a complex transportation network at global scale involves a large number of planning processes, each requiring visibility on how expected demand from customer orders will flow through local regions in the transportation network. For example, planning labour to process warehouses requires visibility on packages arriving at each warehouse. While traditional forecasting approaches rely on statistical or machine learning models independently optimised to forecast for a single granularity and horizon, they can require significant historical data for the granularity in question and cannot adapt well to new time series, inhibiting the expansion of the network to new countries or demand streams. In this work we explore a new approach to time series forecasting, where a single deep neural network trained on large collections of time series datasets is used to provide robust forecasts across many granularities and horizons. Inspired by Large Language Models and the ability of Transformer architectures to model long range dependencies in textual data; we hypothesise that adapting the Transformer architecture to numerical data and training them on large time series datasets can allow a single model to reliably forecast across multiple granularities and horizons and even outperform the traditional multiple-model approach, particularly in cases where little historical data is available.

Keywords

Status: accepted


Back to the list of papers