The encyclopedia · Software & IT · Technical decision · 2021–2024
Amazon forecasted freight market rates 5x better with signature transforms
Amazon's signature-transform model translates rate time series into regression features, beating industry models 5-fold and saving over $50M a year since 2021.
Amazon
The solution
Freight transportation marketplace rates swing with cycles, shocks and heterogeneous lanes, and forecasting them poorly inflates a shipper's network costs and procurement decisions. Amazon's middle-mile trucking team needed predictions robust enough to survive extreme disruption, not just calm quarters.
Their solution rests on the signature transform: a mathematical operator whose universal nonlinearity linearizes the feature space. Instead of training a heavy sequence model, they transform the rate series into signature features and fit a linear regression, while the signature kernel compares time-series similarity so regime changes surface explicitly.
The algorithm was deployed in Amazon middle-mile trucking operations with far superior forecast accuracy and better interpretability than commercially available industry models — more than fivefold better — and estimated annualized savings above $50 million per year since 2021.
Why it worked
- Linearizing the feature space turns a nonparametric problem into a stable, interpretable regression instead of a black-box forecast.
- The signature kernel explicitly tracks seasonality and regime switching, so the model reacts to market regime changes that confound fixed-window models.
- Interpretability mattered as much as accuracy: operators could see why a rate was predicted the way it was and trust it during the pandemic and conflict shocks.
What can be applied
Before adding a bigger model to hard forecasting problems, try translating the input: the signature transform turns a messy time series into features a plain linear regression can read.
Aftermath
The technique was a finalist for the 2024 Daniel H. Wagner Prize for Excellence in the Practice of Advanced Analytics and Operations Research, published in the INFORMS Journal on Applied Analytics. Amazon has kept the model in production for its middle-mile trucking procurement, with documented annualized savings exceeding $50 million since 2021.
Sources
- Transportation Marketplace Rate Forecast Using Signature Transform
- Transportation Marketplace Rate Forecast Using Signature Transform
spotted an error? The archive wants to know.