Logistics ops · Production

DoorDash NextGen ETA system: multi-task deep learning with probabilistic forecasts

The problem

DoorDash's ETA prediction required a cumbersome array of specialized tree-based models, one per delivery type, which became unsustainable to maintain and reached a performance plateau where additional data, features, and enhancements yielded no further improvement. Separately trained models also produced ETA discrepancies across consumer touchpoints, harming the customer experience.

First attempt

Tree-based models could not generalize to unseen or rare delivery scenarios and forced accuracy tradeoffs between earliness and lateness rather than improving overall on-time prediction.

Workflow diagram · grounded in source
1
Customer ETA request
trigger
“DoorDash's ETAs cater to various customer interaction stages and delivery types. Initially, customers can use ETAs on the home page to help them decide between restaurants and other food merchants.”
2
Multi-task model inference
ai_action
“Our MT architecture consists of a shared heavyweight foundation layer, followed by a specialized lightweight task head for each prediction use case”
3
Probabilistic base layer output
ai_action
“The base layer model outputs a predicted distribution to estimate expected ETA time”
4
Decision layer optimization
ai_action
“The decision layer leverages the base model's predictions to solve various multi-objective optimization problems for different businesses”
5
ETA delivered across touchpoints
output
“providing consistent ETA predictions through different stages of the consumer's journey, including on the home page, store page, and checkout page”
Reported outcome

DoorDash developed three versions of NextGen ETA models that improve consumer outcomes on accuracy and consistency, with multi-task learning notably improving predictions for lower-volume markets such as Australia.

Reported metrics
Annual order volume at scale2 billion
NextGen ETA model consumer outcomesall improve consumer outcomes
Australia ETA prediction improvementgreatly improved ETA predictions for deliveries in Australia
Probabilistic prediction progresssignificant strides in the probabilistic prediction and distribution evaluation
Reported stack
Deep Learningneural networksmulti-task
Source
https://careersatdoordash.com/blog/improving-etas-with-multi-task-models-deep-learning-and-probabilistic-forecasts/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

DoorDash developed three versions of NextGen ETA models that improve consumer outcomes on accuracy and consistency, with multi-task learning notably improving predictions for lower-volume markets such as Australia.

What tools did this team use?

Deep Learning, neural networks, multi-task.

What results were reported?

Annual order volume at scale: 2 billion; NextGen ETA model consumer outcomes: all improve consumer outcomes; Australia ETA prediction improvement: greatly improved ETA predictions for deliveries in Australia; Probabilistic prediction progress: significant strides in the probabilistic prediction and distribution evaluation (source-reported, not independently verified).

What failed first in this deployment?

Tree-based models could not generalize to unseen or rare delivery scenarios and forced accuracy tradeoffs between earliness and lateness rather than improving overall on-time prediction.

How is this logistics ops AI workflow structured?

Customer ETA request → Multi-task model inference → Probabilistic base layer output → Decision layer optimization → ETA delivered across touchpoints.

Related logistics ops cases