Logistics ops · Production

DoorDash builds ML forecasting and optimization system to balance Dasher supply and delivery demand

The problem

DoorDash needed an automated system to allocate Dasher incentives ahead of anticipated supply-demand imbalances across thousands of regional markets and time units, but lacked a reliable, maintainable way to forecast and optimize these allocations at the appropriate granularity.

First attempt

Overly complex ML pipelines with long data dependency chains were identified as a reliability risk: initially outperforming naive forecasting but degrading to 'Terrible' performance after three or more months, with high oncall burden. Blind correlation learning also risked learning spurious relationships, such as mistakenly concluding that high incentives lead to fewer Dashers on the road.

Workflow diagram · grounded in source
1
Hours-gap metric computation
Trigger
DoorDash measures expected Dasher hours versus organically available hours per regional-time unit to quantify supply-demand imbalance.
▾ source quote
“For our primary supply and demand measurement metric, we looked at the number of hours required to make deliveries while keeping delivery durations low and Dasher busyness high”
2
ML supply and demand forecasting
Ai action
Gradient boosting via LightGBM generates thousands of regional supply and demand forecasts, reformulated as a regression problem.
▾ source quote
“we primarily reformulated the forecasting problem into a regression problem and used gradient boosting through the Microsoft-developed open source LightGBM framework”
3
Uncertainty quantification via resampling
Validation
A resampling process generates expected estimates of hours gap to measure the likelihood and impact of undersupply, preventing over-allocation in high-variance regions.
▾ source quote
“we generate expected estimates of hours gap from forecasts using a resampling process. By performing resampling, we essentially measure the impact of undersupply in the context of the likelihood of that happening”
4
MIP incentive optimization
Ai action
A mixed-integer program minimizes undersupply by allocating incentives to regional-time units under budget and frequency constraints.
▾ source quote
“The optimizer has a custom objective function of minimizing undersupply with several constraints”
5
Dasher incentive deployment
Output
Dashers receive a guarantee of fixed earnings for deliveries accepted in a specific region-time unit.
▾ source quote
“we provide Dashers a guarantee that they will earn a fixed amount of money on any delivery they accept in a specific region-time unit”
Reported outcome

The mobilization system reduced delivery times, cancellations, and extreme lateness for consumers; drove down merchant order cancellations; enabled more reliable budget adherence with less spending variability; and increased the team's experimentation velocity on the incentive system.

Reported metrics
Delivery timesreductions in delivery times
Order cancellations (consumer)reductions in cancelations
Extreme latenessreductions in extreme lateness
Order cancellations (merchant)order cancellations down
Show all 7 reported metrics
delivery timesreductions in delivery times
order cancellations (consumer)reductions in cancelations
extreme latenessreductions in extreme lateness
order cancellations (merchant)order cancellations down
budget adherence reliabilitymore reliably hit budget expectations
spending variabilityless variability in spending decisions
experimentation velocityrapidly increased
Reported stack
LightGBM
◆ Does this fit your context?

Compare to your context

Tell us your scale, team, and constraints. We'll show what changes at your size, what fails at your scale, and whether this case is a fit, needs adaptation, or won't scale to you. Free demo, no signup.

Compare to your context →
~30 seconds · free
Source
https://careersatdoordash.com/blog/managing-supply-and-demand-balance-through-machine-learning/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The mobilization system reduced delivery times, cancellations, and extreme lateness for consumers; drove down merchant order cancellations; enabled more reliable budget adherence with less spending variability; and in…

What tools did this team use?

LightGBM.

What results were reported?

Delivery times: reductions in delivery times; Order cancellations (consumer): reductions in cancelations; Extreme lateness: reductions in extreme lateness; Order cancellations (merchant): order cancellations down (source-reported, not independently verified).

What failed first in this deployment?

Overly complex ML pipelines with long data dependency chains were identified as a reliability risk: initially outperforming naive forecasting but degrading to 'Terrible' performance after three or more months, with hi…

How is this logistics ops AI workflow structured?

Hours-gap metric computation → ML supply and demand forecasting → Uncertainty quantification via resampling → MIP incentive optimization → Dasher incentive deployment.

Related logistics ops cases

WHAT TO DO WITH THIS

Now compare it to your context

This case is one data point. Whether its pattern fits you depends on your volumes, your stack, and your exception load — that comparison is the step no case study can do for you.

Browse the Logistics ops cases →