Supply chain · Production

DoorDash uses causal inference to improve supply and demand forecast accuracy

The problem

DoorDash's forecasting models could not account for hard-to-measure macroeconomic and calendar factors such as tax refunds and daylight saving because these factors lack a direct feature measure and standard feature engineering fails to capture their heterogeneous, time-varying demand impact.

First attempt

Standard feature engineering and one-hot encoding of concurrent events introduced multicollinearity that prevented disentangling individual factor impacts, and tax refund data alone could not capture the time-varying heterogeneous response across consumer segments, causing models to treat temporary spikes as permanent trends.

Workflow diagram · grounded in source
1
Unmeasurable factor identified
trigger
“some temporary sources of fluctuations (e.g., tax refunds, child tax credits) may not have a direct measure/feature, which makes them hard to detect and account for in our forecast”
2
Market sub-segments selected
ai_action
“Identify sub-segments of the market that are differentially impacted by the factor causing fluctuations in the demand”
3
Difference-in-differences model run
ai_action
“we decided to use the difference-in-differences method, in which we set the treatment group as the affected zip codes and the control group as the non-affected zip codes”
4
Impact removed from demand series
integration
“Remove the calculated impact from the target series so that trend is not affected by the factor”
5
Time-series forecast generated
ai_action
“generate forecasts using the time-series models for submarkets in the United States”
6
Future factor impact applied
output
“we apply post-processing to incorporate the future tax refund impact based on the tax refund phase out in the earlier years”
Reported outcome

Applying causal inference methods to preprocess demand series before time-series modeling produced a sizable improvement in out-of-sample forecast accuracy and a sizable reduction in acquisition costs.

Reported metrics
Forecast out-of-sample performancesizable improvement in the out-of-sample performance of the forecast
Dx acquisition costssizable reduction in the Dx acquisition costs
Reported stack
difference-in-differencestime-series models
Source
https://careersatdoordash.com/blog/leveraging-causal-inference-to-generate-accurate-forecasts/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Applying causal inference methods to preprocess demand series before time-series modeling produced a sizable improvement in out-of-sample forecast accuracy and a sizable reduction in acquisition costs.

What tools did this team use?

difference-in-differences, time-series models.

What results were reported?

Forecast out-of-sample performance: sizable improvement in the out-of-sample performance of the forecast; Dx acquisition costs: sizable reduction in the Dx acquisition costs (source-reported, not independently verified).

What failed first in this deployment?

Standard feature engineering and one-hot encoding of concurrent events introduced multicollinearity that prevented disentangling individual factor impacts, and tax refund data alone could not capture the time-varying…

How is this supply chain AI workflow structured?

Unmeasurable factor identified → Market sub-segments selected → Difference-in-differences model run → Impact removed from demand series → Time-series forecast generated → Future factor impact applied.

Related supply chain cases