Estimating Demand for Online Delivery using Limited Historical Observations

TL;DR

This paper develops algorithms for imputing and estimating online delivery demand using limited and sparse historical data, enabling future demand projections despite data gaps and non-responses.

Contribution

It introduces novel data imputation and synthetic demand estimation methods that leverage existing sparse datasets to improve demand forecasting accuracy.

Findings

Imputation achieves MSE ≈ 0.65 with mean ≈ 1 and SD ≈ 3.5

Synthetic demand estimates align closely with actual data

Identifies key variables influencing delivery volume

Abstract

Driven in part by the COVID-19 pandemic, the pace of online purchases for at-home delivery has accelerated significantly. However, responding to this development has been challenging given the lack of public data. The existing data may be infrequent, and a significant portion of data may be missing because of survey participant non-responses. This data paucity renders conventional predictive models unreliable. We address this shortcoming by developing algorithms for data imputation and synthetic demand estimation for future years without the actual ground truth data. We use 2017 Puget Sound Regional Council (PSRC) and National Household Travel Survey (NHTS) data and impute from the NHTS for the Seattle-Tacoma-Bellevue MSA where delivery data is relatively more frequent. Our imputation has the mean-squared error $\mathsf{MSE} \approx 0.65$ to NHTS with mean $\approx 1$ and standard deviation $\approx 3.5$ and provides a similarity matching between the two data sources' samples. Given the unavailability of NHTS data for 2021, we use the temporal fidelity of PSRC data sources (2017 and 2021) to project the resolution onto the NHTS providing a synthetic estimate of NHTS deliveries. Beyond the improved reliability of the estimates, we report explanatory variables that were relevant in determining the volume of deliveries. This work furthers existing methods in demand estimation for goods deliveries by maximizing available sparse data to generate reasonable estimates that could facilitate policy decisions.