How Uber, Lyft, and DoorDash Optimize Surge Pricing with Two‑Sided Market Experiments

Introduction

AB testing is the gold standard for causal inference, but two‑sided markets introduce network effects that bias measurements. Designing experiments and platforms to address these challenges is essential for companies like Uber, Lyft, and DoorDash.

Surge Pricing Experiment (Part One)

General Pricing Factors

Using Uber as an example, pricing factors are broken down and visualized on a hexagonal H3 grid. When demand spikes in a region, a multiplier (e.g., X.X) is applied and shown to riders and drivers.

Peak‑time Surcharge Details

Uber maps orders onto hexagonal cells; when order volume and driver‑to‑rider ratio rise, a surcharge is triggered and displayed as a multiplier. Drivers can see heat maps of surcharge zones.

Lyft Iterations

Lyft iterated four versions (V0‑V3) focusing on service availability, driver wait time, and network throughput, using queueing models (M/M/c) and spatial Poisson processes to link ETA, driver availability, and pricing.

Time‑Space Slice Experiment (Part Two)

Traditional A/B tests violate the Stable Unit Treatment Value Assumption (SUTVA) in two‑sided markets. DoorDash adopts a time‑space slice design where each slice (e.g., 30 minutes) randomly receives a strategy, and different regions are assigned independently.

Granularity trade‑offs: shorter slices reduce variance but increase bias from competition; longer slices do the opposite.

Bias arises from shared driver pools; variance depends on slice granularity.

Random Saturation Experiments

Two‑stage randomization assigns a saturation level to clusters (e.g., regions) and then randomly treats individuals within clusters, allowing measurement of network‑effect bias.

Uber Experiment System (XP)

XP supports A/B/N tests, causal inference, and multi‑armed bandit experiments across multiple Uber apps. It provides homogeneous group detection, automatic statistical test selection, and continuous experiments.

Features include pre‑existing bias checks, a statistics engine that chooses appropriate tests, and support for continuous experimentation using bandit algorithms and Bayesian optimization.

Summary

Pricing in two‑sided markets requires sophisticated experimental designs to mitigate network‑effect bias. DoorDash’s time‑space slicing, academic optimizations for carry‑over effects, and random‑saturation methods improve measurement accuracy, while Uber’s XP platform standardizes experiment execution and analysis.