Designing Experiments for Peak Surge Pricing in Two‑Sided Markets: Lessons from Uber, Lyft, DoorDash and Didi

AB testing is the gold standard for causal inference, enabling unbiased effect estimation for better business decisions. In two‑sided markets, network effects create unique challenges that require specialized experimental designs and platforms.

The article surveys Uber, Lyft, DoorDash and Didi, focusing on their peak‑surcharge experiments and the underlying scientific systems.

1. Peak‑Surcharge Experiments

Uber maps orders onto H3 hexagonal grids; when order volume or driver‑to‑rider ratio spikes in a grid, a surcharge is triggered and displayed to users. Drivers see heat‑maps of surcharge zones. Lyft iterated its surcharge policy through four versions (V0‑V3), progressively shifting focus from guaranteeing service availability to optimizing network throughput and waiting time.

Key modeling approaches include:

Using M/M/c queueing theory to relate driver utilization and availability.

Modeling conversion rate versus price with exponential functions to predict optimal price adjustments.

Incorporating ETA and driver availability via spatial Poisson processes.

2. Time‑Space Slice Experiments

Traditional A/B testing fails in two‑sided markets due to shared supply pools. DoorDash adopts a time‑space slice design: a 30‑minute time slice is combined with spatial regions, and treatment assignment rotates between experiment and control groups across slices. Granularity of slices balances bias (smaller slices increase competition bias) against variance (larger slices increase variance).

Bias arises from indirect effects of experiment units on control units; variance depends on slice granularity. Optimizing slice granularity reduces both.

3. Random Saturation Experiments

Also known as two‑stage randomization, this design assigns a saturation level (fraction of a cluster receiving treatment) to each independent cluster (e.g., region, time slice). It mitigates network effects by controlling the proportion of treated units within a cluster.

Steps include defining independent clusters, selecting feasible saturation levels, and randomly assigning treatment within clusters.

4. Uber Experiment Platform (XP)

XP supports A/B/N tests, causal inference, and multi‑armed bandit experiments across multiple Uber apps (Rider, Driver, Eats, Freight). Features include:

Pre‑existing bias detection to ensure group homogeneity.

Automatic statistical engine selection for appropriate hypothesis testing.

Continuous experiments using bandit algorithms and Bayesian optimization, allowing simultaneous exploration and exploitation.

Bandit methods reduce opportunity cost but require engineering effort and may not suit long‑term effect evaluation.

5. Summary

Pricing in two‑sided markets is a complex systems problem; Lyft’s surcharge evolved through four iterations, while DoorDash mitigates network‑effect bias via time‑space slice experiments and further reduces carry‑over effects through optimization. Uber’s XP platform standardizes experimental hygiene, supporting efficient scientific iteration across diverse scenarios.