Cold-Start Markets

By the end of this article, you should be able to: say why a brand-new market breaks synthetic control and difference-in-differences, separate the questions a launch review mixes together, design a randomised rollout that avoids the problem, and judge a new market against past launches without overreacting to its first weeks.

Why the usual methods fail

A delivery marketplace opens its 41st city, Boise. Four weeks in, weekly gross order value (GOV) is $180k. Leadership wants to know whether the launch is working.

Every standard method fails in the same place:

• Synthetic control builds a blend of other cities that matches the treated city before treatment. Boise has no before.
• Difference-in-differences subtracts the treated unit's own trend before the change. There's no trend to subtract.
• Causal impact forecasting forecasts the unit from its own history. There's no history.

These aren't three problems. They're one. Each method learns what the treated unit would have done from the unit's own past, and a new market has no past.

Three questions that sound like one

"Is Boise working?" mixes three questions, and they need different answers:

1. "Should we have launched here?" This needs a Boise where you never launched, which doesn't exist and can't be rebuilt from four weeks of data. It's the hardest of the three and rarely the live one.
2. "Is this launch ahead of or behind where it should be?" This is usually what people mean, and it's answerable. The comparison isn't Boise against Boise-without-us. It's Boise against what a launch like this normally looks like at week 4.
3. "Did the thing we did in Boise work?" A promotion, a fee change, a courier incentive. That's an ordinary causal question that happens to sit inside a new city.

Agreeing which question is on the table comes before choosing a method. Promising a causal number for the first question is how a launch ends up judged on noise.