The A/B test point is correct and worth being crisp about: you literally cannot randomize users to a world where the pair exists vs. doesn't, and even if you tried a soft launch to a subset, the order book liquidity would be so thin it wouldn't reflect steady-state quality. So your baseline has to come from somewhere else, either comparable pairs at launch or external reference data.

Spread tightening over time is an underrated success metric. A pair that launches with wide spreads and tightens over 30 days as market makers get comfortable is a healthy trajectory. One that stays wide suggests the market makers aren't finding it worth their while, which is a signal before volume collapses. Retention of traders who came in specifically for the new pair is also worth separating from overall retention because it tells you whether the listing actually expanded your addressable user base or just served people who were already there.

The affiliate fraud callout was the right move. At a crypto exchange the incentive structure is particularly bad because affiliates get paid on funded accounts and motivated fraudsters know exactly how to manufacture a funded account that looks real long enough to clear the attribution window. Last-click attribution on top of that is basically a fraud subsidy.

The cleanest cross-channel comparison framework I've seen is to run the same incrementality logic everywhere you can and then treat channels where you can't do user-level holdout as a separate class that gets geo-based measurement. You end up with two buckets, and you normalize both by incremental cost per funded account. The problem is that search and social often work together in the funnel, so your geo holdout for search is partially capturing social lift if you're running both in the same geos. You'd want to either stagger the holdout windows or suppress both channels in holdout geos, which gets operationally messy but is the honest answer.

Your instinct on volatility as a manipulation proxy was good, and the fact that it sparked a wash trading follow-up tells you they were probing for exactly that depth. For a DS role at Coinbase specifically, market microstructure isn't out of scope at all, their data science work sits right on top of order book data so they expect you to be comfortable there.

On the core metrics: bid-ask spread is the obvious entry point but you want to decompose it a bit. Quoted spread tells you the posted cost, realized spread tells you whether market makers are actually getting filled at those quotes or pulling them under pressure. If realized spread collapses relative to quoted spread, that's a sign the order book is thin and reactive. Depth at multiple price levels (say 0.1%, 0.5%, 1% from mid) matters because a pair can look liquid at the top of the book and then have a cliff. Slippage for a representative order size is the most honest single number you can give a product team.

For wash trading specifically, the signals I'd reach for are: trade size clustering (wash traders tend to be sloppy about using round lots repeatedly), self-trade patterns across account clusters, and volume that doesn't correlate with price movement or external market events. If BTC/USD moves 2% and your new pair moves 0% but volume spikes, something is off. You can also look at the ratio of volume to open interest if derivatives data is available, or compare reported volume against on-chain settlement flows for assets that have that transparency.

One thing I'd add that often gets missed: time-of-day liquidity distribution. A pair might look fine on daily aggregates but be completely illiquid outside of a two-hour window, which is a real problem for users hitting it at 3am.

Fraud rate and chargeback rate are the obvious ones and you were right to lead there. The thing I'd add is that KYC failure rate alone is a lagging signal. By the time it spikes you've already spent the money. A faster proxy is the ratio of signups to identity verification attempts in the first 48 hours, because bot-driven traffic tends to drop off hard at the document upload step. That's something you can monitor in near real-time.

On the mid-flight pause question, the operationally honest answer is that the DS rarely owns that call unilaterally. Pre-defined thresholds are good but you need to be clear about who gets paged, what the escalation path looks like, and whether the default is pause-all or pause-the-specific-channel. I've seen cases where one affiliate was driving all the fraud and killing the whole campaign was the wrong move. Having that logic written down before launch is the thing that actually saves you.

Your instinct on incremental funded accounts was solid. Raw signups are meaningless at a crypto exchange because the friction is in KYC and the first deposit, not the email capture. The attribution window thing you mentioned fumbling is actually where most people lose points, and the cleanest way to handle it upfront is to just state your assumption explicitly before anyone asks: something like 'I'm treating a 30-day post-click window as the conversion boundary, and here's why that fits a 4-week campaign.' Gets it off the table early.

On the geo holdout for broad search, I'd push yourself to be more specific next time about how you pick the holdout geos. Matched market design is the right frame, where you're pairing geos on pre-period volume, demographics, and baseline conversion rate before you ever touch the campaign. The failure mode is picking holdout geos that are structurally different and then your counterfactual is garbage. For Coinbase specifically, crypto adoption is weirdly geographic, so state-level variation in existing user density is a real confounder you'd want to control for. Synthetic control is the more rigorous version of this if you have enough pre-period time series, though explaining that live under pressure is genuinely hard.

The cannibalization piece is where a lot of people wave their hands and the interviewer at Coinbase was right to push. A more specific answer: pull the last 90 days of trading behavior for users who hold or have traded the assets in the new pair, then look at their activity in correlated pairs. If you're listing ETH/BTC and you already have ETH/USD and BTC/USD, you can estimate volume migration by looking at how traders on other exchanges shifted when that pair was listed, or by modeling the implied cross rate and seeing how much of your existing volume is already synthetically replicating it. It's not a perfect estimate but it's a real number you can defend. The point is to show you're thinking about net new volume vs. reshuffled volume, because the fee economics look very different depending on which one it is.