Went through the Two Sigma DS interview loop earlier this year. Not the quant research scientist track (that's a different, heavier process). I interviewed for a data scientist role on the analytics/engineering-support side. Here's exactly what each round covered.
SQL round: This was real SQL, not toy SQL. Two problems in 45 minutes. One was a window functions problem: compute rolling averages and flag anomalies across time-series data segmented by some categorical variable. The other was a joins + aggregation problem with a twist in the business logic. You need to be comfortable with CTEs, window functions (RANK, LAG/LEAD, PARTITION BY), and writing clean, readable queries. I'd practice LeetCode SQL medium to hard if you haven't recently.
Statistics / probability round: This was the round I was least prepared for. Questions ranged from classical stats (hypothesis testing, p-values, Type I vs Type II error) to more applied probability. One question was about designing an A/B test: what's your minimum detectable effect, how do you handle multiple comparisons, what do you do if the randomization is imperfect. Another was a probability problem I'd describe as bayes-adjacent. Not super hard but you can't wing it.
Case study: They gave me a scenario involving a dataset (described verbally, no actual data) and asked how I'd approach a specific business question. For me it was something like: trader performance has changed over this time window, how would you investigate why. I had to talk through data sources, feature selection, model choice, how I'd validate and present findings.
Behavioral: Lighter than I expected. Two questions on cross-functional collaboration, one on a time I had to simplify a complex analysis for a non-technical audience.
Overall: the SQL and stats rounds are the real differentiators. I've done DS interviews at a handful of other companies and Two Sigma's SQL depth was the highest. Come ready for window functions and have your stats fundamentals sharp.