eBay Data Scientist Interview Guide

eBay's data science interview barely touches deep learning, but it will punish you for not understanding two-sided marketplace dynamics. The candidates who struggle aren't missing technical chops. They can't articulate what happens to buyer conversion when you tweak a seller incentive, or why a search ranking win in Electronics might be a loss in Collectibles.

eBay Data Scientist Role

You sit inside a product team (Search & Discovery, Seller Experience, Trust & Safety, Ads & Monetization, or Buyer Success) and own the metrics, experiments, and analytical narratives for that area. The work runs on SQL against eBay's warehouse, Python in Jupyter notebooks, and dashboarding tools like Tableau or Looker. Success after year one means you've shipped experiment readouts that changed a PM's roadmap and earned enough context that stakeholders pull you in before they've decided what to build.

A Typical Week

The breakdown won't shock you until you notice how much time goes to scoping, not executing. Monday afternoon you're translating a vague PM ask ("why did listing completions drop in this category?") into a concrete analysis plan with defined metrics and a timeline. That unglamorous translation work is where experienced DSs at eBay separate themselves from people who just run queries.

Projects & Impact Areas

Trust & Safety produces some of eBay's hardest DS problems because the open marketplace (anyone can list anything) creates adversarial scenarios like fake luxury goods and coordinated seller rings that a first-party retailer never faces. Search ranking presents a different puzzle: matching a buyer searching "1967 Camaro SS hood" to the right listing among billions of unique, long-tail items that commodity recommendation systems aren't designed for. Meanwhile, seller economics work (modeling how a promoted listings fee change affects participation and whether incremental ad revenue cannibalizes organic search quality) leans heavily on causal inference rather than prediction.

Skills & What's Expected

Statistics and business acumen are weighted far above ML sophistication here. The skill most candidates over-index on is deep learning, which matters only for specific ranking or NLP teams, not the broader DS org. The most underrated skill is designing experiments on a two-sided marketplace where treating sellers changes the buyer experience, a SUTVA violation that most textbook A/B testing prep ignores. SQL proficiency means window functions, self-joins for sessionization, and queries that perform well at warehouse scale. ML expectations scale with level and team: a fraud classifier or demand forecast model is sufficient for many roles, but senior loops (T25+) may include a deep dive on model evaluation, feature engineering, and production considerations.

Levels & Career Growth

The widget shows the level bands. What it won't tell you is that the T25-to-T26 jump is where most careers stall. The blocker, based on what the promo criteria describe, isn't technical depth. It's cross-team influence: setting measurement standards other DS teams adopt, or shaping a product roadmap through your analysis. T27 (Principal) roles are scarce and often resemble applied research leads focused on ranking algorithms or marketplace experimentation methodology.

Work Culture

eBay requires three days per week in the San Jose office, with most teams clustering Tuesday through Thursday on-site and keeping Monday or Friday flexible for remote deep work. The pace is calmer than Meta or Amazon: you'll work roughly 9-to-5:30 most weeks, with occasional crunches around quarterly business reviews or major experiment launches. Bureaucracy can slow experiment approvals, but the tradeoff is that your analysis gets the time and scrutiny to actually be defensible before it reaches a VP.

eBay Data Scientist Compensation

eBay's RSUs vest 25% per year over four years, paid out in quarterly chunks of 6.25%. No backloading, no cliff before your first quarterly vest hits. The simplicity is the selling point: your Year 1 comp is predictable, unlike Amazon's 5/15/40/40 structure where early-year cash has to fill the gap.

When negotiating, the offer notes themselves hint at the playbook: equity can often be rebalanced upward if base is capped by band. Ask for a larger initial RSU grant instead of grinding over a few thousand in base, because that grant pays out across four years of vesting. A signing bonus is also worth requesting explicitly, since it bridges the months before your first quarterly RSU vest lands in your brokerage account.

eBay Data Scientist Interview Process

The number one reason candidates wash out is sloppy metric definitions. Not just in the Product Sense round, either. Interviewers across SQL, experimentation, and modeling all probe whether you can specify precise denominators, handle edge cases like returns and bot traffic, and reason about time windows for a marketplace where buyer and seller metrics often conflict. If you can't cleanly define "conversion rate" for a listing page that serves both auction and fixed-price formats, that weakness shows up everywhere.

Most candidates underestimate the Hiring Manager Screen's filtering power. At many companies this round is a vibe check. At eBay, from what candidates report, the hiring manager digs into how you framed ambiguous marketplace problems and chose success metrics in past work. Getting past this screen without a concrete example of driving a decision with data (not just producing an analysis) is unlikely. Prepare for it like a technical round, because it functions as one.

eBay Data Scientist Interview Questions

1WITH first_exposure AS (
2  SELECT
3    user_id,
4    experiment_id,
5    -- Use the variant at the earliest exposure.
6    MIN_BY(variant, exposure_ts) AS variant,
7    MIN(exposure_ts) AS first_exposure_ts,
8    CAST(MIN(exposure_ts) AS DATE) AS exposure_date
9  FROM experiment_exposures
10  GROUP BY 1, 2
11), purchasers_7d AS (
12  SELECT
13    fe.experiment_id,
14    fe.variant,
15    fe.exposure_date,
16    fe.user_id,
17    1 AS purchased_7d
18  FROM first_exposure fe
19  JOIN orders o
20    ON o.buyer_id = fe.user_id
21   AND o.order_ts >= fe.first_exposure_ts
22   AND o.order_ts < fe.first_exposure_ts + INTERVAL '7' DAY
23  GROUP BY 1, 2, 3, 4
24)
25SELECT
26  fe.experiment_id,
27  fe.variant,
28  fe.exposure_date AS event_date,
29  COUNT(DISTINCT fe.user_id) AS exposed_users,
30  COUNT(DISTINCT p.user_id) AS purchasers_7d,
31  CAST(COUNT(DISTINCT p.user_id) AS DOUBLE) / NULLIF(COUNT(DISTINCT fe.user_id), 0) AS conversion_7d
32FROM first_exposure fe
33LEFT JOIN purchasers_7d p
34  ON p.experiment_id = fe.experiment_id
35 AND p.variant = fe.variant
36 AND p.exposure_date = fe.exposure_date
37 AND p.user_id = fe.user_id
38GROUP BY 1, 2, 3
39ORDER BY 1, 2, 3;

1WITH per_listing_day AS (
2  SELECT
3    CAST(event_ts AS DATE) AS event_date,
4    listing_id,
5    -- Flags prevent double counting when there are many events.
6    MAX(CASE WHEN event_type = 'view' THEN 1 ELSE 0 END) AS is_active,
7    MAX(CASE WHEN event_type = 'policy_removed' THEN 1 ELSE 0 END) AS is_removed
8  FROM listing_events
9  GROUP BY 1, 2
10)
11SELECT
12  event_date,
13  SUM(is_active) AS active_listings,
14  SUM(CASE WHEN is_active = 1 AND is_removed = 1 THEN 1 ELSE 0 END) AS removed_active_listings,
15  CAST(SUM(CASE WHEN is_active = 1 AND is_removed = 1 THEN 1 ELSE 0 END) AS DOUBLE)
16    / NULLIF(SUM(is_active), 0) AS defect_rate
17FROM per_listing_day
18GROUP BY 1
19ORDER BY 1;

Experimentation, product sense, and applied statistics together consume over 60% of the interview, which tells you eBay hires for people who can reason about measurement on a two-sided marketplace, not people who can tune XGBoost hyperparameters. These areas compound in practice: a product metrics question about a seller trust badge will slide into designing the A/B test, then into whether your statistical approach holds when buyers browse listings from both treatment and control sellers in a single session. The biggest prep mistake this distribution implies is spending equal time on all six areas when ML and SQL together make up only a quarter of the weight.

Practice eBay-style questions across all six areas at datainterview.com/questions.

How to Prepare for eBay Data Scientist Interviews

eBay's Q4 2025 earnings showed $11.1B in revenue, up 15% YoY, while headcount shrank roughly 6.5%. That math tells you where the company is headed: doing more with fewer people, powered by AI and sharper data science. Their focus category strategy channels investment into luxury, trading cards, refurbished goods, and auto parts, which means DS teams are building category-specific models for search relevance, pricing, and fraud rather than one-size-fits-all pipelines.

The "why eBay" answer that falls flat is any version of "I want to work on a marketplace at scale." eBay's interviewers perk up when you name the specific two-sided tension their platform creates: every listing is user-generated, so Trust & Safety problems look fundamentally different than on platforms with curated inventory. Reference how their seller tools launches aim to reduce friction while maintaining buyer trust, and explain why that tradeoff excites you analytically. Showing you've thought about how a fraud model's false positive rate directly erodes seller retention on an open marketplace will separate you from candidates who rehearsed a generic answer.

Try a Real Interview Question

1WITH base AS (
2  SELECT
3    ea.experiment_id,
4    ea.variant,
5    ea.user_id,
6    ea.exposed_at,
7    CASE
8      WHEN EXISTS (
9        SELECT 1
10        FROM orders o
11        WHERE o.user_id = ea.user_id
12          AND o.order_ts >= ea.exposed_at
13          AND o.order_ts < ea.exposed_at + INTERVAL '7' DAY
14      ) THEN 1 ELSE 0
15    END AS converted
16  FROM experiment_assignments ea
17  WHERE ea.experiment_id = 'trust_checkout_v1'
18),
19agg AS (
20  SELECT
21    variant,
22    COUNT(*) AS n_users,
23    SUM(converted) AS n_converted,
24    CAST(SUM(converted) AS DOUBLE PRECISION) / NULLIF(COUNT(*), 0) AS p
25  FROM base
26  GROUP BY variant
27),
28paired AS (
29  SELECT
30    MAX(CASE WHEN variant = 'control' THEN n_users END) AS n_c,
31    MAX(CASE WHEN variant = 'control' THEN n_converted END) AS x_c,
32    MAX(CASE WHEN variant = 'control' THEN p END) AS p_c,
33    MAX(CASE WHEN variant = 'treatment' THEN n_users END) AS n_t,
34    MAX(CASE WHEN variant = 'treatment' THEN n_converted END) AS x_t,
35    MAX(CASE WHEN variant = 'treatment' THEN p END) AS p_t
36  FROM agg
37)
38SELECT
39  a.variant,
40  a.n_users,
41  a.n_converted,
42  a.p AS conversion_rate_7d,
43  (p.p_t - p.p_c) AS lift_treatment_minus_control,
44  ((p.p_t - p.p_c) /
45    NULLIF(
46      SQRT(
47        ((CAST(p.x_t + p.x_c AS DOUBLE PRECISION) / NULLIF((p.n_t + p.n_c), 0))
48         * (1.0 - (CAST(p.x_t + p.x_c AS DOUBLE PRECISION) / NULLIF((p.n_t + p.n_c), 0)))
49         * (1.0 / NULLIF(p.n_t, 0) + 1.0 / NULLIF(p.n_c, 0))
50        )
51      ),
52      0
53    )
54  ) AS z_statistic
55FROM agg a
56CROSS JOIN paired p
57ORDER BY CASE WHEN a.variant = 'control' THEN 0 ELSE 1 END, a.variant;

eBay's SQL rounds, from what candidates report, lean into transaction-level data where you need to reason about both sides of a marketplace interaction in a single query. The problem above captures that flavor: you're not just aggregating, you're thinking about relationships between entities. Build fluency with similar problems at datainterview.com/coding.

Test Your Readiness

Identify your weak spots, then drill the specific topic areas where you're least confident using datainterview.com/questions.