Warner Bros. Data Scientist Interview Guide

From hundreds of mock interviews for media-company DS roles, one pattern stands out at Warner Bros. specifically: candidates who nail the LightGBM churn model but can't walk a VP of Content through the SHAP plot explaining why single-franchise viewers are highest risk. WBD treats communication as a gating skill, not a soft bonus, and their interview loop reflects it.

Warner Bros. Data Scientist Role

You're sitting between two businesses that run on different clocks: a profitable but shrinking linear TV portfolio (HBO cable, Discovery networks, CNN) and Max, the streaming platform still scaling toward profitability. Your core mandate is time series forecasting and causal modeling for subscription KPIs, things like projecting subscriber growth around an HBO original premiere versus a new sports rights deal, then reporting those projections to leadership in ways that actually shape budget and content decisions.

A Typical Week

What jumps out isn't the modeling time. It's how much of your week goes to writing experiment design docs in Confluence, updating SHAP documentation on model cards, and chasing down broken Snowflake loads that feed someone else's Tableau dashboard on the linear TV side. If you picture this role as heads-down Databricks notebooks all day, recalibrate: stakeholder syncs and pipeline firefighting eat real hours.

Projects & Impact Areas

Subscription forecasting anchors the work, where you're building time series models that project Max subscriber counts across different content release cycles and feeding those numbers into leadership reporting. That forecasting naturally connects to causal inference problems: when marketing launches a win-back campaign or a bundle offer, you're designing the uplift analysis to separate incremental conversions from organic resubscribers. Experimentation rounds it out, with A/B test design for Max product changes (ad-tier funnel tweaks, homepage layout shifts) requiring you to own everything from randomization unit selection to minimum detectable effect calculations alongside product managers and staff engineers.

Skills & What's Expected

ML is rated "expert" level, but the skill most candidates under-prepare for is data pipelines, which WBD rates equally high. You'll build and maintain production workflows across Snowflake and Databricks, implement anomaly detection for data quality, and troubleshoot upstream schema changes from the linear TV side. Meanwhile, GenAI/LLM experience is scored low. WBD wants clean SHAP summary plots you can explain to a non-technical exec, not prompt engineering demos.

Levels & Career Growth

Most external hires land at P2 or P3. The P3-to-P4 jump is where people stall, because the bar shifts from "owns a project end-to-end" to "sets measurement standards across teams and influences roadmaps you don't directly control." If you're eyeing Principal, the path rewards domain ownership (becoming the subscriber forecasting authority) over generalist breadth.

Work Culture

WBD operates hybrid, with in-office expectations varying by location. The pace feels media-industry, not Silicon Valley: steady most weeks, then intense around tentpole launches like a new HBO original premiere or Max picking up live sports rights. Genuine cross-brand exposure (HBO, Discovery, CNN teams who think about audiences in fundamentally different ways) is a real perk, though legacy data silos from the merger still create friction when you're reconciling schemas that were never designed to coexist.

Warner Bros. Data Scientist Compensation

WBD's equity component is worth scrutinizing. Stock grants don't appear until P2, and the vesting schedule isn't publicly documented, so ask your recruiter for the exact cliff, back-loading structure, and refresh grant cadence before you evaluate any offer. The safest move is to negotiate for a higher base or a guaranteed first-year sign-on bonus rather than accepting a larger equity allocation you can't fully model.

The biggest negotiation lever most candidates overlook is leveling, not dollars within a band. WBD's P3 and P4 bands overlap significantly, so framing your experience around the scope descriptors in the next level up (owning measurement frameworks end-to-end, mentoring junior scientists, driving cross-functional adoption of model outputs) can shift you into a higher band entirely. Confirm bonus payout timing, year-one bonus guarantee, and hybrid expectations at the Hyderabad office in writing before you sign.

Warner Bros. Data Scientist Interview Process

Candidates report uneven communication between rounds, and the process can drag past the expected window when recruiters go quiet. Send a short follow-up after each stage referencing something specific you discussed, because WBD's hiring coordination across its streaming and linear segments isn't always tightly synced.

Shallow experimentation fundamentals are among the most common reasons candidates get cut. People over-index on ML prep and stumble when asked to walk through power analysis or explain practical vs. statistical significance for something like a Max free-trial offer test. The other quiet killer: the final panel includes non-technical partners from content, marketing, or product who need to believe you can turn a model output into a decision about, say, whether to renew a Discovery+ original. Strong technicals won't save you if those panelists aren't convinced.

Warner Bros. Data Scientist Interview Questions

1/*
2Daily active paid subscribers over the last 30 days.
3Assumptions:
4- subscriptions.status = 'paid' means paid entitlement (exclude trials).
5- Subscription is active for timestamps in [start_ts, end_ts), end_ts can be NULL for ongoing.
6- watch_events can have many rows per user per day, so de-duplicate to user-day.
7- SQL written to run on common warehouses (Snowflake, Databricks SQL) with minor syntax differences.
8*/
9
10WITH params AS (
11  SELECT
12    /* Use CURRENT_DATE for date grain, adjust if your warehouse uses CURRENT_DATE() */
13    CURRENT_DATE AS as_of_date,
14    DATEADD(day, -29, CURRENT_DATE) AS start_date
15),
16days AS (
17  SELECT DATEADD(day, seq4(), p.start_date) AS day
18  FROM params p,
19  TABLE(GENERATOR(ROWCOUNT => 30))
20),
21watch_user_day AS (
22  SELECT
23    CAST(event_ts AS DATE) AS day,
24    user_id
25  FROM watch_events
26  WHERE CAST(event_ts AS DATE) BETWEEN (SELECT start_date FROM params) AND (SELECT as_of_date FROM params)
27  GROUP BY 1, 2
28),
29paid_coverage_user_day AS (
30  SELECT
31    d.day,
32    s.user_id
33  FROM days d
34  JOIN subscriptions s
35    ON s.status = 'paid'
36   AND d.day >= CAST(s.start_ts AS DATE)
37   AND d.day < CAST(COALESCE(s.end_ts, DATEADD(day, 1, (SELECT as_of_date FROM params))) AS DATE)
38  GROUP BY 1, 2
39)
40SELECT
41  d.day,
42  COUNT(DISTINCT w.user_id) AS active_paid_subs
43FROM days d
44JOIN watch_user_day w
45  ON w.day = d.day
46JOIN paid_coverage_user_day p
47  ON p.day = d.day
48 AND p.user_id = w.user_id
49GROUP BY 1
50ORDER BY 1;

1import numpy as np
2import pandas as pd
3
4
5def add_features_and_wape(df: pd.DataFrame) -> tuple[pd.DataFrame, pd.DataFrame]:
6    """Create leakage-safe time-series features per (country, plan) and compute 28-day WAPE.
7
8    Expected columns:
9      - date, country, plan
10      - trials_started, paid_starts, cancels, active_subs, marketing_spend
11      - y_true, y_pred (for metric)
12
13    Returns:
14      - df_fe: original df with added feature columns
15      - wape_28d: WAPE aggregated per (country, plan) over the last 28 dates in the frame
16    """
17    df = df.copy()
18
19    # Parse and sort for deterministic rolling/lag behavior.
20    df["date"] = pd.to_datetime(df["date"])
21    df = df.sort_values(["country", "plan", "date"]).reset_index(drop=True)
22
23    grp = df.groupby(["country", "plan"], sort=False)
24
25    # Calendar feature.
26    df["dow"] = df["date"].dt.dayofweek.astype("int16")  # 0=Mon ... 6=Sun
27
28    # Leakage-safe rolling means: use shift(1) so today's target does not enter today's features.
29    for col in ["paid_starts", "cancels"]:
30        df[f"{col}_roll7_mean"] = (
31            grp[col]
32            .transform(lambda s: s.shift(1).rolling(window=7, min_periods=1).mean())
33            .astype("float32")
34        )
35
36    # Pure lag feature.
37    df["marketing_spend_lag7"] = grp["marketing_spend"].transform(lambda s: s.shift(7)).astype("float32")
38
39    # Metric: WAPE over the last 28 days per series.
40    # Keep only the last 28 dates per (country, plan) based on ordering.
41    df["_row_num"] = grp.cumcount()
42    df["_n"] = grp["_row_num"].transform("max") + 1
43    df["_is_last_28"] = df["_row_num"] >= (df["_n"] - 28)
44
45    metric_df = df.loc[df["_is_last_28"] & df["y_true"].notna() & df["y_pred"].notna(),
46                       ["country", "plan", "y_true", "y_pred"]].copy()
47
48    metric_df["abs_err"] = (metric_df["y_true"] - metric_df["y_pred"]).abs()
49    metric_df["abs_true"] = metric_df["y_true"].abs()
50
51    agg = metric_df.groupby(["country", "plan"], as_index=False).agg(
52        sum_abs_err=("abs_err", "sum"),
53        sum_abs_true=("abs_true", "sum"),
54        n=("abs_err", "size"),
55    )
56
57    # Guard against divide-by-zero when the 28-day actual sum is zero.
58    agg["wape_28d"] = np.where(agg["sum_abs_true"] > 0, agg["sum_abs_err"] / agg["sum_abs_true"], np.nan)
59
60    # Cleanup helper cols.
61    df_fe = df.drop(columns=["_row_num", "_n", "_is_last_28"])
62    return df_fe, agg
63
64
65# Example usage (df must already exist):
66# df_fe, wape_28d = add_features_and_wape(df)
67# print(wape_28d.sort_values("wape_28d"))
68

The distribution skews heavily toward forecasting and causal reasoning, which tells you WBD wants people who can answer "what will happen to churn when we drop a new HBO original?" and then prove whether that content drop actually caused the retention lift. The compounding trap is that WBD's subscription data carries structural breaks (price tier launches, sports rights deals like the Olympics) that make both forecasting and causal identification harder simultaneously, so prepping these as separate textbook topics leaves you exposed when an interviewer hands you a messy Max scenario that demands both. Pipeline and SQL questions may look like the lighter slice, but they're flavored around subscription event schemas with reactivation edge cases and silent data failures, not the generic warehouse design problems most candidates drill.

Practice Warner Bros. questions across all six topic areas at datainterview.com/questions.

How to Prepare for Warner Bros. Data Scientist Interviews

Warner Bros. Discovery (the full corporate entity behind the "Warner Bros." brand) formally split into two divisions in 2025: Global Linear Networks and Streaming & Studios. That restructuring is the single most important context for your prep, because DS teams now operate within business units that have different data schemas, different KPIs, and different stakeholders. The streaming side posted record viewership during the 2026 Winter Olympics, while the linear side is focused on positioning itself as a one-stop shop for advertisers heading into the 2026/2027 marketplace.

Your "why Warner Bros. Discovery?" answer should center on that dual-segment data problem. Talk about how the DAISY text-to-SQL tool was built to let analysts query across heterogeneous sources, or how the recommendation engine described in their Stack Overflow podcast has to serve both Max originals and Discovery+ unscripted catalogs. That specificity signals you've studied the actual infrastructure, not just the content library.

Try a Real Interview Question

1WITH params AS (
2  SELECT region, CAST(change_date AS DATE) AS price_change_date
3  FROM price_changes
4),
5-- Convert event stream into active intervals per user.
6-- Assumptions:
7-- 1) 'subscribe' starts an active interval at event_date.
8-- 2) 'cancel' ends activity at event_date (user is not active on that date).
9-- 3) Users can have multiple subscribe/cancel cycles.
10ordered AS (
11  SELECT
12    se.user_id,
13    se.region,
14    CAST(se.event_date AS DATE) AS event_date,
15    se.event_type,
16    LEAD(CAST(se.event_date AS DATE)) OVER (
17      PARTITION BY se.user_id, se.region
18      ORDER BY CAST(se.event_date AS DATE)
19    ) AS next_event_date
20  FROM subscription_events se
21),
22intervals AS (
23  SELECT
24    user_id,
25    region,
26    event_date AS start_date,
27    -- If next event is a cancel, treat cancel date as end boundary (exclusive).
28    -- If missing next event, keep it open-ended.
29    CASE
30      WHEN next_event_date IS NULL THEN CAST('9999-12-31' AS DATE)
31      ELSE next_event_date
32    END AS end_date_exclusive
33  FROM ordered
34  WHERE event_type = 'subscribe'
35),
36cohort AS (
37  SELECT
38    p.region,
39    p.price_change_date,
40    i.user_id
41  FROM params p
42  JOIN intervals i
43    ON i.region = p.region
44   AND (p.price_change_date - INTERVAL '1 day') >= i.start_date
45   AND (p.price_change_date - INTERVAL '1 day') <  i.end_date_exclusive
46  GROUP BY p.region, p.price_change_date, i.user_id
47),
48retained AS (
49  SELECT
50    c.region,
51    c.price_change_date,
52    c.user_id
53  FROM cohort c
54  JOIN intervals i
55    ON i.region = c.region
56   AND i.user_id = c.user_id
57   AND (c.price_change_date + INTERVAL '7 day') >= i.start_date
58   AND (c.price_change_date + INTERVAL '7 day') <  i.end_date_exclusive
59  GROUP BY c.region, c.price_change_date, c.user_id
60)
61SELECT
62  c.region,
63  c.price_change_date,
64  COUNT(DISTINCT c.user_id) AS cohort_size,
65  COUNT(DISTINCT r.user_id) AS retained_7d,
66  CASE
67    WHEN COUNT(DISTINCT c.user_id) = 0 THEN 0.0
68    ELSE 1.0 * COUNT(DISTINCT r.user_id) / COUNT(DISTINCT c.user_id)
69  END AS retention_rate
70FROM cohort c
71LEFT JOIN retained r
72  ON r.region = c.region
73 AND r.price_change_date = c.price_change_date
74 AND r.user_id = c.user_id
75GROUP BY c.region, c.price_change_date
76ORDER BY c.region, c.price_change_date;

Warner Bros. Discovery's DS interviews lean into subscription and engagement analytics, so the SQL you'll write looks like real Max business logic: cohort retention, rolling churn windows, funnel breakdowns segmented by content type. Sharpen that muscle at datainterview.com/coding, focusing on queries where the business context matters as much as the syntax.

Test Your Readiness

Identify your blind spots before the real loop does. datainterview.com/questions lets you drill the specific topic areas Warner Bros. Discovery weights most heavily.