Interview experience
Tiger Analytics Interview Questions and Answers (2026)
Overview
Section titled “Overview”Tiger Analytics runs an analytics-consulting-flavored loop - Python/SQL and statistics fundamentals in the OA and early technical rounds, then case-style reasoning over messy data rather than classic DSA-heavy coding interviews.
Tiger Analytics interview process at a glance
Section titled “Tiger Analytics interview process at a glance”| Round | Duration | What they test |
|---|---|---|
| Online Assessment | ~60 min | Python (list comprehensions, Pandas), SQL (joins, window functions), basic DSA |
| Technical Interview 1 | 30-45 min | Python/SQL proficiency, statistics fundamentals, problem solving |
| Technical Interview 2 | 30-45 min | Case-study/project deep-dive, ML fundamentals (data-science track), data reasoning |
| Behavioral/HR Round | 20-30 min | Ambiguity handling, stakeholder communication, culture fit |
Online Assessment
Section titled “Online Assessment”A timed HackerEarth test blending Python scripting, SQL querying, and data-structure challenges - closer to a data-analyst screen than a pure DSA gauntlet.
Common questions
- Python - list comprehensions, Pandas transformations, string/dictionary manipulation
- SQL - joins, GROUP BY/aggregations, window functions
- Basic data-structure problems - arrays, loops, simple algorithmic logic
Technical Interview 1
Section titled “Technical Interview 1”A 30-45 minute round testing hands-on Python/SQL fluency plus core statistics, with problem-solving woven throughout.
Common questions
- Write a SQL query using joins and window functions to answer a business question
- Python - manipulate a dataset using Pandas (filter, group, aggregate)
- Explain a statistical concept you’ve used in a project - correlation, hypothesis testing, distributions
- Basic probability puzzles (e.g. classic dice/coin-style problems)
Technical Interview 2
Section titled “Technical Interview 2”A deeper round mixing a case study or project deep-dive with, for data-science-track roles, ML fundamentals. The interviewer cares more about how you frame an ambiguous problem than whether you reach one “correct” answer.
Common questions
- Walk through a case study analyzing business metrics and recommend an action
- Supervised vs unsupervised learning - explain with examples
- Bias-variance tradeoff and common classification metrics (precision, recall, F1)
- How would you approach a dataset with missing or inconsistent values?
Round-by-round breakdowns, including real candidate reports, are on the Tiger Analytics interview experience page.
Behavioral/HR Round
Section titled “Behavioral/HR Round”A STAR-based conversation on handling ambiguity, stakeholder conflict, and rapid pivots under deadline pressure - realities of consulting-style analytics delivery.
Common questions
- Tell me about yourself
- Why Tiger Analytics?
- Walk me through how you’d approach a case study where the data is messy or incomplete
- Tell me about a time your analysis contradicted what stakeholders expected or wanted to hear - how did you communicate it?
Sample answer frameworks for each of these are on the Tiger Analytics HR interview questions page.
Common technical interview questions and answers
Section titled “Common technical interview questions and answers”Q: Write a SQL query returning the top two products by revenue within each category.
Use a window function: SELECT category, product, revenue FROM (SELECT category, product, SUM(amount) AS revenue, DENSE_RANK() OVER (PARTITION BY category ORDER BY SUM(amount) DESC) AS rnk FROM sales GROUP BY category, product) t WHERE rnk is at most 2. PARTITION BY restarts the ranking within each category, which is what LIMIT cannot express per group. Choose the ranking function deliberately - ROW_NUMBER breaks ties arbitrarily, RANK leaves gaps after a tie, and DENSE_RANK does not - because interviewers usually follow up on ties. Note that window functions run after GROUP BY, so ranking on an aggregate inside the same query level is legal.
Q: How would you compute a 7-day rolling average of daily sales per store?
In SQL: SELECT store_id, sale_date, AVG(total) OVER (PARTITION BY store_id ORDER BY sale_date ROWS BETWEEN 6 PRECEDING AND CURRENT ROW) FROM daily_sales. The critical subtlety is that ROWS counts physical rows, so a store with a missing day silently averages over a wider calendar span - use RANGE BETWEEN INTERVAL ‘6’ DAY PRECEDING AND CURRENT ROW, or join against a calendar table to fill gaps with zero first. In Pandas the equivalent is df.set_index(‘sale_date’).groupby(‘store_id’)[‘total’].rolling(‘7D’).mean(), where the string offset makes it time-based rather than row-based. Mention that the first six rows are a partial window, so you may want min_periods set to 7 to produce NaN instead of a misleading average.
Q: What is a p-value, and what does a result significant at the 0.05 level actually mean?
A p-value is the probability of observing a test statistic at least as extreme as the one you got, assuming the null hypothesis is true. A p-value of 0.03 therefore means such data would arise 3 percent of the time under the null - it is not the probability that the null is true, and not the probability your finding is a fluke. Rejecting at 0.05 fixes the Type I error rate at 5 percent, which is why testing twenty metrics on the same experiment yields a false positive roughly once by chance, and needs a Bonferroni or false-discovery-rate correction. Statistical significance is also not practical significance: with a million users, a 0.01 percent lift will be significant and still worthless, so always report the effect size and confidence interval alongside the p-value.
Q: How would you explain the bias-variance tradeoff to a business stakeholder?
Bias is error from a model too simple to capture the real pattern - it is systematically wrong in the same direction, and it underfits. Variance is error from a model so flexible that it fits the noise in the training sample, so it swings wildly when the data changes and fails on new records - it overfits. Total expected error decomposes into bias squared plus variance plus irreducible noise, so reducing one typically increases the other and there is an optimal middle complexity. The business framing is that a high-bias model gives consistently mediocre predictions while a high-variance model looks excellent in backtesting and disappoints in production, which is exactly why you hold out a validation set and use cross-validation. Regularisation, more training data, and ensembling are the practical levers.
Q: When would you optimise for recall instead of precision, and what does F1 measure?
Precision is the share of predicted positives that are truly positive, so it answers how much you can trust an alert; recall is the share of actual positives you caught, so it answers how much you are missing. Optimise recall when a missed positive is far costlier than a false alarm - fraud screening, disease detection, or churn intervention where the follow-up is cheap. Optimise precision when acting on a false positive is expensive or annoying, such as blocking a legitimate transaction. F1 is the harmonic mean, 2 times precision times recall divided by precision plus recall, which punishes a model that scores well on one and badly on the other. On imbalanced data quote PR-AUC rather than accuracy, because predicting the majority class always gives high accuracy and zero value.
Q: You are handed a dataset with 30 percent missing values in a key column. How do you proceed?
First diagnose the mechanism, because the fix depends on it: missing completely at random means dropping rows is unbiased but wasteful; missing at random means the missingness is explained by other observed columns, so model-based imputation works; missing not at random - income missing precisely because it is high - biases any imputation and needs the missingness modelled explicitly. Check whether missingness correlates with the target, since it is often predictive in itself, and add a binary was_missing indicator. For a numeric column, median imputation is the robust baseline and iterative or KNN imputation the better option; for a categorical column, treat missing as its own level. Never impute using statistics computed over the full dataset before splitting, since that leaks test information into training.
Q: How would you estimate the number of food-delivery orders placed in Bengaluru on a weekday?
Structure it top-down and state assumptions out loud. Bengaluru has roughly 13 million people, about 3.5 million households averaging 3 to 4 people. Restrict to the addressable segment: perhaps 40 percent are urban households with smartphones and disposable income, giving about 1.4 million. Assume an ordering household places roughly 2 orders a week, so about 0.3 orders a day, giving roughly 400,000 to 450,000 orders, with each order serving 1.5 people. Sanity-check the number from the supply side - if there are around 25,000 active restaurants, that is 16 to 18 orders per restaurant per day, which is plausible - and name which assumption the answer is most sensitive to, since the interviewer cares about the structure and the sensitivity check far more than the final figure.
Q: Give an example of correlation being mistaken for causation, and how you would test the causal claim.
Ice cream sales correlate with drowning deaths because temperature drives both - a confounder. In an analytics engagement the same shape appears when a loyalty programme’s members spend more than non-members, since customers who self-select into the programme were already high spenders, which is selection bias rather than a programme effect. The gold standard test is a randomised controlled experiment: randomly assign eligible customers to treatment and control, check that pre-period metrics balance, and measure the difference in the target metric with a confidence interval. When randomisation is impossible, use a quasi-experimental design - difference-in-differences against a comparable control group, a regression discontinuity around an eligibility cutoff, or propensity-score matching - and be explicit that these rest on assumptions the data cannot verify.
Frequently asked questions about Tiger Analytics interviews
Section titled “Frequently asked questions about Tiger Analytics interviews”What is the Tiger Analytics interview process for freshers?
Tiger Analytics typically runs an Online Assessment followed by two technical interviews and a behavioral/HR round, though some drives extend to 5-6 rounds depending on the role. The OA (~60 minutes, often on HackerEarth) covers Python scripting, SQL querying, and data-structure problems; the technical rounds dig into Python/SQL proficiency, statistics and ML fundamentals, and case-style or project deep-dives; the final round is a STAR-based behavioral discussion.
What kind of questions does Tiger Analytics ask?
The OA blends list comprehensions, Pandas transformations, and SQL joins/window-function questions. Technical interviews lean on Python/SQL proficiency, statistics fundamentals (distributions, hypothesis testing), and for data-science-track roles, ML concepts like supervised vs unsupervised learning, bias-variance tradeoff, and classification metrics - plus guided case studies where you reason through messy or incomplete data rather than land a single ‘right’ answer, reflecting the consulting-style nature of the work.
How many rounds are there in Tiger Analytics interviews?
Most candidates go through 4 rounds - an Online Assessment, two technical interviews, and a behavioral/HR round - though some drives run 5-6 rounds depending on seniority and business unit. Freshers can generally expect a shorter path than experienced hires, and the whole process typically spans 2-4 weeks.
How should I prepare for Tiger Analytics interviews?
Be comfortable with Python (list comprehensions, Pandas) and SQL (joins, window functions, GROUP BY), revise core statistics and, for data-science roles, ML fundamentals, and practice explaining your approach to an ambiguous or messy dataset out loud - Tiger Analytics’ case-style rounds care more about how you structure a problem than whether you land a single correct number.
Is Tiger Analytics’ interview different from a pure SDE loop?
Yes. Tiger Analytics is an analytics and AI consulting firm, so its technical rounds emphasize Python/SQL data manipulation, statistics, and case-study reasoning over classic DSA-heavy coding-interview questions. Expect basic data-structure coding in the OA, but the technical interviews look much closer to a data-analyst or data-scientist loop - working through a business problem with incomplete data - than a typical product-company SDE interview.
What does the behavioral/HR round at Tiger Analytics check for?
A STAR-based conversation exploring how you handle ambiguity, stakeholder pushback, and tight deadlines - common realities in consulting-style analytics work. Expect questions about a time your analysis contradicted what a stakeholder expected, alongside standard fit and motivation questions.