Autocoder

Tiger Analytics Interview Questions with Answers

Last updated 2026-11-19

Tiger Analytics is a data science and analytics consultancy, so its interviews are commonly described as a blend of statistics, machine learning, SQL and Python, with data engineering questions for engineering tracks and a close look at the projects on your resume. The aim is to see whether you can turn a messy business question into a defensible analysis.

The twelve questions below are practice in the style of those topics, with worked answers. They are not a record of any interview. For hands-on SQL, pandas and case practice see Tiger Analytics coding round questions, and for the role-level view read how to crack the data science interview.

Rounds (commonly reported)Screening test, technical interviews, project and HR discussion
TopicsStatistics, ML, SQL, Python, Spark, case reasoning
TracksData science, analytics, data engineering

Rounds as commonly reported

We hold no sourced description of the company's selection stages, so this is general guidance for analytics consultancies. Candidates commonly describe a screening stage covering aptitude, SQL or Python, one or two technical interviews that start from your resume and move into statistics, models and queries, and a conversation with a manager or HR about fit and expectations.

Answering like an analytics consultant

State the business question, the data you need, the method, the metric that proves success and the risks. Whenever you name a model, say why a simpler baseline would or would not do. Numbers in your answers should come from a worked example, not from memory of someone else's results.

Hiring status

We hold no sourced hiring announcements, drive dates, openings or process description for Tiger Analytics, so none are stated here. It has no page in our hiring section. Check the company's own careers site for current roles and eligibility.

Tiger Analytics Interview Questions with Answers: questions with worked answers

  1. 1

    Walk me through your most relevant project and the business impact.

    easy

    Use five beats: the business problem, the data (size, quality issues), your approach and why, the evaluation metric and result, and what was deployed or decided. Be precise about your own contribution versus the team's. Include one thing that went wrong, for example label leakage you found, and how you fixed it. Interviewers use this answer to choose their next ten minutes of questions, so mention only what you can defend in depth.

  2. 2

    Explain the bias-variance trade-off.

    easy

    Expected prediction error splits into bias squared, variance and irreducible noise. A simple model (linear regression on a curved relationship) has high bias: it underfits. A very flexible model (a deep unpruned tree) has high variance: it fits noise and changes a lot with the training sample. Total error is usually minimised at a middle complexity, found by cross-validation. Fixes for high variance: more data, regularisation, bagging, fewer features. Fixes for high bias: richer features or a more flexible model.

  3. 3

    What does a p-value mean, and what does it not mean?

    medium

    A p-value is the probability, assuming the null hypothesis is true, of seeing a result at least as extreme as the one observed. A p-value of 0.03 does not mean a 3 percent chance the null is true, nor that the effect is large or important. A small p-value with a huge sample can reflect a trivial effect, so report the effect size and a confidence interval. A 95 percent confidence interval is a procedure that captures the true value in 95 percent of repeated samples.

  4. 4

    You run an A/B test on a checkout change. How do you decide the sample size and avoid common mistakes?

    medium

    Fix the baseline conversion, the minimum detectable effect, the significance level (often 5 percent) and the power (often 80 percent) before starting. A rough rule for a continuous metric is n per group of about 16 times the variance divided by the square of the detectable difference. Decide the duration up front and avoid peeking and stopping on the first significant result, which inflates false positives. Randomise at the right unit (user, not session), check sample ratio mismatch, run across full weekly cycles, and watch guardrail metrics such as refunds.

  5. 5

    A model predicts fraud with 40 true positives, 10 false positives and 20 false negatives. Compute precision, recall and F1, and say which matters.

    medium

    Precision is TP/(TP+FP) = 40/50 = 0.80. Recall is TP/(TP+FN) = 40/60 = 0.667. F1 is the harmonic mean: 2 x 0.8 x 0.667 / (0.8 + 0.667) = 0.727. Which matters depends on cost: if missing fraud is expensive, optimise recall and accept more false alerts; if each alert costs an analyst's time, precision matters more. Choose the decision threshold using the precision-recall curve and business costs, and note accuracy is misleading on imbalanced data.

    python
    tp, fp, fn = 40, 10, 20
    precision = tp / (tp + fp)
    recall = tp / (tp + fn)
    f1 = 2 * precision * recall / (precision + recall)
    print(round(precision, 3), round(recall, 3), round(f1, 3))  # 0.8 0.667 0.727
  6. 6

    How do you handle a heavily imbalanced classification problem?

    medium

    Start by choosing the right metric (precision, recall, PR-AUC), not accuracy. Then try class weights in the loss, resampling only inside training folds (undersample the majority or oversample the minority, SMOTE with care), and threshold tuning on probability outputs. Use stratified cross-validation so each fold has the minority class. Resampling before splitting leaks duplicates into validation and inflates scores, a common mistake.

  7. 7

    Compare L1 and L2 regularisation.

    medium

    L2 (ridge) adds the sum of squared coefficients to the loss; it shrinks coefficients smoothly and handles correlated features by sharing weight. L1 (lasso) adds the sum of absolute values; its geometry drives some coefficients exactly to zero, giving feature selection. Elastic net blends both. Scale features before regularising, since the penalty depends on coefficient size, and choose the strength by cross-validation.

  8. 8

    How do you handle missing values?

    medium

    First understand why they are missing: completely at random, at random given other columns, or informative (a blank income may itself signal something). Options: drop rows or columns when few and random; impute with median or mode, group-wise means, or model-based methods; add a missing-indicator feature; or use algorithms that handle missing values natively. Fit imputers on the training set only, to avoid leakage, and compare the effect on validation performance.

  9. 9

    Write a SQL query for month-over-month revenue growth.

    medium

    Aggregate revenue per month, then use LAG to bring in the previous month and compute the percentage change, guarding against division by zero. Make sure the month series has no gaps or the lag compares the wrong periods. See SQL window function questions for related patterns.

    sql
    WITH m AS (
      SELECT DATE_TRUNC('month', order_date) AS month, SUM(amount) AS revenue
      FROM orders
      GROUP BY 1
    )
    SELECT month, revenue,
           100.0 * (revenue - LAG(revenue) OVER (ORDER BY month))
                 / NULLIF(LAG(revenue) OVER (ORDER BY month), 0) AS mom_pct
    FROM m
    ORDER BY month;
  10. 10

    Explain narrow and wide transformations in Spark, and repartition versus coalesce.

    hard

    A narrow transformation (map, filter) computes each output partition from one input partition and needs no data movement. A wide transformation (groupByKey, join, distinct, reduceByKey) needs data from many partitions and triggers a shuffle: disk and network cost, and a stage boundary. repartition(n) can increase or decrease partitions with a full shuffle; coalesce(n) only decreases, merging partitions without a full shuffle, but can leave them uneven. Prefer reduceByKey or aggregations over groupByKey, and broadcast the small side of a join to avoid shuffling the large side.

  11. 11

    Random forest versus gradient boosting: how do they differ and when would you pick each?

    medium

    A random forest trains many deep trees independently on bootstrap samples with random feature subsets and averages them, which cuts variance and is robust with little tuning. Gradient boosting builds shallow trees sequentially, each fitting the errors of the ensemble so far, which cuts bias and often wins on tabular accuracy but needs careful tuning (learning rate, depth, number of trees) and early stopping to avoid overfitting. Pick a forest for a quick, stable baseline; pick boosting when you need the last bit of accuracy and can tune.

  12. 12

    A manager asks why sales fell 12 percent last quarter. How do you start?

    medium

    Clarify the metric and baseline (year over year or quarter over quarter), then decompose: sales = traffic x conversion x average order value, split by region, product, channel and customer segment. Check data issues first (a tracking outage or a changed definition), then external factors (seasonality, competitors, pricing, stock-outs). Form two or three hypotheses, name the data that would confirm each, and report the likeliest driver with a recommended action and its uncertainty.

Frequently asked questions

What should I prepare for a Tiger Analytics interview?

Statistics (hypothesis tests, distributions), core ML and evaluation, SQL with window functions, Python and pandas, and for engineering tracks Spark. Prepare two projects in depth.

Is there a coding round?

Analytics firms commonly screen with SQL and Python questions, but we hold no sourced description of the current format. Confirm it with your recruiter.

How important are case studies?

Business reasoning is usually valued for consulting-style analytics roles. Practise structuring a problem, naming metrics and stating assumptions.

Where do I find current openings?

On the company's own careers site. We hold no sourced hiring dates for Tiger Analytics.

Are these real Tiger Analytics questions?

No, they are practice questions in the style of analytics interviews, not a record of any interview.

Related guides

More in Company Interview Questions

See all in this section