← Intuit Interview Insights

Intuit·Software Engineer·Technical Phone Screen·Intermediate

Intermediate
Jun 2026

Summary

Intuit product analyst interview with a meaty ML scenario around a tax-filing prediction model. The whole thing revolved around one feature and the questions kept pulling at different threads of the same problem, which I actually liked more than I expected.

Questions Asked (3)

Q1

A key feature in your model has a lot of zeros and missing values. How do you figure out why, and how do you handle each case differently during modeling?

Product Analytics & MetricsRoot Cause AnalysisData Modeling
Author's notes

This is where I spent the most time and probably rambled a bit.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by systematically diagnosing the root causes of zeros and missing values, distinguishing between true zeros, missing at random, and missing not at random. Then, for each case, apply appropriate handling strategies during modeling, such as imputation, feature engineering, or model-based approaches, while validating the impact on model performance.

Pro tip: Always investigate the data generation process first—zeros and missing values often carry business meaning (e.g., zero usage vs. missing data). Document your assumptions and test multiple handling strategies to see which yields the best model performance.

1. Understand the Data Generation Process

Examine how the data was collected and what zeros and missing values represent in the business context. For example, a zero might mean 'no activity' while missing might mean 'data not tracked'.

2. Diagnose Patterns and Causes

Analyze the distribution of zeros and missing values across features, time, and segments. Use visualizations and statistical tests to determine if they are random or systematic.

3. Classify Missingness Mechanisms

Categorize missing values as Missing Completely at Random (MCAR), Missing at Random (MAR), or Missing Not at Random (MNAR). This guides the appropriate handling strategy.

4. Apply Different Handling Strategies

For true zeros, consider leaving as is or transforming; for MCAR/MAR, use imputation (mean, median, model-based); for MNAR, use domain-specific imputation or include missingness indicators.

5. Validate and Iterate

Test the impact of each handling method on model performance using cross-validation. Monitor for bias and adjust as needed.

Key Points to Mention

  • Distinguish between true zeros and missing values—they often have different meanings.
  • Use missingness indicators as features when missingness itself is informative.
  • Consider model-based imputation (e.g., KNN, MICE) for MAR data.
  • Avoid dropping data unless missingness is minimal and MCAR.
  • Evaluate the impact of handling strategies on model metrics and business outcomes.
  • Document assumptions and communicate with stakeholders about data limitations.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q2

If a feature isn't available at inference time but exists in training data, what are your options and how do you choose between dropping it, building a proxy, or reframing when the prediction happens?

Technical Trade-offsStakeholder ManagementProduct Analytics & Metrics
Author's notes

The training-serving skew angle was the part I found genuinely tricky.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business impact and constraints: how critical is the feature, what's the cost of dropping it, and what's the latency/accuracy trade-off. Then systematically evaluate each option (drop, proxy, reframe) against data leakage risks, implementation complexity, and stakeholder needs. Conclude with a recommendation that balances model performance with product feasibility, and suggest validation via offline simulation or A/B test.

Pro tip: Emphasize that you'd quantify the feature's importance (e.g., permutation importance) and simulate the impact of each option before deciding—this shows you're data-driven and mitigate risk. Also, mention that reframing the prediction time might require cross-team alignment, so loop in product and engineering early.

1. Clarify the problem and constraints

Ask questions to understand why the feature is unavailable at inference (e.g., latency, data pipeline, regulatory) and what the business impact of losing it is. Identify success metrics and constraints like latency, cost, and stakeholder expectations.

2. Assess feature importance and leakage risk

Quantify how much the feature contributes to model performance using importance scores or ablation tests. Check if using it in training but not inference causes leakage or training-serving skew.

3. Evaluate options: drop, proxy, or reframe

For each option, estimate impact on model accuracy, implementation effort, and maintainability. Dropping is simplest but may hurt performance; a proxy requires finding a correlated, available feature; reframing (e.g., predicting earlier or later) changes the prediction task and may need product buy-in.

4. Choose based on trade-offs and validate

Select the option that best balances performance, feasibility, and stakeholder needs. Validate offline with a holdout set simulating inference conditions, and if possible, run an online A/B test to measure real impact.

5. Communicate and iterate

Document the decision, rationale, and expected impact for stakeholders. Monitor post-deployment and be ready to iterate if the chosen approach underperforms.

Key Points to Mention

  • Training-serving skew and data leakage: ensure the feature's absence at inference doesn't invalidate the model.
  • Feature importance analysis (e.g., permutation importance, SHAP) to quantify the cost of dropping.
  • Proxy feature selection: find a correlated feature available at inference, but validate its predictive power and stability.
  • Reframing prediction time: consider shifting when prediction happens (e.g., earlier with less data or later with more) and its product implications.
  • Trade-off analysis: compare accuracy loss, engineering effort, latency, and business impact for each option.
  • Stakeholder alignment: involve product, engineering, and business teams to ensure the chosen approach meets needs and is feasible.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q3

The feature is negatively correlated with success. Give at least two plausible explanations including a contrasting interpretation, and describe how you'd validate or rule out each one.

Product Analytics & MetricsRoot Cause AnalysisA/B Testing & Experimentation
Author's notes

More sessions meaning less success sounds backwards at first.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the metric and the correlation context, then propose at least two distinct explanations—one causal (feature harms success) and one non-causal (confounding or reverse causality)—and explicitly contrast them. For each, describe a validation method such as a controlled experiment, stratified analysis, or instrumental variable, and explain how you'd rule it out.

Pro tip: Acknowledge that correlation doesn't imply causation, but also note that negative correlation could be a sign of a valuable feature that attracts users who are already struggling—so investigate user intent and segment-level effects before concluding the feature is harmful.

1. Clarify the metric and correlation

Define what 'success' means (e.g., retention, revenue) and how the feature usage is measured. Confirm the correlation is negative and statistically significant, and consider the time window and population.

2. Propose causal explanation: feature harms success

Hypothesize that using the feature directly causes worse outcomes (e.g., confusing UI, performance issues). Suggest validation via A/B test or randomized controlled trial, or by analyzing feature usage intensity and outcomes.

3. Propose non-causal explanation: confounding or reverse causality

Offer an alternative: a third variable (e.g., user skill) drives both feature use and lower success, or struggling users seek out the feature (reverse causality). Validate by stratifying on confounders, using propensity score matching, or checking temporal order.

4. Contrast interpretations and design validation

Explicitly compare the causal and non-causal explanations, noting that they imply different interventions. Describe how you'd rule out each: e.g., if A/B test shows no effect, causal explanation is weakened; if effect disappears after controlling for confounder, non-causal is supported.

5. Recommend next steps and decision

Summarize which explanation is more likely based on evidence, and propose a path forward: iterate on the feature, target it to the right users, or deprecate it. Emphasize the importance of experimentation and monitoring.

Key Points to Mention

  • Correlation vs. causation and the need for controlled experiments
  • Confounding variables (e.g., user tenure, platform, acquisition channel)
  • Reverse causality (users with low success are more likely to use the feature)
  • Simpson's paradox or segment-level analysis to uncover heterogeneous effects
  • A/B testing or quasi-experimental methods (e.g., difference-in-differences, instrumental variables)
  • Practical implications: feature iteration, targeting, or deprecation based on validation

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.