← Two Sigma Interview Insights

Two Sigma·Data Scientist·Technical Phone Screen·Senior

SeniorPrefer not to say
Jun 2026

Summary

Two Sigma data scientist interview, heavy on ML design and statistical rigor. The whole session was basically one long open-ended question about building a signal from Google Trends data, with the interviewer pushing deeper at every step. Felt less like a quiz and more like a research discussion where they were actively probing for where your thinking breaks down.

Questions Asked (5)

Q1

You have access to relative search volume data over time plus standard daily stock market data. Design a model end-to-end that uses search data to predict stock prices, covering target framing, feature engineering, model choice, training/validation, and signal evaluation.

Data ModelingTechnical Trade-offsA/B Testing & Experimentation
Author's notes

This is the kind of question where you can sound smart for five minutes and then completely fall apart when they keep asking 'what else.' I started with predicting forward returns instead of price levels, which felt right, but I underestimated how much they'd push on the cross-sectional vs time-series framing.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the prediction target and horizon, then discuss how to engineer features from search data (e.g., lags, changes, sentiment) and combine with market data. Outline a robust validation scheme that respects time series structure and evaluate the signal's economic value beyond statistical accuracy.

Pro tip: Emphasize that search data is often a leading indicator but noisy; use techniques like cross-correlation and Granger causality to identify predictive lags, and always compare against a strong baseline like ARIMA or simple momentum to demonstrate incremental value.

1. Define target and prediction horizon

Clarify whether predicting next-day return, direction, or volatility, and the horizon (intraday, daily, weekly). This determines the modeling approach and evaluation metrics.

2. Engineer features from search and market data

Create features like lagged search volumes, percentage changes, moving averages, and interactions with market variables. Consider dimensionality reduction (PCA) and stationarity transformations.

3. Select and train model with proper validation

Choose models from linear (LASSO) to tree-based (XGBoost) or deep learning (LSTM), and use time-series cross-validation (expanding window) to avoid lookahead bias. Tune hyperparameters.

4. Evaluate signal and backtest

Assess predictive performance (RMSE, IC) and economic value via backtesting with transaction costs. Compare to baseline models and check robustness across subperiods.

5. Iterate and monitor

Discuss potential overfitting, feature importance, and how to monitor model decay. Suggest online learning or periodic retraining.

Key Points to Mention

  • Avoid lookahead bias by using only past data and proper time-series validation
  • Feature engineering: lags, differences, rolling statistics, and interaction terms
  • Model choice: start simple (linear) and increase complexity; consider regularization
  • Evaluation: use both statistical metrics (IC, RMSE) and economic metrics (Sharpe ratio after costs)
  • Baseline comparison: always compare against a naive or market model
  • Data challenges: search data may be noisy, sparse, and subject to regime changes

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

Q2

With thousands of stocks and basically unlimited query terms to choose from, how do you pick which search queries to use without the whole thing becoming data snooping?

Data ModelingA/B Testing & ExperimentationTechnical Trade-offs
Author's notes

Blanked for a second here.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Frame the problem as hypothesis-driven research with strict pre-registration and out-of-sample validation. Emphasize that the goal is to avoid overfitting by limiting the search space based on economic theory, data constraints, and multiple testing corrections. Discuss how you would balance exploration and confirmation to ensure findings are robust and not just noise.

Pro tip: Mention that you would simulate the entire research pipeline on synthetic data to calibrate false discovery rates before touching real data—this shows you understand the multiple comparisons problem deeply and can quantify the risk of data snooping.

1. Define the hypothesis space with economic rationale

Start by restricting queries to those grounded in economic theory or market microstructure, such as value, momentum, or liquidity effects. This reduces the universe of possible queries and prevents aimless data mining.

2. Pre-register the analysis plan

Specify in advance the exact queries, data splits, and evaluation metrics to avoid post-hoc cherry-picking. This includes holding out a final test set that is never used until the very end.

3. Apply multiple testing corrections

Use methods like Bonferroni, Holm, or false discovery rate (FDR) control to adjust p-values for the number of hypotheses tested. This quantifies and controls the likelihood of false discoveries.

4. Validate out-of-sample and across regimes

Test the selected queries on unseen data and across different market regimes (e.g., bull, bear, high volatility) to ensure robustness. Use walk-forward analysis or cross-validation with purging/embargo to avoid leakage.

5. Iterate with a feedback loop but maintain discipline

If results are promising, refine hypotheses based on economic intuition, but always re-validate on new out-of-sample data. Keep a log of all tests to track the cumulative multiple testing burden.

Key Points to Mention

  • Multiple testing problem and corrections (e.g., Bonferroni, FDR)
  • Out-of-sample validation and cross-validation techniques (e.g., walk-forward, purged K-fold)
  • Economic theory as a guide to limit hypothesis space
  • Pre-registration of analysis to prevent p-hacking
  • Data snooping bias and how to detect it (e.g., White's Reality Check, Hansen's SPA test)
  • Trade-off between exploration and confirmation, and the need for a hold-out set

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

Q3

Your backtest shows a strong in-sample Sharpe ratio that collapses out of sample. How do you diagnose what went wrong?

Root Cause AnalysisA/B Testing & ExperimentationTechnical Trade-offs
Author's notes

Classic.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by acknowledging the classic overfitting problem and then systematically walk through potential causes: data leakage, look-ahead bias, regime shifts, and multiple testing. Emphasize the importance of out-of-sample validation and propose concrete diagnostic steps like checking feature stability and performing sensitivity analysis.

Pro tip: At Two Sigma, they value rigorous quantitative thinking—mention that you would compute the deflated Sharpe ratio to account for multiple testing, and always maintain a hold-out set that you never touch until final validation.

1. Verify Data Integrity and Leakage

Check for look-ahead bias, survivorship bias, or data leakage in features or labels. Ensure that all data used in training is strictly point-in-time and that no future information is inadvertently included.

2. Assess Overfitting and Multiple Testing

Evaluate the number of trials and the possibility of overfitting due to multiple hypothesis testing. Use techniques like deflated Sharpe ratio or White's Reality Check to adjust for selection bias.

3. Analyze Regime Shifts and Non-Stationarity

Investigate whether market regimes changed between in-sample and out-of-sample periods. Test for structural breaks and consider whether the strategy relies on relationships that are no longer stable.

4. Perform Sensitivity and Robustness Checks

Conduct sensitivity analysis on hyperparameters, feature importance, and transaction costs. See if small changes drastically alter performance, indicating fragility.

5. Re-evaluate Validation Methodology

Review the cross-validation scheme (e.g., walk-forward vs. k-fold) and ensure it respects temporal order. Consider using purged cross-validation to avoid leakage.

Key Points to Mention

  • Look-ahead bias and data leakage
  • Multiple testing and deflated Sharpe ratio
  • Regime shifts and non-stationarity
  • Cross-validation techniques for time series (walk-forward, purged)
  • Sensitivity analysis and robustness checks
  • Transaction costs and market impact

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

Q4

Search volume often reacts to price moves rather than leading them. How would you actually establish that your feature leads returns rather than just lagging them?

Data ModelingA/B Testing & ExperimentationTechnical Trade-offs
Author's notes

This one tripped me up more than I expected.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Acknowledge the endogeneity problem and propose a causal inference framework that isolates exogenous variation in the feature. Use methods like instrumental variables, difference-in-differences, or randomized experiments to test whether the feature predicts future returns beyond contemporaneous price movements.

Pro tip: Emphasize that in financial data, establishing lead-lag requires careful handling of look-ahead bias and time-series stationarity; consider using Granger causality with proper lag selection and out-of-sample validation to avoid overfitting.

1. Define the causal question and identify confounders

Clarify what 'leads' means: does the feature predict returns at t+1 after controlling for returns at t? Identify potential confounders like market regimes, liquidity, or other contemporaneous signals.

2. Find exogenous variation or design an experiment

If possible, run an A/B test where the feature is randomly assigned. Otherwise, use natural experiments, instrumental variables, or regression discontinuity to isolate exogenous changes in the feature.

3. Model lead-lag relationship with time-series methods

Use Granger causality tests, vector autoregressions (VAR), or lead-lag regressions with appropriate lags and controls. Ensure stationarity and avoid look-ahead bias by using only past information.

4. Validate out-of-sample and assess economic significance

Test the predictive relationship on holdout data and evaluate whether the lead translates into profitable trading after costs. Check robustness across sub-samples and alternative specifications.

5. Address reverse causality and alternative explanations

Rule out that returns drive the feature (e.g., via feedback loops) by testing for reverse Granger causality and using placebo tests or falsification exercises.

Key Points to Mention

  • Endogeneity and reverse causality: price moves can cause changes in search volume, so need exogenous variation.
  • Granger causality and VAR models: test if past values of feature improve prediction of future returns beyond past returns.
  • Randomized experiments or natural experiments: gold standard for causal inference, but often infeasible in finance; quasi-experimental methods as alternatives.
  • Look-ahead bias and stationarity: ensure no future information leaks and handle non-stationary time series properly.
  • Out-of-sample testing and economic significance: predictive power must hold on unseen data and survive transaction costs.
  • Placebo tests and falsification: test for spurious relationships by shuffling or using unrelated features.

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

Q5

If the firm already runs momentum and reversal signals, how would you test whether your search-based signal adds genuine incremental value versus just repackaging what they already have?

A/B Testing & ExperimentationProduct Analytics & MetricsTechnical Trade-offs
Author's notes

Straightforward in theory: regress your signal on the existing factors and check if the residual still predicts returns.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Frame the answer around orthogonalization and incremental predictive power: first residualize your signal against existing momentum and reversal factors, then evaluate whether the residual adds out-of-sample alpha. Emphasize rigorous backtesting with proper controls, transaction cost analysis, and statistical significance to rule out data mining.

Pro tip: Mention that you would test the signal's incremental value in a live paper-trading environment with realistic execution constraints, because backtest overfitting is the biggest risk in signal research. Also, highlight the importance of economic intuition—if the signal's edge disappears after controlling for known factors, it's likely just repackaging.

1. Orthogonalize against existing signals

Regress your search-based signal on momentum and reversal factors (and other known risk factors) to obtain the residual signal. This isolates the unique component not explained by existing strategies.

2. Measure incremental predictive power

Evaluate the residual signal's ability to predict future returns using out-of-sample tests, such as cross-sectional IC, portfolio sorts, and Fama-MacBeth regressions. Compare performance metrics (e.g., Sharpe, t-stat) against a baseline model with only momentum and reversal.

3. Control for multiple testing and data mining

Apply techniques like deflated Sharpe ratio, White's Reality Check, or cross-validation to account for the number of signals tested. Ensure the incremental value is robust to different time periods, universes, and parameter choices.

4. Assess economic significance and implementability

Analyze turnover, transaction costs, capacity, and correlation with existing strategies. Simulate a combined portfolio to see if adding the signal improves net-of-cost performance and diversification.

5. Validate with live or paper trading

Run the signal in a paper-trading or small live allocation to confirm that the incremental value persists in real market conditions, free from backtest biases.

Key Points to Mention

  • Orthogonalization: residualizing your signal against momentum and reversal to isolate unique alpha.
  • Out-of-sample testing and cross-validation to avoid overfitting.
  • Statistical significance: t-stats, Sharpe ratios, and multiple testing corrections (e.g., deflated Sharpe).
  • Economic significance: transaction costs, turnover, capacity, and correlation with existing signals.
  • Incremental value in a combined portfolio: does adding the signal improve risk-adjusted returns?
  • Live validation: paper trading or small live allocation to confirm real-world performance.

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