Machine learning has moved from niche to standard in quantitative finance. Firms use it for signal research, execution optimization, alternative data, and risk. But the ML you are tested on is not the ML of Kaggle leaderboards — it is a smaller, sharper toolkit, examined for failure modes as much as for accuracy.
The core checklist
You should be able to discuss each of these fluently, with examples from market data:
- Bias-variance tradeoff. Be ready to connect it to overfitting in financial signals — where the noise-to-signal ratio is brutal and the effective sample size is smaller than the row count suggests.
- Regularization. L1 versus L2, what sparsity buys you in a feature-rich signal setting, and when you would reach for elastic net.
- Cross-validation, honestly done. Purged and embargoed splits for time series — why random k-fold leaks information in financial data and what to do instead.
- Tree ensembles versus linear models. Gradient boosting remains the workhorse for tabular financial data; be able to say why, and when a linear model would win.
- Neural nets, proportionately. Where they help (sequence models, NLP on filings and news) and where they underperform trees on limited, noisy data.
The questions that separate candidates
"How would you backtest a model on price data?"
This is the question that exposes whether someone has actually done the work. Strong answers mention: point-in-time correctness (no using restated or revised data), walk-forward or purged splits rather than random folds, transaction costs and slippage in the evaluation, and sanity checks against trivial baselines. Weak answers describe a single train-test split and a high Sharpe ratio.
"Your model has 0.8 AUC on validation. Is it good?"
The right instinct is suspicion. A 0.6 AUC daily-return classifier in equities would be extraordinary; 0.8 usually means leakage. Common leaks the interviewers are probing for: lookahead in feature construction, survivorship bias in the universe, and target contamination through scaling before splitting.
"Class imbalance — 99% of days the market does nothing. Now what?"
Reframe the objective. Discuss precision/recall tradeoffs for the actual decision, threshold tuning against a cost function, and whether resampling helps or just distorts calibration. Mention that metrics should map to the trading decision, not to an abstract accuracy number.
Financial-flavored ML questions
- How does non-stationarity break standard ML assumptions, and what would you do about it — rolling retraining, regime features, distributional checks?
- How would you build features from earnings calls, filings, or news flow? What are the pitfalls of NLP on financial text?
- Reinforcement learning shows up for execution and routing — be able to explain at a conceptual level why deterministic control sometimes beats RL in production.
How to prepare
- Rebuild the classics from scratch once: linear and logistic regression, a small gradient boosting model, and one cross-validated pipeline with leakage checks. Interviewers can tell who has done this.
- Read one post-mortem about a backtest that lied. Internalizing failure modes is worth more than memorizing algorithms.
- Practice describing tradeoffs out loud. The rubric rewards structured thinking: state assumptions, propose an approach, name the failure modes, iterate.
Firms are not looking for ML researchers in these interviews. They are looking for someone who knows which tool reaches for which problem, and — more importantly — every way that tool can silently fail on market data.