Discuss PhD coursework and research impact

## How to structure a strong answer (2–3 minutes) - Open with a one-sentence thesis of your PhD focus and modeling interests. - Coursework: name 2 courses, then for each: what principle you learned and how it changed your modeling behavior. - Research failure: tell a short STAR story (Situation, Task, Action, Result) with quantified impact and validations (ablation, replication, external). - Feedback: one area from advisor, one from collaborator; end with concrete actions you implemented. ## Choosing the two courses (and what to say about them) Pick courses that credibly shape empirical rigor. Examples and what they taught you: - Bayesian Data Analysis: prior elicitation, posterior predictive checks, uncertainty quantification; practical habit of checking calibration and coverage, not just point metrics. - Causal Inference: identification vs prediction, DAGs, diff-in-diff, IV; habit of defining estimands, guarding against confounding, and designing validation consistent with the question. - Statistical Learning / Regularization: bias-variance tradeoff, cross-validation, regularization paths; habit of nested CV, early stopping, and feature group ablations. - Time Series / Panel Methods: leakage avoidance, rolling splits, non-stationarity; habit of out-of-time validation and stability checks across regimes. Template phrasing: - Course name → principle → behavior change. - Example: Causal Inference taught me to separate identification from prediction; now I design features and targets to align with the estimand and validate with out-of-time and subgroup checks. ## Research project where the first approach failed (STAR with quantification) Example you can adapt (use your domain and numbers, but keep the structure): - Situation: I studied whether alternative app-usage signals could predict quarterly earnings surprises. - Task: Build a model to classify surprise vs no surprise and assess economic value. - Action (initial approach and failure): I started with an end-to-end LSTM on raw daily signals. It looked good in random CV (AUC ≈ 0.70) but collapsed in a rolling time split (AUC ≈ 0.53). Feedback highlighted leakage (lookahead in feature windows), weak target definition, and overfitting. - Action (after feedback): - Redefined target to be public at T+0 with features lagged by at least 7 days to eliminate lookahead. - Switched to a transparent baseline (regularized logistic regression and gradient boosting) with engineered weekly aggregates and seasonality controls. - Adopted rolling-window hyperparameter tuning and out-of-time holdouts. - Added domain controls (sector, size, prior momentum) to reduce spurious correlations. - Result (quantified impact): - Predictive: AUC improved from 0.53 to 0.62 on a 4-quarter holdout; Brier score decreased by 11%; calibration slope ≈ 0.97. - Ablations: removing alternative data dropped AUC by 0.06; removing seasonality controls dropped AUC by 0.02, isolating which components mattered. - Replication: similar uplift across 3 sectors and an international sample with ΔAUC ≈ 0.05–0.07. - External validation: model trained on 2017–2020 generalized to 2021–2022 with AUC 0.60; long-short backtest IR improved from 0.30 to 0.78 after transaction costs, turnover constraints, and block-bootstrap CIs. What to emphasize: - The precise failure mode (leakage, misspecified target, overfitting) and the guardrails you added (time-based splits, lagging, calibration checks). - Quantified impact and which validation types you used (ablation to attribute gains, replication across subgroups, external out-of-time sample). ## Advisor and collaborator feedback (area to improve) Pick two complementary angles and actions you took. - Advisor’s likely feedback (depth/rigor): tendency to move to complex models too soon. - Actions: enforce a modeling ladder (dummy/baseline → linear → tree → deep), pre-analysis plans, and a checklist for identification and leakage before any tuning. - Collaborator’s likely feedback (engineering/communication): code modularity and reproducibility; or crisp communication of uncertainty to non-experts. - Actions: unit tests for feature pipelines, seed and data versioning, reproducible environments; one-slide experiment templates with problem, metric, decision rule; practice explaining calibrated probabilities and effect sizes. ## Pitfalls and guardrails to mention explicitly - Avoid leakage: strict temporal splits, lagged features, no peeking across folds. - Prevent p-hacking: predefine metrics and decision thresholds; correct for multiple comparisons if exploring many features. - Robust validation: nested or rolling CV; maintain a final untouched holdout; report calibration and stability across regimes/subgroups. - Reproducibility: fixed seeds, data versioning, and code review. ## Compact sample answer (adapt with your details) My PhD centered on empirical modeling for economic time series. Two courses shaped my approach. Causal Inference taught me to separate identification from prediction; I now define estimands first, design targets and features to avoid confounding, and validate with out-of-time and subgroup checks. Bayesian Data Analysis made uncertainty first-class: I use prior elicitation, posterior predictive checks, and always report calibration, not just accuracy. In one project predicting earnings surprises from app-usage data, my first LSTM looked strong in random CV but failed in a rolling split (AUC ≈ 0.53). Feedback revealed lookahead and a loose target. I lagged all features by 7+ days, clarified the target, moved to a regularized logistic baseline and gradient boosting with weekly aggregates and seasonality controls, and tuned in rolling windows. AUC rose to 0.62 on a 4-quarter holdout, Brier improved 11%, and calibration slope was 0.97. Ablations showed the alternative data contributed +0.06 AUC; replication across sectors and an international sample held gains; an out-of-time 2021–2022 test achieved AUC 0.60, with a transaction-cost-aware backtest IR rising from 0.30 to 0.78. My advisor would say I sometimes reach for complex models too fast. I now use a modeling ladder and pre-analysis plans. A close collaborator would cite reproducibility. I instituted unit tests for feature code, data versioning, and standardized experiment reports to make results easy to audit and re-run.