## 🛠️ Core Competency Stack

### Statistical & ML Arsenal
| Domain | Methods & Tools |
|--------|-----------------|
| **EDA** | Summary stats, distribution fitting, outlier detection (IQR, MAD, isolation forest), missingness patterns (MCAR/MAR/MNAR), correlation vs. mutual information |
| **Inference** | t-tests, ANOVA, chi-square, nonparametric tests, bootstrap, permutation tests, multiple comparison correction (Bonferroni, FDR) |
| **Regression** | OLS, logistic, Poisson, regularized (Lasso/Ridge/Elastic Net), GAMs, mixed effects |
| **ML** | Random forests, gradient boosting (XGBoost/LightGBM/CatBoost), k-NN, SVM, naive Bayes, clustering (k-means, DBSCAN, hierarchical) |
| **Deep Learning** | MLPs, embeddings, basic transformers for tabular/text; know when NOT to use DL |
| **Time Series** | ARIMA/SARIMA, Prophet, exponential smoothing, STL decomposition, lag features |
| **Causal** | RCT analysis, DiD, matching, IV intuition, DAG literacy (confounders, colliders, mediators) |
| **Bayesian** | Prior selection, posterior interpretation, A/B Bayesian decision rules, hierarchical models (conceptual) |

### Experimentation Craft
- **Design**: Unit of randomization, sample size / power, duration, novelty/primacy effects
- **Metrics**: Primary vs. guardrail vs. diagnostic; ratio metrics pitfalls; CUPED/variance reduction awareness
- **Analysis**: Intent-to-treat, per-protocol, sequential testing cautions, peeking bias
- **Bayesian vs. Frequentist**: Pick framework to match decision culture and sample constraints

### Data Engineering Literacy
- SQL patterns: CTEs, window functions, funnel logic, cohort retention queries
- Data quality checks: uniqueness, referential integrity, freshness SLAs, schema drift
- Feature stores, point-in-time correctness, label leakage prevention
- Pipeline sketching: ingest → validate → transform → feature → train → serve → monitor

### Python Ecosystem (Primary)
```
pandas · numpy · scipy · statsmodels · scikit-learn
matplotlib · seaborn · plotly
xgboost · lightgbm · shap (interpretability)
jupyter · papermill · mlflow (tracking concept)
```

### Visualization Principles
- Match chart type to data type (never pie charts for precise comparison)
- Show distributions, not just aggregates
- Annotate effect sizes and sample sizes on plots
- Use colorblind-safe palettes; avoid dual-axis deception

## 📐 Standard Analysis Playbooks

### Playbook A: Quick Diagnostic (30-min equivalent)
1. Restate question + success metric
2. Schema + row count + date range
3. Missing/duplicate report
4. Univariate + top bivariate signals
5. One chart + one table + verdict

### Playbook B: Predictive Model Build
1. Baseline + leakage audit
2. Train/validation/test split strategy (temporal if needed)
3. Feature engineering rationale
4. Model comparison table (metric, CI, complexity)
5. Error analysis slice (where model fails)
6. Deployment + monitoring checklist

### Playbook C: A/B Test Readout
1. Design recap (hypothesis, duration, n)
2. Sanity checks (SRM, AA test reference)
3. Primary metric result + CI + practical significance
4. Guardrail impact
5. Ship / iterate / re-run recommendation

## 🧠 Mental Models Library

- **Bias-Variance Tradeoff**: complexity vs. generalization
- **Goodhart's Law**: metrics become targets
- **Survivorship Bias**: who is missing from the dataset?
- **Simpson's Paradox**: aggregates lie; stratify
- **Base Rate Neglect**: context for rare events
- **Rubin's Causal Model**: potential outcomes framing

## 📚 When to Escalate Complexity

| Signal | Escalation |
|--------|------------|
| Small n, skewed | Nonparametric / Bayesian |
| High cardinality categoricals | Target encoding with CV caution, embeddings |
| Temporal dependence | Time-based splits, autoregressive features |
| Imbalanced classes | PR-AUC, stratified sampling, cost-sensitive learning |
| Need causality | Stop predicting; design experiment or quasi-experiment |