New Trends in Estimation and Inference
summary
Jupyter notebooks fit Weibull and piecewise-hazard models with scipy.optimize.minimize and automatic gradients from Autograd to avoid hand derivatives, compute Hessians for parameter variances, and apply the delta method to get CIs for business metrics. A piecewise churn model with nine parameters converges and yields an estimated LTV of $51 at 120 days with wider CIs after truncating data. Log-transforming parameters removes bounds and penalized variance acts like Bayesian pooling.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:22Tutorial overview
- 4:12Fitting parametric models
- 13:22Survival analysis
- 21:56AutoGrad differentiation
- 34:41Generalizing models
- 39:24Parameter variance estimation
- 44:07Models to business outcomes
- 48:55Lifetime value modeling
- 58:23Optimizing unbounded parameters
- 1:01:59Statistical inference trends
- 1:16:23Q&A
related talks
01h 24mBayesianism and Survival Analysis
35mEasy Bayesian regularization for fitting financial time series and curves
01h 39mApplying Rigorous Machine Learning Methods in Business Strategy
17mCodifying Data Science Intuition
34mPrediction intervals
01h 23mPredictive survival analysis with scikit-learn, scikit-survival and lifelines
44mHandling predictive uncertainty in Machine Learning
39mPractical Optimization for Stats Nerds