- calibration/parameter estimation - nonlinear regression (time delay, nonlinearities, accumulations, missing data, autocorrelations, hetereoskedasticity) -> method of least squares, method of simulated moments - e.g. VCR (three series of sales, price, adopters) which has installed base (includes learning curve), SIR model, price - how sensitive is the payoff function for calibration to changes in parameters - ![[Pasted image 20221205150416.png]] - use smallest calibration probems possible (e.g. learning curve = cumulative production) ## 1. out of sample prediction (e.g. covid) 2. modeling human growth and weigh dynamics (not used in the model) e.g. 3. synthetic data (add process and measurement noise to the model) - covid: acuity (estimated on synthetic data) 1. have ground truth 2. p% coverage to p% coverage - ![[Pasted image 20221205150746.png]] Accumulated structural bias: endogeneity, processs, measurement noise account for noise coming in - define policy function; u = P(s,r) - better save (high level of capability, ) - turn on and off is better; don't continue production (invest in bringing crews; produce) - SD가 blackbox model보다 10~20 배 of data ~ estimated parameter: calibration, surrogate,