Compare SDA and BATS

[hyunjimoon]

Assume:

• $\theta$ are model parameters
• Stock represents System behavior

SDA

$\frac{\partial Stock}{\partial Loop} * \frac{\partial Loop}{\partial \theta}$

BATS

$\frac{\partial \theta{policy}}{\partial Stock}$ based on mapper constructed from behavior classifier $\frac{\partial Stock}{\partial \theta}$
i.e. learn $\psi$ for $P{\psi} = \frac{\partial Stock}{\partial \theta}$ then take the inverse.

From Oliva20_SDA paper,

In addition to formulating the equation to express decomposition weights as a function of the right eigenvectors of the A matrix, we developed graphs that decompose the behavior of a stock variable in terms of the projections of the eigenvectors (see Figure 3) and applied the same logic of assessing elasticities, in this case of the weights on the stock variable (wij ), and the eigenvalues (λi ), to each model parameter. We termed this approach dynamic decomposition weight analysis (DDWA).

Having access to both approaches, we found that although LEEA yielded intuitive endogenous explanations for the observed behavior (based on the loops in the SILS), for policy analysis it was more productive to use DDWA and simultaneously explore the impact of model parameters on the eigenvalues and their projections on specific stocks.

[hyunjimoon]

Classify or measure distance, N+/N-/P+/P- string-based distance.