In this paper, we incorporate diagnostic expectations into the standard small open economy model with an occasionally binding constraint. Diagnostic expectations are a forward-looking model of belief formation characterized by an overreaction to recent news. We find that diagnostic expectations exacerbate crisis dynamics and generate realistic boom-bust credit cycles through the interaction of overreaction to news, neglect of risk, and the borrowing constraint. Optimal macroprudential policy mainly addresses the pecuniary externality and corresponding overborrowing in normal times, while mitigating underborrowing induced by over-pessimism during financial crises. Policymakers often advocate for a macroprudential policy of “leaning against the wind,” while the standard quantitative model typically recommends “fueling the economic boom.” By introducing extrapolative belief distortions, our framework reconciles this disconnect and aligns the quantitative policy prescription with conventional policymaking wisdom.

Using the DSGE model as the data-generating process (DGP), we assess how three key modeling choices influence government spending multiplier estimates: (1) the econometric method-vector autoregressions (VARs) versus local projections (LPs); (2) the identification strategy for government spending shocks—such as recursive, Blanchard–Perotti (BP), or forecast error (FE) methods; and (3) the variable transformation-log versus Gordon–Krenn (GK). Our results demonstrate that even when using the same data set, these choices can lead to substantially different multiplier estimates. Furthermore, we find that the choice of econometric method should align with the shock identification strategy and targeted estimation horizon. For the short-run, LP method produces the most accurate government spending multipliers when the true shock sequence is known. When there is no strong candidate for the shock, BP-type shocks are preferable, with both VAR and LP methods being more suitable for short-run analysis, while VAR models yield more reliable estimates for long-run horizons. Additionally, using the GK transformation instead of the log transformation reduces the upward bias commonly observed in VAR and LP estimates.