What a Practical Buffer Can and Cannot Promise
The practical question raised by the Federal Reserve paper is what to do when the theoretically best capital rule requires information a regulator cannot reliably possess. This chapter compares the paper's no-failure benchmark, optimized and baseline static buffers, simple indicator rules, and risk-weighted thought experiment on four criteria: preventing failures, preserving deposit liquidity, supporting welfare, and remaining implementable. The result is not a universal winner. It is a conditional judgment about when a modest buffer is a robust fallback and when more precise targeting could justify its information burden.
The distinction matters because the no-failure rule is easy to admire. It sets capital exactly high enough to keep banks from switching into socially inefficient risk. In the model, that is the right safety boundary. But a policy can be optimal for a regulator who sees every relevant shock and unusable for a regulator who must act with delayed, noisy, or incomplete information. Practical policy begins at that gap.
The benchmark is precise because it knows too much
The no-failure rule is an information-rich benchmark. At every date, it sets capital at the lowest level that prevents the bank from entering the risky regime given the current constellation of shocks and the model's state. It does not ask for a permanently high capital ratio. It asks for the threshold that keeps safe lending privately attractive at that particular moment.
That precision is valuable for two reasons. First, it avoids the discrete losses associated with a switch into risky lending. Risky firms have a lower average productivity in the model, risky banks can fail, equity can be destroyed, and deposit insurance losses are passed through to households as taxes. Second, it avoids charging the household for capital above the threshold. Once safe lending is secured, extra equity mainly displaces deposits and their liquidity services.
The paper tests whether this threshold is merely a local trick. It optimizes a linear rule that can respond to every shock process, the current innovations to those processes, and the lagged capital requirement. That rule becomes nearly identical to the no-failure benchmark. Its static buffer falls to about 1 basis point, its average bank-failure rate falls to zero, and its consumption-equivalent variation is below one-tenth of a basis point. When the rule is also allowed to use all the model's state variables, the correlation between its capital requirements and the no-failure rule is 1 in the reported comparison.
Within the class of constant-coefficient linear rules tested by the authors, this is strong evidence that the no-failure construction captures the best available response in the model. It does not establish unrestricted global optimality. Nonlinear rules with coefficients that change when banks fail are outside the paper's scope, and the result depends on the model's shock processes and welfare criterion.
More importantly, the benchmark is not implementable as a direct formula. It requires the regulator to know the shocks affecting the economy, including innovations that may not be observable in real time, and to know how those shocks alter the relative return on many possible forms of lending. Even in a model with only a safe and a risky project, that is a demanding informational requirement. In an actual banking system, the relevant categories are more numerous and the distinction between a temporary decline in safe returns and a lasting deterioration in credit quality is not cleanly visible.
The benchmark should therefore be used as a yardstick, not as a promise. It tells us how much welfare is available if the regulator can locate the threshold accurately.