The current expected credit losses (CECL) model stipulates that loan loss provisions should be forward-looking. I document that banks increasingly rely on macroeconomic forecasts following the implementation of CECL, yet most forecasts are not rational. I build a theoretical model to study the implications of CECL for bank provisions, lending, aggregate output in the economy, and bank stability. Additionally, I derive the optimal minimum capital requirement within the CECL framework. I first explore the implications under rational expectations and then consider bank expectations influenced by Kahneman and Tversky's (1972) representativeness heuristic to capture the empirical properties of macroeconomic forecasts. My model demonstrates that the representativeness heuristic results in overreaction to news, leading to underprovisioning, excessive lending and risk-taking in reaction to good news, and the opposite in reaction to bad news. Due to timely loan loss provisioning under CECL, the optimal capital constraint is time-varying in response to the macroeconomic conditions and the underlying risk in the economy. In contrast to rational expectations, the representativeness heuristic necessitates a binding capital constraint even when bank equity is high and the social cost of bank failure is low.