The Bank of Italy’s ICAS rating process has two stages that combine the statistical model with an expert assessment, performed by two analysts and the rating committee, to obtain the final rating for the firm. Every month, the statistical model produces the probability of default (PD) over a one-year horizon for 370,000 non-financial firms, using a fully automated procedure. This paper illustrates the methodology underlying the Bank of Italy’s ICAS statistical model and its validation process. The model preserves simplicity and ‘readability’ by relying on a logit regression, while it tries to improve predictive performance with machine learning components for some variables that display non-linear behaviour towards default prediction. The model shows robust properties, as it discriminates between healthy and risky firms with fairly stable results. The discriminatory power is rather high and it improves as the size of the company increases, thus ensuring a proper evaluation of the largest exposures in monetary policy operations.