Forecasting Central-Bank Analytical Agendas: Exploratory evidence from 372 boxes in Banca d'Italia's Annual Reports, 2014–2025

Banca d'Italia's Annual Report has a stable chapter structure, but its analytical boxes (riquadri) provide a flexible editorial space for new evidence and emerging risks. This paper builds a dataset of 372 boxes published in twelve reports (data years 2014–2025) and classifies their titles by theme and transmission channel with a locally run instruction-tuned language model (qwen2.5:7b-instruct, temperature 0). It then evaluates a double-filter transmission hypothesis: international macrofinancial topics are first developed by agenda-setting institutions and subsequently normalized by bodies such as the FSB, ECB, BIS and ESRB before entering a national central bank's analytical agenda. Annual signal matrices, manually coded from a 107-document institutional watchlist, are compared retrospectively with the themes observed in 2021–2025. Mean F1 is 0.60 (annual range 0.47–0.76), compared with 0.54 for a keyword-based classification baseline. The result is exploratory rather than causal or genuinely out of sample: the evaluation covers five years, signal coding involves judgement, and no human-labelled gold standard is available for the LLM taxonomy. The channel classification nevertheless gives an important scope condition: 78% of boxes are domestic and 2% institutional, leaving 20% in the international channel addressed by the hypothesis. The paper closes with explicitly uncalibrated, pre-verifiable scenarios for the report due in 2027, led by private credit/NBFI, financial fragmentation, tokenisation, AI-related financial risk and cyber resilience.