Replication package for: Public Money and Private Investment in Unlisted Firms: Within-Firm Associations, Recognition Timing and Sensitivity

This deposit contains the replication package for the article “Public Money and Private Investment in Unlisted Firms: Within-Firm Associations, Recognition Timing and Sensitivity” by Angelo Leogrande, Mauro Di Molfetta, Valeria Notarnicola, Maria Giovanna Trotta, Antonio Volpe Plantamura and Gianluca Zaza, submitted to the International Journal of Financial Studies (manuscript ijfs-4646066). The article studies how operating contributions recorded in the statutory accounts of Italian firms co-move with gross investment within firms. It compares that co-movement with the co-movement of investment and the firm’s own cash flow net of those contributions. The data are 62,506 firm-year observations on 9,492 unlisted firms over 2015–2024: innovative start-ups, innovative SMEs and ordinary SMEs. All estimates are conditional associations, not causal effects. The package contains three items. The first is the Python code that builds the estimation sample from AIDA (Bureau van Dijk) exports and reproduces every table and figure of the article and its Supplementary Materials: panel estimators, timing and distributed-lag regressions, sensitivity analysis, scaling and equivalence tests, instrumental-variable attempts, clustering and flexible learners. The second is the AIDA search strategies used for the exports. The third is the list of tax codes of the firms in the estimation sample, with their population. A README file gives the order of execution, the order in which netting, scaling and winsorisation are applied, the cross-validation protocol, and the script that produces each table. The AIDA data are licensed by Bureau van Dijk and are not included. Users with an AIDA licence can re-export them using the search strategies and the list of tax codes, together with the Business Register files of the special sections for innovative start-ups and innovative SMEs, dated 27 July 2026. The code runs in Python 3 with pandas, numpy, scipy, linearmodels, scikit-learn and matplotlib.