Beyond the Aggregate: Provenance, Communication and Scope in Machine-Learning FDI Forecasting for Policy

Abstract: Machine-learning forecasts of national-aggregate FDI are increasingly offered to policymakers as decision support. This paper identifies three risks that accompany such forecasts and are rarely stated alongside the accuracy figures, and grounds each in a documented case. The first is provenance. A forecast assembled from central bank, statistical agency and multilateral sources inherits their definitional choices and revision practices, and downstream users usually cannot trace which source supplied which observation. In the case examined here, half the assembled series was linearly interpolated from annual figures and two further construction defects were present, none of which was documented or detectable from the published output. The second is communication. An out-of-sample R-squared says nothing about which quarters a model handles well; the best model in the companion forecasting study achieved 0.0257 and was the only one of twelve to beat a mean benchmark, a context that a headline figure omits. The third is aggregation. National-aggregate FDI in Nigeria is dominated by a small number of large transactions, so a forecast validated at that level supports no inference about sectoral or regional dynamics. We propose three safeguards: a provenance table including construction method, mandatory disclosure of failure modes alongside any accuracy figure, and explicit scope limits restricting aggregate forecasts to aggregate questions.

Keywords: Data Governance; Model Documentation; Forecast Communication; Policy Use; Nigeria.