Abstract
Does access to transaction-level records affect which forecasting method public-finance analysts should choose? To answer this question, we compare four aggregation strategies and a structural one applied to a short, highly seasonal time series of child-benefit payments: direct-aggregate, bottom-up, data-enrichment, and minimum-trace reconciliation. Our application concerns monthly child-benefit expenditure in Republika Srpska (Bosnia and Herzegovina), where we have approximately a hundred observations of total payments, but more than a million observations of individual payments. We also consider four different ways to deal with the recurring near-zero observations typical of administrative payment data. Under an expanding-window design, where each model can only use information available at the time of each forecast origin, exponential smoothing attains the lowest error across every treatment, though once the overlapping blocks and multiple testing are accounted for, it is statistically indistinguishable from minimum-trace reconciliation and several other benchmarks. Enrichment with microdata-derived features offers no advantage once those features are restricted to information known in advance, and machine-learning methods do not outperform the benchmarks; the treatment of the near-zero observations is as consequential for measured accuracy as the choice of forecasting method. The administrative microdata nonetheless prove valuable in a different way: an entitlement-basis (accrual) series reconstructed from them is far smoother than the cash series, suggesting that their main contribution may lie in defining what is measured rather than in feeding machine-learning models; we examine this accrual view, and its limitations, in an appendix. Overall, on short, strongly seasonal public-finance series, well-established benchmarks remain hard to beat, and additional data changes which object is measured rather than improving the cash forecast itself, unless it carries genuinely forward-looking information.