Revenue forecasting underpins fiscal planning and macroeconomic policy, yet no consensus exists on which modelling paradigm, data source, or evaluation protocol delivers the most reliable projections. This paper provides a systematic review of the revenue-forecasting literature from 1970 to 2024, synthesising 312 peer-reviewed studies and 48 institutional reports across national, sub-national, and sector-specific jurisdictions.

Six families of approaches:

We taxonomise approaches into six families:

Univariate time-series

Multivariate macro-econometric

Micro-simulation and tax-buoyancy models

Machine-learning and deep-learning systems

Judgmental and hybrid ensembles

Now-casting and real-time frameworks leveraging high-frequency data

For each family, we summarise theoretical foundations, typical covariates, estimation techniques, and reported forecast-error metrics.

Meta-analysis findings:

Meta-analysis of 1,174 reported forecast horizons shows that hybrid methods combining macro-econometric structure with machine-learning adjustments reduce mean absolute percentage error (MAPE) by 15–30% relative to pure time-series benchmarks at horizons up to two years. Conversely, gains dissipate beyond three years unless judgmental overrides are incorporated.

Emerging practices:

We highlight emerging practices—vector-based tax-gap modelling, transformer architectures with fiscal attention layers, and narrative-augmented now-casting—that exploit novel data such as satellite imagery and web-scraped prices.

Persistent gaps:

The review also identifies persistent gaps: limited out-of-sample testing, inconsistent uncertainty quantification, and scarce cross-country transferability studies

Conclusion and proposal:

We close by proposing a reproducibility checklist and an open repository schema to accelerate comparative research and policy uptake.