Big-Data Machine Learning for Inflation Forecasting — the South African Reserve Bank's Rigorous, Mixed-Results Evaluation

The South African Reserve Bank (SARB) is the national central bank responsible for South Africa's monetary policy and official inflation forecasts. In 2022, five SARB economists (Byron Botha, Rulof Burger, Kevin Kotzé, Neil Rankin and Daan Steenkamp) tested a battery of machine-learning models against the bank's own published inflation forecasts, using large disaggregated consumer-price datasets. The exercise was published as SARB Working Paper WP/22/01 and later peer-reviewed in the journal Empirical Economics.

The study set out to test whether linear (LASSO, Ridge) and nonlinear (XGBoost, neural network) machine-learning models, trained on large disaggregated consumer-price data, could produce more accurate South African headline and core inflation forecasts than SARB's own official published forecasts and a simple random-walk benchmark, across horizons from 1 to 24 months.

Researchers built and trained multiple ML model families on disaggregated CPI data and generated forecasts for headline and core inflation at horizons of 1 to 24 months. Forecast accuracy (RMSE) for each ML model was then benchmarked directly against SARB's own published forecasts and against a naive random-walk model, and differences were tested for statistical significance using Diebold-Mariano tests.

At short horizons SARB's official forecast, which benefits from expert judgement and within-month data updates, was substantially more accurate: for 1-month-ahead core inflation its RMSE was 0.13 versus 0.25–0.27 for the best ML models. The ranking reversed at longer horizons — at 12 months a random-forest model's core-inflation RMSE (0.71) and a neural network's (0.86) were both well below SARB's own forecast (1.19), and at 24 months a neural network (1.09) again beat SARB's forecast (1.32). For headline inflation, XGBoost overtook SARB's own forecast from the 12-month horizon onward (1.28 vs 1.42, and 1.48 vs 1.63 at 24 months).

Despite these longer-horizon gains, the authors' own headline conclusion was candid: Diebold-Mariano tests found that the machine-learning models were 'certainly competitive' but not consistently superior to a simple random-walk model, and traditional benchmarks often outperformed the statistical-learning models. No evidence was found that SARB has adopted any of these models into its actual production inflation forecast; the paper stands as a research evaluation of where machine learning does and does not yet outperform expert-augmented central-bank forecasting.