SSCIF Gets Smarter — Kazakhstan's Central Bank Adds Ridge, Lasso and Elastic-Net Machine Learning to Its Inflation Forecasting System

The National Bank of the Republic of Kazakhstan (NBK) has long run a Selective-Combined Inflation Forecasting System (SSCIF) built on traditional econometric models (OLS, LTAR, BVAR, Random Walk) inside its Monetary Policy Department.
In a December 2024 NBK working paper (NBRK-WP-2024-13), analysts Zarina Adilkhanova and Yerzhan Islam added Ridge Regression, Lasso Regression and Elastic Net machine-learning algorithms to SSCIF and re-ran pseudo out-of-sample accuracy tests, measured by root-mean-squared error (RMSE), across four historical test windows (2016, 2018, 2020-21 and 2023-24).
The updated, ML-augmented model reduced RMSE for overall inflation by 18.6% (January-September 2018), 15.8% (July 2020-March 2021) and 15.3% (May 2023-January 2024) versus the previous econometric-only version, though the authors note the earlier model still forecast some sub-components (non-food and service inflation) more accurately during a March-November 2016 window that had shaped its original design. In August 2024 the IMF's Caucasus, Central Asia and Mongolia Regional Technical Assistance Center (CCAMTAC) ran a scoping mission at the NBK's request to help embed a more robust near-term forecasting system across the wider Monetary Analysis Division, with roll-out planned for the IMF's fiscal year 2026.
The authors explicitly label SSCIF's results as internal, model-only pseudo-out-of-sample estimates that exclude expert judgement and external shocks, and state the paper reflects their personal views rather than the NBK's official position; the CCAMTAC-backed scale-up was, as of the source reporting, still at the planning/roll-out stage rather than a completed institution-wide deployment.

Read the full analysis: https://nationalbank.kz/file/download/106589