Nowcasting Domestic Liquidity with Machine Learning — Bangko Sentral ng Pilipinas' Regularization and Tree-Based Models

Bangko Sentral ng Pilipinas (BSP), the Philippine central bank, published BSP Working Paper Series No. 2022-04, 'Nowcasting Domestic Liquidity in the Philippines Using Machine Learning Algorithms', by Juan Rufino M. Reyes of the Department of Economic Statistics, in March 2022. Domestic liquidity (M3) growth is a leading indicator BSP uses to help set monetary policy, but like many official statistics it is published with a lag and subject to revision.

To test whether machine-learning nowcasting methods could produce more accurate, timelier estimates of domestic liquidity growth than BSP's traditional time-series benchmarks.

The study applied Ridge Regression, LASSO, Elastic Net, Random Forest and Gradient Boosted Trees to high-frequency monetary, financial and external-sector indicators from January 2008 to December 2020, evaluating one-step-ahead nowcasts under an expanding-window process against ARIMA, random-walk and dynamic-factor-model benchmarks.

The regularisation methods performed best: Elastic Net produced the lowest overall error of the models tested, with a root-mean-square error of 0.529 and mean absolute error of 0.391, and the machine-learning models as a group produced consistently lower monthly forecast errors than the traditional benchmarks, including during the months of early 2020 when liquidity suddenly expanded because of COVID-19-related national-government borrowing.

BSP's own conclusion is modest and specific: the paper recommends the models as a 'supplementary tool' to the Bank's existing suite of forecasting models for GDP, inflation and policy simulation, rather than claiming they replace it. The paper is mirrored by the Philippine Institute for Development Studies, alongside companion BSP papers applying similar methods to regional inflation and disaggregated CPI nowcasting - indicating a sustained, multi-year internal research programme rather than a one-off exercise.