Accuracy of long short-term memory model in predicting YoY inflation of cities in Indonesia

Abstract

Our  research  evaluates  the  effectiveness  of  the long  short-term  memory (LSTM) model in forecasting annual year-on-year (YoY) inflation across 82 cities in Indonesia based on time series data from BPS economic reports for 2014-2024. This study tests the accuracy of the model in reconstructing past inflation patterns, then evaluates the capabilities and limitations of the model in  various  urban  area  contexts  with  the root  mean  square  error (RMSE), mean  absolute  percentage  error (MAPE),  and coefficient  of  determination(R2)  metrics.  The  findings  show  that  LSTM  performs  well  in  metropolitan areas  such  as  Jakarta,  Bandung,  and  Surabaya  with R2values  >0.8  and  the lowest  MAPE  of  10.91%  in  Jakarta.  However,  in  small  cities  with  higher economic  volatility  such  as  Tanjung  Pandan,  the  model  shows  significant prediction   errors   (R²<0.50   and   MAPE   up   to   283.11%).   Moderate performance  (0.50≤ R²≤0.80)  was  found  in  cities  such  as  Palembang, Semarang, and Makassar, reflecting the model's adaptive ability to moderate inflation  patterns.  These  results  emphasize  the  important  role  of  structured economic data in improving the reliability of predictions, so that the policy implications  of  this  study  include  the  use  of  the  LSTM  model  as  an  early warning system by fiscal and monetary authorities, as well as the need for a data-based  inflation  control  strategy  to  strengthen  regional  and  national economic    resilience    in    supporting    sustainable    development    towards Indonesia Emas 2045.