As a key indicator of macroeconomic performance, inflation trends significantly influence monetary policy, macroeconomic regulation, and financial market stability. However, macroeconomic time series often contain both linear trends and complex nonlinear fluctuations, which limit the accuracy and stability of traditional statistical models. To address this, the paper proposes an improved ARIMA-LSTM hybrid forecasting model for inflation rate prediction. The ARIMA component extracts the linear structure of the series, while the residual sequence captures unexplained nonlinear information. A multi-scale LSTM network then learns deep features from the residuals, and a dynamic weight fusion mechanism adaptively combines linear and nonlinear predictions. Experiments using CPI data from IMF, World Bank, and FRED databases show that the proposed model achieves an RMSE of 0.564 on the test set—12.1% lower than the traditional ARIMA-LSTM and 17.1% lower than ARIMA alone. It also outperforms models such as SVR, random forest, LSTM, and GRU in MAE and MAPE. In multi-step forecasting, error growth remains around 12% over six steps, notably lower than comparison models. Ablation studies and Diebold–Mariano tests confirm the effectiveness of the multi-scale module and dynamic fusion mechanism. Overall, the improved ARIMA-LSTM model enhances inflation prediction accuracy and stability, offering practical value for macroeconomic forecasting and policy analysis.