Forecasting the Future: Neural Networks vs. Time Series in Pakistan’s CPI

In this paper, diverse models used in the prediction and projection of the consumer price index in Pakistan are reviewed. The data used for the analysis include monthly data from July 2016 to February 2024. We employed both basic time series models like; ARIMA and SARIMA, as well as advanced models; Neural Network Autoregressive (NNAR) and the Multilayer Perceptron (MLP) models. We used differencing as was justified by the Augmented Dickey-Fuller test to handle non-stationary features of the Consumer Price Index series. The correctness of the model attracted the use of performance measures including Mean Correct Squared and Root Mean Correct Squared Error. The SARIMAX (4,1,4)(0,1,[1],12) was found to outperform other more conventional models considerably. The MLP model yielded the lowest RMSE and MSE therefore it is said to have a higher prediction accuracy. Forecasting showed the potential to increase the CPI, an aspect of inflation. In light of these realizations, it becomes clear that only fiscally positive economic initiatives should be used as inflation checkers plus control mechanisms to foster economic stability. In this study, the authors evaluated the power of machine learning models for CPI projection and found their results invaluable for policymakers. Next studies should consider other macroeconomic variables and examine the mixed models to enhance the expected performance of the CPI.