Accurate forecasting of prices of agricultural commodities has paramount importance, as it enables farmers, policymakers, and the government to make well-informed decisions. While stochastic models like autoregressive integrated moving average (ARIMA) and its components have gained popularity in modelling linear dynamics, they fall short when it comes to capturing the inherent nonlinearity present in the datasets. Machine learning (ML) techniques, such as artificial neural networks (ANNs), have rapidly gained prominence in the field of forecasting, as they are better suited to handle the nonlinearity present in the data. Therefore, a crucial step considered was the preprocessing of the time series data to extract the underlying signal. The present studies focus on the application of the hybrid models and also analyses the prediction accuracy of different models in forecasting monthly prices of maize in Andhra Pradesh, India. The models used for the study were ARIMA, GARCH, ANN, waveletARIMA, waveletGARCH and waveletANN models were employed to compare the accuracy performance of different models in price forecasting. Empirical evidence clearly demonstrates that substantial improvements over conventional techniques were achieved by implementing a wavelet-based combination approach in conjunction with machine learning (ML) techniques. This approach capitalises on the strengths of both wavelet transformations and ML algorithms, resulting in enhanced forecasting performance and accuracy. The ARIMA (2, 1, 1) model was identified as one of the appropriate models to forecast the prices of maize in Andhra Pradesh. The significant p- value indicates the strength of evidence against the null hypothesis. A small p-value in ARCH-LM test reveals the presence of conditional heteroskedasticity. The non-linear time series, NNAR (3-10-1) model was identified as the best model, 51 weights with linear output units. The study revealed that the wavelet based hybrid models forecasted the maize price series better than the individual stochastic models, confirmed based on the performance metrics. Among the proposed models the best performed model identified was waveletANN, with the least MSE and RMSE values.
Swarnalatha et al. (Mon,) studied this question.