A Comparative Study of Neural Network and Wavelet Decomposition Models for Price Forecasting of Tomato

Abstract

Tomato price forecasting in Bihar marketplaces is the subject of a study. The study compares and contrasts many time series models. Three models are compared: WT-TDNN (Wavelet Transform with TDNN), TDNN (Time-Delay Neural Network), and ARIMA. This study evaluates the efficacy of advanced and traditional time series models for forecasting tomato prices in Bihar markets, with a focus on ARIMA, Time Delay Neural Networks (TDNN), and Wavelet-TDNN (WT-TDNN). ARIMA models were optimized using the Akaike Information Criterion (AIC), identifying (0, 1, 2) and (1, 1, 1) as optimal configurations, and validated through Wald's test. TDNN models with 2:4s:1l and 1:6s:1l architectures exhibited robust performance based on RMSE, MAE, and MAPE metrics. The WT-TDNN model, which integrates wavelet decomposition with TDNN, demonstrated superior predictive accuracy over both ARIMA and standard TDNN models. By employing the Haar filter for series decomposition into orthogonal components and applying TDNN to each component, WT-TDNN effectively captured the complex, non-linear dynamics of agricultural price series. This innovative approach enhances the precision of price forecasting, offering significant implications for agricultural economics and machine learning applications. The findings provide valuable insights for policymakers, farmers, and market analysts, enabling improved decision-making and fostering resilience in the agricultural sector. The study highlights WT-TDNN as a transformative tool for addressing the challenges of agricultural price prediction.