Abstract
In this study, we proposed a forecasting model for non-linear and non-stationary potato prices by integrating singular spectrum analysis (SSA) with a time delay neural network (TDNN). The price series was initially decomposed into independent components using SSA. Each component was then individually forecasted using TDNN, and the final forecast was obtained by combining the forecasts of all components. The performance of the proposed model is evaluated against several benchmark models, including the autoregressive integrated moving average (ARIMA), TDNN, recurrent SSA (SSA-R), and SSA-based ARIMA (SSA-ARIMA) models, using monthly wholesale price data (₹/quintal) of potato crops from three major Indian markets: Agra, Delhi, and Bengaluru. The proposed model also demonstrated superior performance in capturing directional changes, as evidenced by the D-statistic results. The statistical significance of the models was further assessed using the Diebold-Mariano test, which identified the proposed model as the top performer among those evaluated.
