Volatility modelling is an important area of time series analysis. Agricultural commodity prices tend to be more volatile due to seasonality, inelastic demand, production uncertainty and also because many agricultural commodities are perishable. An increase in price volatility implies higher uncertainty about future prices, a fact that can affect producer’s welfare. Therefore, understanding the nature of agricultural commodity price volatility and the ability to accurately forecast the price volatility are important concerns among both policy makers and farming community. Price volatility provides a measure of the possible variation or movement in the price variable. Wide price movements over a short period of time are considered as high volatility. The importance of volatility has led to the development and applications of many significant time series models. In this regard, GARCH (Generalized Autoregressive Conditional Heteroscedasticity) model after its introduction has been widely used owning to its applicability in variety of domains. In recent years, the machine learning technique is introduced for developing learning models and is used to forecast time series with deep learning algorithms. Deep learning algorithms such as Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN) have performed satisfactorily in many disciplines of science, especially in the financial sectors. Deep learning mechanism is used to visualize the pattern and structure of the series, mainly complexity and non-linearity by extracting hidden layers from the target network in time series forecasting. Thus, this study focuses on using these well-established deep learning models to efficiently model and forecast volatile price series of edible oils in India namely rapeseed and mustard, soyabean and sunflower. The datasets were collected from April 1982 to March 2025 from https://eaindustry.nic.in/. As it is a well-known fact that efficiency of these models largely depends upon optimal hyperparameter settings along with appropriate input features. In our case the input features refer to the ideal number of lags. Hence, we approach this problem in two phase manner, first by selecting ideal number of lags using random forest algorithm and then optimizing the hyperparameters of the LSTM-CNN model using Genetic Algorithm (GA). We then implement LSTM, CNN, LSTM-CNN and optimized LSTM-CNN to our real datasets. We document superior results for the proposed optimized LSTM-CNN model. The gain in efficiency ranged between 5-10 (%) in case of RMSE and 7-13 (%) for MAPE, when compared to conventional LSTM-CNN model. This study highlights the importance of hyperparameter optimization and ideal lag selection for improving the efficiency of deep learning models. This study not only provides an effective methodology for forecasting volatile agricultural prices but also presents a framework that can be generalized to other similar time series datasets. Moreover, the findings contribute toward achieving Sustainable Development Goal (SDG) 2.c, which aims to limit extreme food price volatility, thereby aiding policymakers in developing informed and resilient agricultural policies.
