India plays a significant role in global agriculture production, including spices. Spices play an important role
in cultural significance and economic trade relations
while providing nutritional and medicinal benefits.
The volatility and complexity of the price of spices require improved forecasting methods to support informed decision-making in agricultural markets. Recently,
researchers have focused on using the traditional time
series model as well as machine learning (ML) model
to forecast the price of agricultural commodities. Using
a standalone model struggles to capture the complex
pattern in time series data. To overcome this challenge,
ensemble machine learning approaches based on fixed
weight (FW-ensemble) have been proposed. The ML
models like artificial neural networks (ANN), random
forest (RF), k-nearest neighbours (kNN), extreme gradient boosting (XGBoost), support vector regression
(SVR) and the stochastic model, e.g. autoregressive integrated moving average (ARIMA) model have been
used. The outputs of these models are ensembled using
optimised fixed weights. In this study, the prices of two
important spices, namely turmeric and coriander, from
2010 to 2024, collected from AGMARKNET (https://
agmarknet.gov.in/) were considered. The MCS algorithm was used to select the better-performing model.
The empirical performance of the ensemble method
was compared with that of the stochastic model
(ARIMA), ML techniques (ANN, RF, kNN, XGBoost,
SVR) and deep learning techniques, e.g. long shortterm memory (LSTM) and gated recurrent unit
(GRU), based on several accuracy measures. It revealed that the FW-ensemble approach significantly outperformed the other candidate models in terms of
prediction accuracy.