Forecasting black pepper prices in India: a comparison of time series and machine learning models
Shripad Bhat, T.M. Kiran Kumara, Shiv KumarPurpose
This study aims to evaluate the predictive performance of traditional econometric time series models and modern machine learning (ML) approaches for forecasting black pepper prices in India's Cochin market, a key benchmark market in the Indian spice trade, with the aim of identifying superior forecasting methods to support price risk management in the Indian spice sector.
Design/methodology/approach
Monthly average prices of Malabar Garbled 1 (MG1) grade black pepper from the Cochin market (April 1997–August 2025; n = 341 observations) were used to compare six forecasting models: Autoregressive Integrated Moving Average (ARIMA), Exponential Smoothing (ETS), Multiple Linear Regression (MLR), Random Forest (RF), Extreme Gradient Boosting (XGBoost) and Support Vector Regression (SVR). The dataset was partitioned into a training set (April 1997–August 2023) and an out-of-sample test set (September 2023–August 2025). Forecast accuracy was assessed using root mean square error (RMSE) and mean absolute percentage error (MAPE), complemented by model confidence set (MCS) and Diebold–Mariano (DM) tests.
Findings
The Random Forest model achieved the lowest out-of-sample forecast error (MAPE = 4.82%), representing an approximately 42% reduction relative to the ARIMA benchmark (MAPE = 8.35%). Statistical tests confirm that RF significantly outperforms ARIMA, MLR and SVR, while its performance is statistically comparable to ETS. The MCS procedure excludes ARIMA from the superior model set, highlighting the limitations of purely linear autoregressive structures in capturing the complex dynamics of black pepper price series.
Research limitations/implications
Integration of ensemble-based ML models into decision-support platforms administered by institutions such as the Spices Board of India can enhance procurement planning, inventory management, and price risk mitigation for exporters, processors and smallholder pepper farmers.
Originality/value
This study represents one of the earliest attempts at a long-horizon, rigorous out-of-sample comparison of classical and machine learning forecasting frameworks for Indian black pepper prices, spanning nearly 3 decades of market data, and applies formal statistical tests of predictive superiority to guide model selection in agribusiness decision-making.