Enhancing Predictions for Indian Coffee Exports with Hybrid Statistical and LSTM Models

Abstract

Accurate forecasting of Indian coffee exports is essential for optimizing supply chains, stabilizing prices, and mitigating market risks. Traditional statistical models like ARIMA and SARIMA struggle with capturing nonlinear dependencies and demand fluctuations, necessitating an advanced hybrid approach. This study proposes a Hybrid ARIMA-SARIMA-LSTM Forecasting Model (HASL-FM) that integrates statistical methods with deep learning for improved predictive accuracy. The objective is to enhance export forecasting precision by leveraging historical export data, macroeconomic indicators, and climate variables. The model employs ARIMA for linear trends, SARIMA for seasonal variations, and LSTM to capture complex, long-term dependencies. The proposed method is implemented in Python on an edge server (16–20 cores, 10TB–15TB storage) and evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and R² score. The dataset consists of Indian coffee export records, macroeconomic factors, and climate data. Experimental results demonstrate that HASL-FM significantly outperforms traditional models (ARIMA, SARIMA) and machine learning models (SVM, ANFIS), achieving the lowest RMSE (0.0256) and MAPE (0.0754), with the highest R² (0.9852). These findings confirm that AI-driven forecasting enhances accuracy, supporting producers, exporters, and policymakers in decision-making, resource allocation, and risk management. Unlike traditional models, HASL-FM effectively captures seasonal trends and nonlinear dependencies, leading to more stable trade and optimized market strategies. Future research could integrate global trade policies, sentiment analysis, and reinforcement learning to further enhance predictive capabilities