Recurrent neural network architecture for forecasting banana prices in Gujarat, India

Abstract

Objectives: The forecasting of horticulture commodity prices, such as bananas, has wide-ranging impacts on farmers, traders and end-users. The considerable volatility in horticultural commodities pricing estimates has allowed farmers to exploit various local marketplaces for profitable sales of their farm produce. Despite the demonstrated efficacy of machine learning models as a suitable substitute for conventional statistical approaches, their application for price forecasting in the context of Indian horticulture remains an area of contention. Past attempts to forecast agricultural commodity prices have relied on a wide variety of statistical models, each of which comes with its own set of limitations.

Methods: Although machine learning models have emerged as formidable alternatives to more conventional statistical methods, there is still reluctance to use them for the purpose of predicting prices in India. In the present investigation, we have analysed and compared the efficacy of a variety of statistical and machine learning models in order to get accurate price forecast. Autoregressive Integrated Moving Average (ARIMA), Seasonal Autoregressive Integrated Moving Average model (SARIMA), Autoregressive Conditional Heteroscedasticity model (ARCH), Generalized Autoregressive Conditional Heteroscedasticity model (GARCH), Artificial Neural Network (ANN) and Recurrent Neural Network (RNN) were fitted to generate reliable predictions of prices of banana in Gujarat, India from January 2009 to December 2019.

Results: Empirical comparisons have been made between the predictive accuracy of different machine learning (ML) models and the typical stochastic model and it is observed that ML approaches, especially RNN, surpassed all other models in the majority of situations. Mean Absolute Percent Error (MAPE), Root Mean Square Error (RMSE), symmetric mean absolute percentage error (SMAPE), mean absolute scaled error (MASE) and mean directional accuracy (MDA) are used to illustrate the superiority of the models and RNN resulted least in terms of all error accuracy measures.

Conclusions: RNN outperforms other models in this study for predicting accurate prices when compared to various statistical and machine learning techniques. The accuracy of other methodologies like ARIMA, SARIMA, ARCH GARCH, and ANN falls short of expectations.

Fig 1. Architecture of Artificial Neural Network (ANN).

The diagram was created using NN-SVG (Lenail, 2015) [43].

Fig 2. Architecture of Recurrent Neural Network (RNN).

The diagram was created using PyTorch and graphviz library [–45].

Fig 3. Time series scattered plot for banana prices in Gujarat.

Fig 4. Time series decomposition of banana prices in Gujarat.

Fig 5. Pair-plot of multiple lag prices of banana in Gujarat.

Fig 6. ACF& PACF lags of time series data for banana price in Gujarat.

Fig 7. Performance of time series models for banana price on test dataset.

Fig 8. Polar chart of accuracy measures on testing data of time series models for banana prices in Gujarat.

Fig 9. Performance of time series models for banana price for forecasted period.

Fig 10. Polar chart of accuracy measures on forecasted data of time series models for banana prices in Gujarat.