Hybrid Deep Learning with Denoising Stacked Autoencoders for Agricultural Price Forecasting

Abstract

The complex nature of agricultural price data, characterized by perishability, seasonality, and non-stationarity, often renders conventional forecasting 
models inadequate. To address these challenges, this study proposes two hybrid learning frameworks: Stacked Autoencoder-based Deep Learning 
(SAE-DL) and Denoising Stacked Autoencoder-based Deep Learning (DSAE-DL). These models represent a significant departure from existing 
deep learning (DL) models, such as Multilayer Perceptron (MLP), Recurrent Neural Network (RNN), 1D Convolutional Neural Network (1D CNN), 
Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). Unlike these standard models, which process raw input directly and often 
struggle with data volatility, our proposed hybrid approach integrates an autoencoder bottleneck to perform hierarchical feature extraction and 
dimensionality reduction prior to forecasting. The SAE-DL model focuses on capturing the underlying compressed representation of the price series, 
while the DSAE-DL introduces a stochastic denoising mechanism to specifically reconstruct core features from noisy data and outliers. Empirical 
analysis using a real-world onion price dataset reveals that the proposed DSAE-DL models, particularly the DSAE-RNN configuration, consistently 
outperform all other models across metrics, including RMSE, MAE, and MAPE. The Diebold-Mariano (DM) test further confirms that both proposed 
hybrid frameworks, SAE-DL and DSAE-DL, significantly improve prediction accuracy compared to standalone DL models. These findings highlight 
the superior robustness of hybrid autoencoder-based architectures in contending with market volatility.