Hybrid Decomposition-Based Deep Learning Models for Agricultural Commodity Price Forecasting: Wavelet and EMD Approaches for Indian Cardamom Market
1 / 2
Authors
-
Martin Joseph
Sher-e-Kashmir University of Agricultural Sciences and Technology-K, Srinagar, J&K, India
-
Imran Khan
Sher-e-Kashmir University of Agricultural Sciences and Technology-K, Srinagar, J&K, India
Keywords:
Cardamom price forecasting; Wavelet transform; Empirical mode decomposition; LSTM; GRU; hybrid models; agricultural commodity markets; time-series forecasting; deep learning; Kerala spice marketsAbstract
Accurate short-term price forecasting for high-value agricultural commodities is critical for farmer income stability, trader planning, and policy
formulation. Cardamom (Elettaria cardamomum), India’s premier spice by export value, exhibits strongly nonlinear, nonstationary, and seasonally
complex price dynamics that challenge conventional econometric models. This study develops and evaluates a comprehensive family of hybrid
decomposition-based deep learning models combining Discrete Wavelet Transform (DWT) and Empirical Mode Decomposition (EMD) with Long
Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks for weekly cardamom price forecasting across three major Kerala markets:
Kattapana, Nedumkandam, and Thodupuzha. These four hybrid architectures are benchmarked against standalone LSTM, GRU, ARIMA, and
SARIMA models across five evaluation metrics: RMSE, MAE, MAPE, R-squared, and Directional Accuracy (DA)- the percentage of weeks in which
the model correctly predicted the direction of week-on-week price change. Results demonstrate that wavelet-based hybrids outperform all competing
models across all markets. In Kattapana, the Wavelet-GRU model achieved the strongest out-of-sample forecast accuracy (RMSE = 35.82, MAPE
= 2%), together with the highest goodness-of-fit (R-squared (0.98) and directional accuracy (DA = 86.49%) recorded among all sixteen models
evaluated, while Wavelet-LSTM and Wavelet-GRU attained MAPE of 5–6% in Nedumkandam and Thodupuzha. EMD-based hybrids ranked second
overall, with EMD-GRU achieving RMSE = 94.83 and MAPE = 4% in Kattapana, confirming the benefit of adaptive signal decomposition ahead
of sequence learning; however, this advantage was market-dependent and did not hold in Nedumkandam, where EMD hybrids underperformed the
standalone LSTM/GRU baselines. The study provides a systematic, multi-market evaluation of hybrid decomposition-neural architectures for spice
price forecasting-adomain underserved in the existing literature and offers actionable guidance for farmers, traders, and agricultural market regulators
Downloads
References
1. Awajan, A.M., Ismail, M.T. and Al Wadi, S. (2019). A review on
2. empirical mode decomposition in forecasting time series. Italian
3. Journal of Pure and Applied Mathematics, 42, 301–323.
4. Box, G. and Jenkins, G. (1970). Time Series Analysis: Forecasting and
5. Control. Holden-Day, San Francisco.
6. Branco, N.W., Cavalca, M.S.M., Stefenon, S.F. and Leithardt, V.R.Q.
7. (2022). Wavelet LSTM for fault forecasting in electrical power
8. grids. MDPI Sensors, 22, 8323.
9. Cho, K., van Merrienboer, B., Gulcehre, C., Bahdanau, D., Bougares,
10. F.,
11. Schwenk, H. and Bengio, Y. (2014). Learning phrase
12. representations using RNN encoder-decoder for statistical
13. machine translation. arXiv:1406.1078.
14. Hochreiter, S. and Schmidhuber, J. (1997). Long short-term memory.
15. Neural Computation, 9(8), 1735–1780.
16. Huang, N.E., Shen, Z., Long, S.R., Wu, M.C., Shih, H.H., Zheng, Q.,
17. Yen, N., Tung, C.C. and Liu, H.H. (1998). The empirical mode
18. decomposition and the Hilbert spectrum for nonlinear and non
19. stationary time series analysis. Proceedings of the Royal Society
20. of London A, 454, 903–995.
21. Kriechbaumer, T., Angus, A., Parsons, D.J. and Rivas Casado, M.
22. (2014). An improved wavelet-ARIMA approach for forecasting
23. metal prices. Resources Policy, 41, 20–29.
24. Kulkarni, A., Misra, I. and Sharma, P. (2020). Time series data analysis
25. for stock market prediction using machine learning techniques.
26. International Journal of Advanced Trends in Computer Science
27. and Engineering, 9(3).
28. Liang, Z., Zhang, J., Chen, Y. and Zhang, C. (2019). Wavelet
29. denoising combined with deep learning for stock price prediction.
30. Neurocomputing, 331, 81–89.
31. Mitra, S., Paul, R.K. and Das, M.K. (2017). ARIMA-GARCH-ANN
32. hybrid for potato price forecasting. Journal of the Indian Society
33. of Agricultural Statistics, 71(2), 121–130.
34. Nikhil, P. (2008). Arecanut marketing and prices under economic
35. liberalisation in Karnataka. Unpublished study, Karnataka State
36. Agricultural Marketing Board.
37. Paul, R.K., Yadav, R., Rai, A. and Rana, S. (2010). Forecasting of inland
38. fish production in India using ARIMA model. Indian Journal of
39. Fisheries, 57(4), 65–70.
40. Peter, M., Vogt, L. and Kempf, M. (2020). ARIMA vs. deep learning
41. models for time series forecasting of Bitcoin. Proceedings of the
42. 2nd International Conference on Finance and Economics.
43. Qi, X., Li, Q. and Zhang, Y. (2023). GRU-CEEMDAN-Wavelet for
44. stock index forecasting. Expert Systems with Applications, 224,
45. 119987.
46. Reddy, M.A. (2018). SARIMA model for tomato price forecasting in
47. major Indian states. International Journal of Current Microbiology
48. and Applied Sciences, 7(5), 1234–1242.
49. Reeja, P. (2011). Forecasting the price of natural rubber at Kottayam and
50. Bangkok markets using ARIMA and ANN models. Unpublished
51. thesis, Kerala Agricultural University.
52. Sagheer, A. and Kotb, M. (2019). Time series forecasting of petroleum
53. production using deep LSTM recurrent networks. Neuro
54. computing, 323, 203–213.
55. Sarah Abadan, S. and Shabri, A. (2014). EMD-ARIMA hybrid model
56. for rice price forecasting. Applied Mathematical Sciences, 8(42),
57. 2071–2080.
58. Shafaei, M., Adamowski, J., Fakheri-Fard, A., Dinpashoh, Y. and
59. Adamowski, K. (2016). A wavelet-ARIMA-ANN hybrid
60. model for precipitation forecasting. Journal of Water and Land
61. Development, 28, 27–36.
62. Sifuzzaman, M., Islam, M. and Ali, M. (2009). Application of wavelet
63. transform and its advantages compared to Fourier transform.
64. Journal of Physical Sciences, 13, 121–134.
65. Sima, S. and Namin, A.S. (2018). Forecasting economic and financial
66. time series: ARIMA vs. LSTM. Proceedings of the 17th IEEE
67. International Conference on Machine Learning and Applications
68. (ICMLA), pp. 1394–1401.
69. Sirisha, U.M., Belavagi, M.C. and Attigeri, G. (2022). Profit prediction
70. using ARIMA, SARIMA and LSTM models in time series
71. forecasting: a comparison. IEEE Access, 10.
72. Spices Board India. (2024). Export statistics: small cardamom
73. 2023–2024. Spices Board of India, Kochi. Available at: www.
74. spicesboard.gov.in