Comparing RMB Exchange Rate Forecasting Accuracy based on Dynamic BP Neural Network Model and the ARMA Model

Zhiqiang Ye, Xiang RenJournal of Stock & Forex Trading 2016 · DOI

TL;DR

This paper compares the forecasting accuracy of the ARMA model and a dynamic BP neural network for the RMB exchange rate. The study finds that the BP neural network significantly outperforms ARMA in predicting both the direction of exchange rate fluctuations and the magnitude of deviations, highlighting the advantage of nonlinear models for complex financial markets.

What Problem Does This Paper Address?

Predicting exchange rates is crucial for businesses and policymakers, especially for currencies like the Chinese Renminbi (RMB) which has undergone significant reforms. Following China's shift to a managed floating exchange rate system in 2005, the RMB's value has become more dynamic and influenced by market forces, making accurate forecasting a challenge. Traditional time-series econometric models like ARMA are effective for linear relationships, but financial markets often exhibit complex, nonlinear dynamics. This paper addresses the need to compare the effectiveness of a linear model (ARMA) against a nonlinear model (dynamic BP neural network) for forecasting the RMB exchange rate, particularly in the context of its post-reform complexity.

Methods

The study employs two primary modeling approaches:

  • ARMA (Autoregressive Moving Average) Model: This is a standard linear time-series model that captures linear dependencies between past values of the exchange rate and past error terms. The authors use the autocorrelation criterion (AIC) to determine optimal model orders (n and m) and estimate parameters using the least squares method.
  • Dynamic BP Neural Network Model: This approach utilizes a backpropagation (BP) neural network designed to handle time-series data. The model learns nonlinear relationships between past exchange rate values and future ones. The authors explore different network architectures, settling on an input layer of 3 neurons, a hidden layer of 7 neurons, and an output layer of 1 neuron. They use the Levenberg-Marquardt algorithm for training and employ a specific criterion to predict the direction of fluctuation.

Datasets

The study uses historical RMB exchange rate data against the U.S. dollar from January 1, 2011, to October 10, 2012. This dataset, comprising 321 samples, is divided into training and testing sets to evaluate the models' performance.

Key Insights and Ideas

  • Nonlinearity Matters: The core insight is that the inherent complexity and nonlinear nature of the RMB exchange rate market make nonlinear models like the BP neural network more suitable for accurate forecasting than linear models like ARMA.
  • Directional Accuracy is Key: The paper emphasizes that predicting the direction of exchange rate fluctuations is as important, if not more so, than predicting the exact value. The BP neural network demonstrates superior performance in this regard.
  • Model Complexity vs. Data: The authors acknowledge the potential for overfitting in neural networks but argue that with a sufficiently large dataset (321 samples) and an appropriately chosen network architecture (3-7-1), the risk is mitigated, and the model's generalization ability is maintained.
  • AIC for Model Selection: The use of the Autocorrelation Criterion (AIC) is highlighted as an effective method for determining the optimal orders for the ARMA model, ensuring a robust linear model baseline.

Key Results

  • Directional Accuracy: The BP neural network achieved a directional accuracy of 59.62% for predicting RMB exchange rate fluctuations, significantly outperforming the ARMA model's 46.73%.
  • Deviation Accuracy: The root mean square deviation (RMSE) for the BP neural network was 0.0519, which is lower than the ARMA model's RMSE of 0.0665.
  • Statistical Significance: A Diebold-Mariano test indicated that the BP neural network model's performance was statistically superior to the ARMA model (S(1) = 3.167, p-value = 0.0015).
  • Training Efficiency: The BP neural network training process was efficient, converging quickly (less than 1 second) and achieving a low error function.

Limitations and Future Work

  • Model Specificity: The study focuses on the RMB/USD exchange rate. While the findings suggest the superiority of nonlinear models, their performance might vary for other currency pairs or under different market conditions.
  • Data Period: The analysis is based on a specific historical period (January 2011 - October 2012). Future work could explore longer time horizons and more recent data to assess model robustness.
  • Parameter Tuning: While the authors experimented with network parameters, the paper notes that there isn't a uniform standard for determining the number of neurons in the input and hidden layers, suggesting this could be an area for further optimization.
  • Market Complexity: The paper acknowledges that China's foreign exchange market is more complex than mature markets, implying that while the BP neural network shows promise, further research into more sophisticated nonlinear models or hybrid approaches might yield even better results.

Summary updated

Abstract

This paper uses the dynamic back propagation (BP) neural network model and the autoregressive moving average (ARMA) model to forecast the RMB exchange rate based on the data from January 1, 2011 to October 10, 2012. The results show that the dynamic BP neural network model works better than the ARMA model in evaluating both the trend and the deviation of RMB exchange rate.
BibTeX
@article{ye2016comparing,
  author = {Zhiqiang Ye and Xiang Ren},
  title = {{Comparing RMB Exchange Rate Forecasting Accuracy based on Dynamic BP Neural Network Model and the ARMA Model}},
  journal = {Journal of Stock \& Forex Trading},
  year = {2016},
  doi = {10.4172/2168-9458.1000161},
}
Figure