Hybrid Wavelet-ARNN-ARIMA Model for Forex Price Forecasting

1. Introduction

The foreign exchange (Forex) market, with a daily trading volume exceeding $5 trillion, presents significant opportunities and risks. Accurate price forecasting is crucial for effective trading strategies. However, Forex data is characterized by high volatility, noise, and complex non-linear patterns, making prediction exceptionally challenging. Traditional linear models like ARIMA often fall short in capturing these dynamics. This paper proposes a novel hybrid methodology that synergistically combines Wavelet Denoising, an Attention-based Recurrent Neural Network (ARNN), and the Autoregressive Integrated Moving Average (ARIMA) model to address both the linear and non-linear components of Forex time series, aiming for superior predictive performance.

2. Related Literature

2.1 Wavelet Denoising

Wavelet Transform is a powerful tool for time-frequency analysis, effectively separating signal from noise in non-stationary financial data. By decomposing a time series into approximation and detail coefficients, it allows for the selective removal of high-frequency noise components that can obscure underlying trends and autocorrelation structures, a preprocessing step critical for improving model input quality.

2.2 Neural Networks in Finance

Neural Networks, particularly Recurrent Neural Networks (RNNs) and their variants like LSTMs, have shown promise in modeling complex, non-linear financial time series. The integration of attention mechanisms, as seen in models like the Transformer, allows the network to focus on the most relevant past observations for making a prediction, enhancing sequence modeling capabilities.

2.3 Hybrid Forecasting Models

The "decomposition and ensemble" paradigm is well-established. The core idea is to use different models to capture different data characteristics (e.g., linear vs. non-linear, trend vs. seasonality) and then combine their forecasts. This paper's contribution lies in the specific combination of wavelet denoising for preprocessing, ARNN for non-linear patterns, and ARIMA for residual linear components.

3. Methodology

3.1 Data Preprocessing & Wavelet Denoising

The original Forex price series $P_t$ is decomposed using Discrete Wavelet Transform (DWT): $P_t = A_J + \sum_{j=1}^{J} D_j$, where $A_J$ is the approximation coefficient (low-frequency trend) and $D_j$ are detail coefficients (high-frequency noise at level $j$). A thresholding function (e.g., soft thresholding) is applied to the detail coefficients to suppress noise, followed by reconstruction to obtain the denoised series $\tilde{P}_t$.

3.2 Attention-based RNN (ARNN) Architecture

The model uses an encoder-decoder RNN framework with an attention layer. The encoder (LSTM cells) processes the input sequence $\tilde{P}_{t-n:t-1}$ and produces a sequence of hidden states $h_i$. The attention mechanism computes a context vector $c_t$ as a weighted sum of these encoder states: $c_t = \sum_{i=1}^{n} \alpha_{t,i} h_i$, where the attention weights $\alpha_{t,i}$ are learned by a feed-forward network. The decoder LSTM then uses $c_t$ and its previous state to predict the non-linear component $\hat{N}_t$.

3.3 ARIMA Model Specification

The ARIMA(p,d,q) model fits the linear relationship in the time series. After the ARNN captures the non-linear part, the residual series $R_t = \tilde{P}_t - \hat{N}_t$ is modeled by ARIMA: $\phi(B)(1-B)^d R_t = \theta(B) \epsilon_t$, where $\phi$ and $\theta$ are AR and MA polynomials, $B$ is the backshift operator, $d$ is the differencing order, and $\epsilon_t$ is white noise. This yields the linear forecast $\hat{L}_t$.

3.4 Hybrid Integration Strategy

The final prediction $\hat{P}_t$ is a simple additive combination of the forecasts from the two constituent models: $\hat{P}_t = \hat{N}_t + \hat{L}_t$. This assumes the linear and non-linear components are additive and have been effectively separated by the modeling process.

4. Experimental Results

4.1 Dataset & Experimental Setup

Experiments were conducted on high-frequency USD/JPY five-minute exchange rate data. The dataset was split into training, validation, and test sets. Baseline models for comparison included standalone ARIMA, standard LSTM, and other neural network architectures from related literature.

4.2 Performance Metrics & Comparison

The proposed hybrid model achieved a Root Mean Square Error (RMSE) of 1.65 and a directional accuracy (DA) of approximately 76%. This outperformed all baseline models. For instance, a standalone ARIMA model might achieve a DA of ~55-60%, while a standard LSTM might reach ~65-70%, highlighting the value of the hybrid approach and preprocessing.

4.3 Result Analysis & Discussion

The significant improvement in directional accuracy is particularly noteworthy for trading applications, where predicting the correct price movement direction (up/down) is often more critical than the exact price point. The reduction in RMSE indicates overall forecast error minimization. The results validate the hypothesis that wavelet denoising stabilizes the input and that the hybrid model effectively captures both linear and non-linear dependencies.

5. Technical Analysis & Expert Insights

6. Analysis Framework & Case Example

Scenario: Predicting the next 5-minute candle for EUR/USD during a major central bank announcement (e.g., ECB press conference).

Framework Application:

1. Wavelet Preprocessing: The raw 5-min price series from the last 4 hours (48 data points) is decomposed. The high-frequency "detail" coefficients spiking during the announcement are heavily thresholded, smoothing out micro-noise while preserving the major directional jump.
2. Model Decomposition:
   • ARIMA Component: Models the underlying momentum and mean-reversion tendency that existed before the news. Its forecast might be a slight continuation of the pre-news trend.
   • ARNN Component: The attention mechanism focuses heavily on the most recent, volatile post-announcement price bars. It learns from similar historical "news shock" patterns to predict the likely short-term overreaction and subsequent partial retracement.
3. Hybrid Forecast: The final prediction = (ARIMA's trend-based forecast) + (ARNN's news-impact adjustment). This is more nuanced than either model alone, which might either underreact (ARIMA) or overfit to noise (a standard RNN on raw data).

7. Future Applications & Directions

• Multi-Asset & Cross-Market Prediction: Extend the framework to model correlations between Forex pairs, equities, and bonds. The ARNN encoder could process multiple related time series simultaneously.
• Integration with Reinforcement Learning (RL): Use the hybrid model's predictions as the state representation for an RL agent that learns optimal trading execution policies, directly optimizing for profit rather than prediction error.
• Explainable AI (XAI) Enhancements: Develop methods to attribute the final forecast to specific linear trends (via ARIMA coefficients) and specific past time points (via ARNN attention maps), providing traders with actionable reasons for the prediction.
• Adaptive Online Learning: Implement mechanisms for the model to continuously update its parameters with new data in a streaming fashion to adapt to changing market regimes, moving beyond static train-test paradigms.

8. References

1. Bank for International Settlements (BIS). (2019). Triennial Central Bank Survey of foreign exchange and OTC derivatives markets.
2. Box, G. E. P., Jenkins, G. M., Reinsel, G. C., & Ljung, G. M. (2015). Time series analysis: forecasting and control. John Wiley & Sons.
3. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I. (2017). Attention is all you need. Advances in neural information processing systems, 30.
4. Zhang, G. P. (2003). Time series forecasting using a hybrid ARIMA and neural network model. Neurocomputing, 50, 159-175.
5. Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural computation, 9(8), 1735-1780.
6. Wolpert, D. H., & Macready, W. G. (1997). No free lunch theorems for optimization. IEEE transactions on evolutionary computation, 1(1), 67-82.
7. DeepMind. (2023). Research in Adaptive Agents. Retrieved from https://www.deepmind.com/research/highlighted-research/adaptive-agents