Abstract
Forecasting foreign exchange (FX) rates is a complex task due to the inherent noise and non-linearity of financial markets. This study investigates the impact of input data frequency—daily, weekly, and monthly—on the predictive accuracy of Back-Propagation Neural Networks (BPNN). By analyzing historical data for the British Pound (GBP) and Japanese Yen (JPY) against the U.S. Dollar, we demonstrate that weekly input data provides an optimal balance between noise reduction and signal retention. Our results show that BPNNs trained on weekly data consistently outperform both daily and monthly models, as well as the standard Random Walk benchmark.
Introduction
Predicting foreign exchange rates is notoriously difficult due to their non-linear, chaotic behavior. While FX markets generate data at high frequencies, researchers must choose a specific input frequency—daily, weekly, or monthly—to build their models. The common practice is to match the input frequency to the output horizon (e.g., using daily data to predict tomorrow’s rate). However, this approach faces a fundamental tension: high-frequency (daily) data is often riddled with "noise" that obscures underlying trends, while low-frequency (monthly) data is too sparse, smoothing over essential volatility.
This paper explores whether a "Goldilocks" frequency exists that maximizes the predictive power of neural networks across different forecasting horizons.
Methodology
The researchers employed a standard three-layer Back-Propagation Neural Network (BPNN). To ensure robustness, several configurations were tested:
- Input Nodes: The number of lagged observations was varied (3, 5, 7, and 9) to determine the optimal "memory" for the network.
- Benchmarks: Every neural network was compared against a Random Walk (RW) model, the industry-standard "naive" benchmark which assumes the best predictor for tomorrow is today's price.
- Data: Historical exchange rate data for the U.S. Dollar against the British Pound (GBP) and the Japanese Yen (JPY) from January 1990 to December 2002 was used. The dataset included 3,291 daily observations, 678 weekly averages, and 156 monthly averages.
Results and Analysis
The performance was measured using Root Mean Squared Error (RMSE), where lower values indicate higher accuracy. The study first established the performance of the Random Walk benchmark across different horizons.
The "Goldilocks" Frequency
The most significant finding is that weekly data consistently outperformed other frequencies. As shown in the results for the 5-input node configuration, weekly input data achieved the lowest RMSE, significantly beating the Random Walk benchmarks for 1-day ahead predictions.
Specifically, for 1-day ahead predictions, neural networks using weekly input data achieved an RMSE of approximately 0.0045 for GBP and 0.0063 for JPY. In contrast, models using daily or monthly inputs often failed to beat the Random Walk model. This suggests that weekly data is slow enough to filter out the "jitter" of daily trading but fast enough to capture the medium-term momentum that monthly data ignores.
Complexity and Horizons
The study also found that network complexity matters. Networks with 5 or 7 input nodes generally performed better than those with 3 or 9, indicating a "U-shaped" relationship where too few inputs lead to under-fitting and too many lead to over-fitting on noise.
Long-term (1-month ahead) forecasting remained the most difficult challenge. While weekly-input NNs showed a slight edge, the Random Walk model became much harder to beat as the forecasting horizon increased and market volatility introduced more uncertainty.
Conclusion
In the volatile world of FX forecasting, more data is not always better. This research demonstrates that weekly data provides the optimal input frequency for neural networks, offering a superior balance of signal and noise compared to daily or monthly data. Future work may explore more modern architectures like LSTMs or the inclusion of exogenous economic indicators to further refine these predictive models.
Summary updated