## Forecasting Foreign Exchange Rates using Recurrent Neural Networks (1996)

Venue: | Applied Artificial Intelligence |

Citations: | 18 - 0 self |

### BibTeX

@ARTICLE{Tenti96forecastingforeign,

author = {Paolo Tenti},

title = {Forecasting Foreign Exchange Rates using Recurrent Neural Networks},

journal = {Applied Artificial Intelligence},

year = {1996},

volume = {10},

pages = {567--581}

}

### OpenURL

### Abstract

This article proposes the use of recurrent neural networks in order to forecast foreign exchange rates. Artificial neural networks have proven to be efficient and profitable in forecasting financial time series. In particular, recurrent networks, in which activity patterns pass through the network more than once before they generate an output pattern, can learn extremely complex temporal sequences. Three recurrent architectures are compared in terms of prediction accuracy of futures forecast for Deutsche mark currency. A trading strategy is then devised and optimized. The profitability of the trading strategy, taking into account transaction costs, is shown for the different architectures. The methods described here, which have obtained promising results in real-time trading, are applicable to other markets. For years, opposing views have existed between the trading and academic communities about the statistical properties of foreign exchange rates. Traders considered exchange rates to have persistent trends that permitted mechanical trading systems (systematic methods for repeatedly buying and selling on the basis of past prices and technical indicators) to consistently generate profits with relatively low risk. Researchers, on the other hand, presented evidence supporting the random

### Citations

2980 |
R.J.: Learning Internal Representations by Error Propagation
- Rumelhart, Hinton, et al.
- 1986
(Show Context)
Citation Context ...OL The potential advantages and limitations of an artificial neural network (ANN), and, in particular, of a multilayer feedforward neural network (Werbos, 1974;sForecasting Foreign Exchange Rates 569 =-=Rumelhart et al., 1986-=-) over other statistical methods or expert systems are well known. They are universal function approximators and, being inherently nonlinear, are notoriously good at detecting nonlinearities, but suff... |

562 |
Beyond regression: new tools for predictions and analysis in the behavioral science
- Werbos
- 1974
(Show Context)
Citation Context ... method. RECURRENT NEURAL NETWORK AS A FORECASTING TOOL The potential advantages and limitations of an artificial neural network (ANN), and, in particular, of a multilayer feedforward neural network (=-=Werbos, 1974-=-;sForecasting Foreign Exchange Rates 569 Rumelhart et al., 1986) over other statistical methods or expert systems are well known. They are universal function approximators and, being inherently nonlin... |

381 |
Modeling Financial Time Series
- Taylor
- 1986
(Show Context)
Citation Context ...pendence, the null hypothesis of independence can be strongly rejected, demonstrating the existence of nonlinearities in exchange rates (Brock et al., 1991; De Grauwe et al., 1993; Fang et al., 1994; =-=Taylor, 1986-=-). It would seem to be very difficult to predict exchange rates, characterized by nonlinearities and high noise, using only high-frequency (weekly, daily, or even intraday) past prices. Surprisingly t... |

186 | Serial order: A parallel distributed processing approach - Jordan - 1989 |

109 | Neural Net Architectures for Temporal Sequence Processing - Mozer - 1993 |

86 | Beating the foreign exchange market - Sweeney - 1986 |

82 | Exchange Rate Theory: Chaotic Models of Foreign Exchange Markets - Grauwe, Dewachter, et al. - 1993 |

67 |
The significance of technical trading-rule profits in the foreign exchange market: a bootstrap approach
- Levich, Thomas
- 1993
(Show Context)
Citation Context ..., new evidence has emerged, which reinforces previous tests, on the profitability and statistical significance of mechanical trading systems in currency markets (Bilson, 1992; LeBaron, 1993; Levich & =-=Thomas, 1993-=-; Taylor, 1994). It seems that technical trading rules are able to pick up some of the hidden patterns in the inherently nonlinear price series, contradicting the conclusions reached by many earlier s... |

39 |
Stock Price Pattern Recognition: A Recurrent Neural Network Approach
- Kamijo, Tanigawa
- 1990
(Show Context)
Citation Context ...researchers have confirmed the superiority of RNNs over feedforward networks when performing nonlinear time series prediction (Connor & Atlas, 1993; Logar et al., 1993; Adam et al., 1994). (See also (=-=Kamijo & Tanigawa, 1990-=-) for an application of RNNs to recognition of stock price patterns.) The main disadvantage of RNNs is that they require substantially more connections, and more memory in simulation, than standard ba... |

25 |
Futures using a Channel Rule: A Study of the Predictive Power of Technical Analysis with Currency Examples
- Taylor
- 1994
(Show Context)
Citation Context ... has emerged, which reinforces previous tests, on the profitability and statistical significance of mechanical trading systems in currency markets (Bilson, 1992; LeBaron, 1993; Levich & Thomas, 1993; =-=Taylor, 1994-=-). It seems that technical trading rules are able to pick up some of the hidden patterns in the inherently nonlinear price series, contradicting the conclusions reached by many earlier studies that fo... |

16 |
Nonlinear Dynamics, Chaos and Instability
- Brock, Hsieh, et al.
- 1991
(Show Context)
Citation Context ...evidence has clearly shown that while there is little linear dependence, the null hypothesis of independence can be strongly rejected, demonstrating the existence of nonlinearities in exchange rates (=-=Brock et al., 1991-=-; De Grauwe et al., 1993; Fang et al., 1994; Taylor, 1986). It would seem to be very difficult to predict exchange rates, characterized by nonlinearities and high noise, using only high-frequency (wee... |

16 | Currency exchange rate prediction and neural networks design strategies - AN, Azema-Barac, et al. - 1993 |

15 |
Fractal structure in currency futures price dynamics
- Fang, Lai, et al.
- 1994
(Show Context)
Citation Context ...is little linear dependence, the null hypothesis of independence can be strongly rejected, demonstrating the existence of nonlinearities in exchange rates (Brock et al., 1991; De Grauwe et al., 1993; =-=Fang et al., 1994-=-; Taylor, 1986). It would seem to be very difficult to predict exchange rates, characterized by nonlinearities and high noise, using only high-frequency (weekly, daily, or even intraday) past prices. ... |

12 | A Test of Futures Market Disequilibrium Using Twelve Different Trading Systems - Lukac, Brorsen, et al. - 1988 |

6 |
The profitability of Currency Trading
- Bilson, Hsieh
- 1983
(Show Context)
Citation Context ... the risk of overfitting (Pardo, 1992), new evidence has emerged, which reinforces previous tests, on the profitability and statistical significance of mechanical trading systems in currency markets (=-=Bilson, 1992-=-; LeBaron, 1993; Levich & Thomas, 1993; Taylor, 1994). It seems that technical trading rules are able to pick up some of the hidden patterns in the inherently nonlinear price series, contradicting the... |

5 |
Recurrent neural networks and time series prediction
- Connor, Atlas
- 1991
(Show Context)
Citation Context ...w output pattern, can learn extremely complex temporal patterns. Several researchers have confirmed the superiority of RNNs over feedforward networks when performing nonlinear time series prediction (=-=Connor & Atlas, 1993-=-; Logar et al., 1993; Adam et al., 1994). (See also (Kamijo & Tanigawa, 1990) for an application of RNNs to recognition of stock price patterns.) The main disadvantage of RNNs is that they require sub... |

5 | Local feedback multilayered networks, Neural Computation 4(1 - Fasconi, Gori, et al. - 1992 |

5 |
W.J.B.: A comparison of recurrent neural network learning algorithms
- Logar, Corwin, et al.
- 1993
(Show Context)
Citation Context ...learn extremely complex temporal patterns. Several researchers have confirmed the superiority of RNNs over feedforward networks when performing nonlinear time series prediction (Connor & Atlas, 1993; =-=Logar et al., 1993-=-; Adam et al., 1994). (See also (Kamijo & Tanigawa, 1990) for an application of RNNs to recognition of stock price patterns.) The main disadvantage of RNNs is that they require substantially more conn... |

4 |
Identification and Prediction of NonLinear Models with Recurrent Neural Networks’, Laboratoire de Robotique de
- Adam, Zarader, et al.
- 1994
(Show Context)
Citation Context ...lex temporal patterns. Several researchers have confirmed the superiority of RNNs over feedforward networks when performing nonlinear time series prediction (Connor & Atlas, 1993; Logar et al., 1993; =-=Adam et al., 1994-=-). (See also (Kamijo & Tanigawa, 1990) for an application of RNNs to recognition of stock price patterns.) The main disadvantage of RNNs is that they require substantially more connections, and more m... |

4 | Trading Market Cycles - Ehlers, MESA |

4 |
A Complete Guide to the Futures Market
- Schwager
- 1984
(Show Context)
Citation Context ...tion accuracy is not the goal in itself, and it should not be used as the guiding selection criteria in the tests. (While this simple concept is part of the wealth of knowledge of mechanical traders (=-=Schwager, 1984-=-), it is rarely considered in tests undertaken by academicians.) The three recurrent architectures are compared in terms of prediction accuracy of futures forecast of Deutsche mark currency and its co... |

3 |
Design, testing and optimization of trading systems
- PARDO
- 1992
(Show Context)
Citation Context ...m, presumably because they used less than rigorous methodology. While it is true that mechanical trading systems need to be designed and optimized with care in order to avoid the risk of overfitting (=-=Pardo, 1992-=-), new evidence has emerged, which reinforces previous tests, on the profitability and statistical significance of mechanical trading systems in currency markets (Bilson, 1992; LeBaron, 1993; Levich &... |

2 |
The Complete Investment Book
- Bookstaber
- 1985
(Show Context)
Citation Context ... the running standard deviation of the k last periods (where k = 13, 21, 34), and technical indicators such as the average directional movement index (ADX) (Wilder, 1978), trend movement index (TMI) (=-=Bookstaber, 1985-=-), rate of change (ROC) (Kaufman, 1980), and Ehlers leading indicator (ELI) (Elhers, 1992). Inputs were normalized to zero mean and two standard deviations for all three data sets. The output was norm... |

2 | The new commodity trading systems and methods - Kaufmann - 1980 |

1 |
Practical comparisons of foreign exchange forecasts
- LeBaron
- 1993
(Show Context)
Citation Context ...verfitting (Pardo, 1992), new evidence has emerged, which reinforces previous tests, on the profitability and statistical significance of mechanical trading systems in currency markets (Bilson, 1992; =-=LeBaron, 1993-=-; Levich & Thomas, 1993; Taylor, 1994). It seems that technical trading rules are able to pick up some of the hidden patterns in the inherently nonlinear price series, contradicting the conclusions re... |

1 |
Optimal selection of parameters in technical trading systems. Boston University±Rome
- Tenti
- 1991
(Show Context)
Citation Context ...t base of technical indicators. This selection was based on previous work performed on the optimal choice of parameters of different technical indicators and on their combined use in trading systems (=-=Tenti, 1991-=-). Inputs included the compound returns of the last n periods (where n = 1, 2, 3, 5, 8), the running standard deviation of the k last periods (where k = 13, 21, 34), and technical indicators such as t... |

1 |
Forecasting Foreign Exchange Rates 581
- Weigend, Huberman, et al.
- 1992
(Show Context)
Citation Context ...fitting problem. The first is to train the model on the training set and to evaluate the model’s performance on the test set. The second approach is to use one of the many network pruning algorithms (=-=Weigend et al., 1992-=-) to reduce the network size, thereby limiting the number of hidden neurodes and hence the number of parameters to be estimated. The solution I adopted is based on a parsimonious choice of the number ... |