Based on GA-BP neural network model of futures price forecasts and Analysis

Abstract: This paper selected January 5, 2009 to October 29, the main A1001 soybean futures contract transaction data as a total of 200 training data, October 30 - November 12 of the 10 data for the test data, the use of BP neural network prediction model for the establishment of futures prices, and corrected using genetic algorithms, in order to achieve the forecast for soybean futures price analysis. The results show that the improved GA-BP neural network model fitting accuracy was significantly higher than the BP neural network model, and the futures price movements have a good prediction, to the futures market can provide investment advice. In addition, the use of the improved model can be manipulated on the futures market, the phenomenon of early warning, the regulators have a certain reference value.

Keywords: futures price forecast neural network genetic algorithm BP


Introduction and literature review

Since the 20th century, China's futures market has made considerable progress, but relatively speaking, because of China's futures market is still at a low level, market manipulation and serious investors to invest in the concept of scientific issues such as the market risk events continue to occur, directly hindered China's futures the market to mature. a lot of risk events in the final analysis, is the volatility of futures prices, futures prices and therefore the analysis and forecast trends in the futures market has naturally become the most important risk control study, at the same time, futures prices also help to understand help investors reduce risk, increase revenue, the overall stability of financial markets and coordination.

Early start in foreign futures market, futures market forecasts in the study and practice a lot of valuable work carried out, Shaikh A. Hamid, Zahid Iqba (2004 with a neural network to predict the S & P 500 index futures price volatility, Shahriar Yousefi, Ilna WEinrEIch et al (2005 proposed a wavelet-based forecasting process and used for crude oil futures to predict. In China, scholars have also tried to measure the model of futures prices to predict: Zhangfang Jie Hu Yanjing (2005 ARMA model, Xi Wang Tao (2005 ARIMA model, Liu Yifang, late Cathay Pacific (2006 GARCH-EWMA futures price forecasting model, Yangxi Liang, Dong-Hua Zhu, Liu Yi Fei (2006 BP neural network model and so the futures price forecasting has been applied.

Summary of domestic and foreign prices of futures prediction, can be found on the futures forecast a series of problems, such as: futures data with high noise, the correlation between the factors complex, futures prices are non-linear characteristics, etc. In this case, the artificial neural network method shows its unique advantages, so this choice of the BP neural network model as the basic causal futures short-term forecasting model, based on the actual application of the need to do a creative improvements.

Empirical analysis 1 variable selection and data sources this choice of Dalian Commodity Exchange soybean futures contract for the study, as a relatively stable types of transactions, it can reflect the trend to some extent the situation where the exchange transaction, its forecast to a certain extent, also may reflect other futures exchanges predict its feasibility into account data availability, integrity and other factors, this paper selected January 5, 2009 to October 29, the main A1001 soybean futures contract A total of 200 transactions data as training data, October 30 - November 12 of the 10 data for the test data from the Dalian Commodity Exchange data.

As the futures price changes is affected by many factors, in order to maximize the accuracy of the forecasts, input variable selection for the opening day, the day the highest price, lowest price that day, the day's closing price, settlement price, the day trading volume, turnover and the day positions, a total of eight inputs.

2. BP neural network model and implementation

Back-propagation model (BP neural network model can approximate any nonlinear function, the operation is divided into the forward signal propagation and error back propagation in two phases: the first stage, the sample passed from the input layer, through the hidden layer of processing, transfer to the output layer. If the actual output and the output layer does not match the expected output, then transferred to the second stage, the output error to some form of hidden layer by layer by layer to the input layer back propagation, and error apportioned to all layers of cells, resulting in cell layers of the error signal, and thus the right to fix the value of each unit.

According kolmogorov theorem, a three-level BP neural network is sufficient to complete any of the n-dimensional m-dimensional mapping, which generally requires the use of a hidden layer is sufficient. The number of hidden layer nodes using trial and error method to determine this is 20. In order to improve the training accuracy, this paper will set the initial learning rate of 0.05, and using adaptive learning rate function, the training process in the future based on the training error to automatically adjust the learning rate. Meanwhile, the paper selected continuously differentiable function that is tansig S-tangent function as transmission function, which can be derivative of the nonlinearity with saturation, enhanced network nonlinear mapping ability. Links to free download http : / / eng.hi138.com
Based on the above model with the parameters set in matlab to achieve the results shown in Figure 1, can be seen from the figure, the soybean futures price forecast is roughly the same trend, but the overall error is larger, although the use of adaptive learning to improve the convergence rate, but the BP gradient descent algorithm there is a big limitation. In order to improve the neural network weight adjustment, so the genetic algorithm with BP neural network optimization.

3 models to improve and achieve global optimization genetic algorithm is a search algorithm, its basic idea is to first solve the problem represented by genotype, by selection, crossover and mutation selected from the individual to adapt to the environment, and seek the optimal solution problem, have better global search performance. Genetic algorithm applied to the neural network model, to achieve the two complement each other, played a neural network mapping capabilities and extensive global search ability of genetic algorithm, but also to accelerate the learning speed of the network, integrated improve the entire learning process model approximation ability and generalization ability.

In MATLAB, the results shown in Figure 2, can be seen, adding weight and genetic algorithm to optimize the threshold after the GA-BP model results closer to the true value can be better and more accurate fit in the GA-BP model case, the advantages of outstanding performance in convergence speed, short cycle above. Compared with the traditional BP algorithm to reduce the weight threshold of random initialization, GA-BP can significantly reduce the convergence time convergence are given in Table 1 below:
Table 200 in the previous iteration, for example, genetic algorithms mse can be reduced to e-5 orders of magnitude, while the simple BP neural network can only achieve e-3, and the decrease in the latter part of the iteration more and more obvious.
Conclusion Based on BP neural network model for empirical analysis, the results from the forecast, forecast and actual values ??of the trend is the same, but the predicted and actual values ??with a large deviation, which is due to BP neural network of their own problems caused by.

To solve the BP neural network model can not accurately predict the price of the issue, taking into account the adjustment of BP algorithm is due to the weight limitations, this paper, genetic algorithm optimization model using genetic algorithm global search for the right values, to avoid the BP algorithm limitations from the results, the predicted and actual values ??of error is small, can accurately predict futures prices.

Using this model, we can also predict the futures price manipulation in the predicted value and the collection of the history of futures trading on the normal real data to estimate model parameters, in order to predict the future period of futures prices, forecasts and real value through interval between test error and the error compared to the forecast range can determine whether the price to be manipulated if the forecast data and continuous with the market to deviate from the real data, and a larger variety than the mean error of normal transactions, can be manipulated to determine the existence of the phenomenon, namely the establishment of the moment type:

That price is being manipulated in the prediction interval in which the real and the predicted range of values ??and predicted values, Oj and Dj are the true test interval and predicted values, n is the length of prediction interval data, m is the length of inspection interval data , to allow the maximum volatility. At this point the regulatory authorities should be careful observation, to stop the manipulation, to maintain price stability and market order.

Links to free download http://eng.hi138.com

Futures Market Papers