Development of Artificial Intelligence Algorithm Trading Systems

Abstract

With easier fund management due to computational trading algorithms, which can learn from multiple sources of information, technical trading rules evolved to include artificial intelligence machine learning method using variants of neural networks. Backpropagation neural network outperforms common technical analysis indicators and traditional statistical models. Neural networks are used in machine learning by inputting past historical price data and technical indicators to predict the next output. The training is performed to achieve the lowest mean error between the predicted output and the target which is the actual close. This chapter applies a neural network enhanced technical indicator (N-CAMA′) to crude light oil futures (CLOF) which experienced high volatility in recent years.

Appendix

In this study, MATLAB’s “narxnet” NARNNX is used to establish a one-step-ahead prediction model. The NARNNX model is built on the linear ARX method, which is generally applied in time series modeling.

The fundamental equation for the NARX model is:

$$\hat{c}_{t + 1} = f\left( {C_{t,} C_{t - 1} , \ldots ,C_{t - d,} {\text{AMA}}_{t,} {\text{AMA}}_{t - 1} , \, \ldots ,{\text{AMA}}_{t - d} } \right)$$

(5.3)

where the obtained value of the dependent output signal ĉ_t+1 is regressed on d former values of the target signal C_t and d previous values of exogenous (independent) input signals AMA_t. NARNNX model by applying a FBNN to estimate the function f. Moreover, weights and biases in an FBNN will be adjusted continuously to minimize the error term between output ĉ_t+1 and target C_t to achieve the lowest mean of the error terms.

The architecture of a NARX network includes the number of hidden layers, the number of delays (the number of past data of that network that account for training), and portions of training, validation, and testing. NARX networks divide the data into three subsets: training set, validation set, and testing set, which sets will be spread randomly along the time series, with a configured percentage for each of them; in this study, the proportions are training 75%, validation 15%, and testing 15%.

In this experiment, the close and AMA′ are processed as inputs. Using 10 neurons, 2 delays, 1 hidden layer, and Levenberg-Marquardt optimization, 2000–2014 data is used for training. The output is then used as the predicted close for the next day.

The training period is from 2 January 2000 to 31 December 2014 while the validation period is from 2 January 2015 to 31 December 2017 and the out-of-sample period is from 2 January 2018 to 31 December 2018. The method used is similar to the approach used by Yao et al. (1999) which is to long the futures when the predicted output is higher than the current close and to short otherwise.

We ran tests from one to 15 neurons. Interestingly, 10-neurons ANN model is selected in accordance with the earlier tests to employ the model with the least NMSE.

Although the best architecture to apply depends on the type of the problem to be solved by the network, there is no rule of thumb to select the number of hidden layers and delays (Kaastra & Boyd, 1996). In this study, Levenberg-Marquardt optimization is used as the training algorithm, which is a built-in.

The objective of this exercise is to determine whether the abnormal return from utilizing the forecast of next period’s price is significantly higher than the passive buy-and-hold control. If ĉ_t+1 > C_t, where, ĉ_t+1 is the predicted closing price output for the next period, and C_t is the current closing price, the trading strategy is to buy, and if otherwise, the strategy is to sell.

The results of the prediction accuracy shown in hit rate and the profitability performance will be evaluated.