Future of Trading – Neural Networks

Author: Sahitha Abdulla, NMIMS

Technological changes has revolutionized the way financial assets are traded, be it equity or forex or commodities. The first and the major technological intervention in the financial services markets started in the mid 1990`s with advent of Algorithm trading. Algorithm trading soon gave way to the next big and the most HAPPENING technologies in the financial markets - the Neural Networks.

Numerous sophisticated financial institutes have been reported using Neural Network to make trade decisions. And because of its effectiveness and scope, many institutions have highly invested in the area of artificial neural networks for financial trading.

“Apart from the U.S. Department of Defense, the financial services industry has invested more money in neural network research than any other industry or government body,”

According to a 1993 book by Robert R. Trippi and Efraim Turban.

An Artificial Neural Network (ANN) is built in such a way so as to model the human brain. A typical human brain consists of a huge network of neurons with around 1000000000000 neurons and 10000000000000000 connections amongst the neurons. The memory cells of brain is represented by the neurons in ANN and the connections between the memory cells in brain is represented by the weight value. The output from a neuron is calculated using the weights of the connections and the input values. Once a network is trained correctly, it discovers the rules between the input and output rather than memorizing them. The network will be able to give outputs from a different set of input. In a sense, the network has become an expert for this particular type of input. A neural network can perform tasks that a linear program like an Algorithm trading cannot. The biggest advantage (and disadvantage) of Neural network is that they are not emotional and hence they are able to discover the hidden rules and hidden numbers without getting affected by emotions and hence are well suited for financial applications. In addition, neural networks can adapt very quickly to the changing environment by retraining unlike humans who adapt to changes very slowly.

What is Neural Network and how are they different from conventional programming?

Neural network is a information processing system consisting of a number of interconnected processing elements called neurons. These neurons possess unique learning capabilities. Unlike a computer programming where the algorithm is pre determined and fed into the computer , the neural networks is trained to arrive at a logical conclusion based on past data. It is trained to work much like the human brain which is an expert in this.

How to train a neural network?

There are various steps involved in teaching a neural network

1.  Supervised Learning

The neurons are modeled using a set of data called training set. The training set includes factors like stock index variations, the macroeconomic development, the news about stocks, inflation, unemployment etc. After successful training neural network will be able to extrapolate i.e classify, predict and estimate values in future.

2.  Back Propagation Learning Algorithm

The neurons learn though back propagation learning algorithm which seeks to minimize the difference between the output of the neural net and the desired output. The difference is calculated as error component and is added back into the system with a different weight in order to minimize the error. This is a continuous process until minimum deviation is achieved. 

The outcome of supervised learning might not always be as desired, because some important parameter like the microeconomic factors in which the industry the firm operates is not taken into consideration. The learning is so immense that the neural network finds links with various external factors that are not clearly visible to an individual.

3.  Text Mining

In order to use AI in the chaotic and nonlinear financial markets the learning has to be beyond supervised learning. The enhancement can be in terms of Text mining. This technique involves scanning large chunks of news pertaining to the script or markets in totality and look for words or cues, something that a typical analyst looks out for which is not very clearly evident to the common reader. 

The Text Mining algorithm includes the following attributes:

a.   Train the network:

The network needs to be trained to look for aspects , data, words which a typical stock analyst would look for while reading news. A attribute dictionary is designed in this process in order to increase efficiency as the complexity increases.

b. Filter the redundant data:

While analyzing millions of articles there will be more redundant data than the useful ones. These need to be filtered out at early stage so as to minimize the burden on the network.

c. Rearrange and Classify the data:

The data needs to be properly classified in order to establish trends among various inter related articles and to forecast probable outcomes based on these parameters.

For instance a stock is generally considered overpriced if it has a high P/E ratio compared to the industry in which the form is operating. The converse will hold true for a undervalued stock. This is only one method of estimating the price, there are many other techniques as well

1.    Discounted cash flow valuation:

Here the value is estimated based on the present value of some parameters in the cash flow statements, which includes dividends, free cash flow and operating cash flow.

2.    Relative Cash flow :

The value of the stock is estimated by comparing the price at which the script is trading in comparison with its fundamentals.

3.    Cost of capital

4.    Capital budgeting

A Dividend discounted model (DDM) can help us analyze a script based on the above parameters and this can be applied as an input to the neural network to verify if the script is stable and fairly priced to be purchased by the brokers.

Value of the stock =D1+D2+D3+T (1+k) (1+k)2 (1+k)3…… (1+k)∞

Where

Dt   = Dividend during time period t

K = the required rate of return of the stock 

T = Terminal value of the stock.

DDM and random walk theory are some of the popular methodologies used to analyse stocks using neural networks.

Do neural network actually work or is it a work of fiction?

Neural networks are more than 80 % accurate for short trades, for example it can very accurately predict if a stock will move up by 1% in the next 10 minutes or so. The network can analyze more than 10,000 times the data an analyst can read .

Analysts tested a neural network called Arizona Financial Text System (AZFinText) under unusual market conditions that prevailed in 2005 by providing the system with 9,000 news articles and 10 million stock quotes. The predictions of the network gave returns of 8.5% which was higher than the returns offered by S&P 500 index and other top 10 funds. A human – machine collaboration was also tried out during the same period where the analyst would select the a portfolio based on various analysis and the AZFinText had the liberty to choose the stock to invest in. this experiment gave a result of 20% returns which was surprising for the very creators of the system.

The creators of these networks though agree that the system is quite effective only in the short run as the variables involved as comparatively less when compared to devising long term strategies. 

How different are the neural networks from Algorithmic trading?

The underlying objective of both these estimation techniques is the same, but the modulus operandi is very different. Under algorithmic trading the trader first needs to perform a pre trade analysis and define the algorithm which includes parameters like the bench mark price, when the trade needs to be executed and any deviation rules i.e when not to execute the trade. It is a method which gives the trader liberty to set the rules and the system will execute the instructions when the desired criterion is attained. This is particularly useful for investors involved in large volume trades.

Neural networks are an extension of Algorithmic trading with less human intervention. The network is fed with data which include the index, news articles pertaining to the market, global financial and economic news. The network is then allowed to correlate all the data and pick out stocks which might perform well during a intra trade session. The success percentage of these networks is 70% and further research is ongoing to include many more parameters to increase the success ratio.

Conclusion

Neural networks is a mathematical model which helps in identifying the underlying patterns and trends within stocks and stock market among diverse data and information , in order to predict highly accurate short term forecasts. Although neural networks doesn’t give 100 % accuracy it is still a tool which needs to be further developed and fine tuned to increase the success ratio. The applications of neural networks in finance is immense and is as much applicable to forex, commodities and options as much as it is for stocks as discussed in the paper. Traders can use this tool to spot hidden gems in the market and increase their returns.

“If ever there were a field in which machine intelligence seemed destined to replace human brainpower, the stock market would have to be it. Investing is the ultimate numbers game, after all, and when it comes to crunching numbers, silicon beats gray matter every time.”