Sphere of neural networks’ financial applications is actually boundless. Any goal concerning manipulation of financial tools, whether it is currency or securities, entails risks and needs to be estimated and predicted thoroughly. How quotations of the basic currencies will change tomorrow? Will outwardly successful company give credit back? How to choose profitable and at the same time reliable “investor’s portfolio”? Analytic departments of financial (and not only financial) companies have to solve all these and hundreds of other problems, using all kinds of analytic instruments. It is no mere chance that financial applications make up quarter of neural networks products market (let’s remind that size of global market of neural networks increases on 40% annually and exceeded 600 million dollars in 1994).

Neural networks appeared in Russian market exactly in financial sphere. Fifty neural networks software suites have been sold in Russia for the last quarter, and overwhelming majority (about forty) have been bought for financial use (twenty – for banks). And after such “big shots” as Inkombank, Autobank, and  Eleksbank turned their favorable eyes on neural networks, apparently, that the process of further distribution of neural networks came into self-generation mode (it seems that arguments like “I want the same toy Alexander has” are in common use not only among children). Many companies use neural networks software suites not only for solving concrete tasks, but also for not to be behind competitors in mastering of new weapon.

But why are neural networks so attractive? What makes them so good for solving various tasks on prediction and discernment? Not going into details, we can say that currently there are four essentially different approaches for solving tasks of analysis.

Firstly, if data are interdependent and their volume is rather small, you can use classic analysis methods (for example, correlation methods). Secondly, you can create an expert system, using rules “if…then”. Thirdly, you can use methods of fuzzy logic (they are in fashion today) and operate with qualitative characteristics, such as “majority”, “reliable”, “some”, etc. And, at last, fourthly, in cases, when input data volume is vast, you have no idea about their interconnections, some part of information is distorted and some part of information is lost, then neural networks will help you. You only have to enumerate factors, which essentially influence the forecasted value, and to find enough examples describing behavior of these values in past. Neural network will “be adjusted” on the given totality of examples by itself, minimizing summary error of prediction. Moreover, analysis of adjusted network allows to find hidden dependences between input and output data, which often remain “behind the scenes”, when using traditional methods. Assuming that character of interconnection between specified parameters won’t essentially change for some time, you can use adjusted and trained neural network for short-term (and sometimes long-term) prediction.

“Well, it is not for us…”, - disappointedly says reader, who was nearly fascinated by cold shine of “perfect weapon”. Our financial market is formed solely by ruble rate, and ruble is unpredictable. And also we have mysterious parliament, peculiar taxes, grandiose financial pyramids, extremely simple advertisement, and many other things that make business look like roller coaster “in Russian style”. It is the real truth, but let us give some more precise definitions.

Firstly, ruble as well as any other currency (including various coupons, levs, tugriks, etc) can be predicted quite well – we only need high-end computers and one dozen of experts. Game on the world currency market has turned into a war of supercomputers a long time ago. Stories like legend about Soros, whose daring game brought him one billion during one day, gradually become a thing of the past. And when you read in the “Financial Times” that yen “dropped” dollar on two points, it means that yesterday supercomputer of some “Sakura Bank” beat its competitor from “Chase Manhattan” (or vice versa).

Secondly, we are not as exceptional as we think we are. Our history saw many events: local conflicts, revolts of interregnum, big scandals around false corporations, etc. Scenarios of their development and influence on financial market are studied well enough. Besides, an entire system of macroeconomic indicators, like Dow-Jones average or S&P 500, has been worked out for many years; they are sui generis “barometers” of current state of market. Many of these parameters have been punctiliously recorded since 1901, and databases of reports for several last months are the object of brisk trade. You might answer that there were no such indicators in Russia. Believe us – they will appear in the nearest future, before you collect dividends from your voucher.

At last, thirdly (and it is the most important thing), state of financial market is not determined by one (even dominating) parameter, but the sum total of processes of different nature that have various speed of response. For example, dollar rate can fall down within one hour (it happened more than once), but broker’s expectations, reflected in futures’ quotes for the following month, will change much more slowly. And if you have a tool for assessment of speed of this process (neural networks) and iron nerves, you will have enough time to prepare and play competently and skillfully. Eventually, what is more important for you – to predict “black Tuesdays” or to get stable profit?

Let’s study the use of neural network in financial prediction, using concrete example – prediction of currency futures rate. Company that had successfully used neural networks software suite Brain Maker Pro for its internal tasks, decided to try it for prediction of futures quotes on Moscow Commodity Exchange in March, 1995. Futures contract quote on $1000 with June date of payment was chosen as the object of prediction (input parameter of neural network). Input parameters for training of network were changes of futures rates during May, June, and July for the last four trading days (dynamics of the last day was taken into account as a separate parameter, evaluated using special formula), and ruble-to-dollar rate for four days. Training data included forty last trading days (two months). After six and a half thousand steps of training algorithm (it took 3 minutes of calculations) the neural network showed quite adequate reaction on all set of given parameters, i.e. it was trained.

Then the neural network was used for ten days for prediction of today’s futures rate for June. The calculations were made when the current dollar rate and the last quote of May futures had become known. As a rule, there was about one hour left till trading on June futures. The result was unexpectedly accurate: the network didn’t once make mistakes in predictions of change tendencies (decline or growth), and deviation of real rate from predicted rate made up no more than 10 rubles in nine cases from ten.

Of course, such method can also be used for playing on GKO, for currency dealing, and for many other applications. But the given example is quite typical, because it shows some interesting rules of work with neural networks. Firstly, experience has shown that in spite of simple interface of neural networks they are very delicate and began to obey to their owners only after 2-3 weeks of intensive study and “adaptation”. Secondly, when choosing toolkit, pursuit of low prices does not prove its value. Of course, you can use so called “student” version of neuron network suite for $300, but for new task adjustment you will need high-end professional suite like BrainMaker Pro, OWL or something similar. Thirdly, analytic tools of neuron network suites open new opportunities for study of parameters of tasks, as an adjusted network accumulates in itself hidden regularities of a subject field. For example, in the given example with futures the neuron network suite was also used for analysis of influence of current dollar rate’s fluctuation on futures quotes with delayed date of performance. There was found quite an unusual regularity: on the level +30 rubles for bid there comes a kind of “saturation” of brokers’ expectations, and further growth of rate does not affect futures quotes (generally, it can be explained, however it is quite problematic to find such dependence, using traditional analytic methods).

MetaTrader Expert  " Neuro 1" showed moderate results for more than 4 months real account tests. It is based on simple Artificial Intelligence model and few simple indicators, which feed the network with information. It has small TP to cut profit very fast and the optimization showed that with current parameters of market if he finds the wave he can cut allot! If market conditions doesn't met the profit needs it will sleep. The sleep time also being controlled by Neural Network and you can never say when it get "the good mood" again. Read more>>>

And what is about responsibility for decisions making? It is clear that the price for mistake in financial operations is too high! We advice you (following the Americans) to use the following method: if neural network suite shows approach of “black Tuesday”, and, on the contrary, your broker is sure in success – trust broker. If he makes mistake, you won’t lose (because your broker is, probably, quite experienced, and together with him majority of competitors will make mistakes, too) – your neural network suite, which predicted financial collapse correctly, will also predict a winning game strategy. If the neural network makes mistake, you won’t lose, too; you’ll just once more note that computers can’t be trusted. Besides, there are a lot of tasks, where the price of one mistake is not so high, and you have some time to make additional adjustments. For example, Inkombank specialists are seriously thinking over use of neural network suites for selection of the most optimal places for opening of new branch offices. From methodological point of view such approach is safe: undoubtedly, the net will make choice among good variants easier, and knowingly bad variants can be set aside by the expert himself.

If to formulate the place of neural networks in arsenal of your financial tools in one phrase, we can say, that the neural network is a prompter for skilful analyst.  Neural network won’t help looser, but good broker can improve his game in many times.

BJF Trading Group

http://iticsoftware.com