Hello, Leonid. We would like to congratulate you on such good results of your Expert Advisor. How long ago did you start trading?

Thank you for congratulations. Well, it is a long story. I had two interests from my childhood - radio engineering and music. In different times one of them was prevailing. At school I was playing in a school music group and assembled radios, transmitters, TV-sets and even assembled a synthesizer for myself. After school I entered Moscow State University of Radio-Engineering, Electronics and Automation.

There I also "worked" in both spheres - music and radio engineering. I played in the institute group and studied engineering of radio sets. After getting a diploma in 1989 I devoted all my time to music, thus radio engineering was forgotten for a time. I started to compose music professionally.

Tehnologia group. Leonid is second from left.

I played in different groups - Bioconstructor, Tehnologia. I composed music for projects like Tehnologia, Comissar, N. Gulkina and Zvezdy group, Lada-Dance, Strelki and many others. And this was up to 2004. In 2004 for some reasons I had to leave the show business. I was looking for another occupation that would be interesting to me. And I came across Forex and Stocks, NeuroShell and MetaTrader 4. This is when my engineering education was of great help.

Please, tell us more about your trading experience.

First of all I decided to study trading in order not to make wild guesses. I attended various trading courses and took private lessons. I was a student of Veniamin Safin (ForexClub), Sergey Belyaev (Alpari). After that I started trading manually. After a year of such trading I understood that manual trading does not suit me, because it requires too much time, power and nerves, while success depends much on one's psychological state. So I started to think about automated trading.

In the internet I found a very interesting program NeuroShell Trader that allows getting market entering signals in the automated mode. Having tried it for half a year, having visited various forums about it and talked to the "native population" of these forums I found out that no one understood or really knew how to use it. Or if anyone knew, they were concealing it, which could hardly be. After that I made a conclusion that further appeared to be the only one correct: help can be got from nowhere but developers.

So I decided to learn everything from developers in order to understand how neural networks should be correctly used in trading. In Moscow State University, in the Neuroproject company I found courses in using neuronets in financial markets. There I got acquainted with Sergey Dolenco (Sergey Dolenco (senior research officer, PhD of physico-mathematical sciences, the official representative of NeuroShell in Russia, he took a direct part in the creation of NeuroShell2) who delivered me lectures in this topic, in use of neural networks, particularly NeuroShell in financial markets. After that I went to the USA to visit Steve Ward, the developer of NeuroShell series who also shared his experience with me and gave me a few lessons.

I should say this was of great help to me in the understanding of the subject and I completely changed my views that I had got communicating in various neuronet forums. It should be noted that use of neuronets in financial markets differs much from their use in other spheres. The fact is that we do not know future market behavior, so our task is to teach a network by the lessons of the past to forecast future. And not only forecast, but also make profit, which is a little different task. And not only to make profit, but to get a smoothly rising equity curve. Summing it all up, our purpose is to transform an unpredictable "chaotic" Close (or Open - it is not important) line into a quite predictable, smooth, upgoing Equity line. This task is non-trivial and its solution is not trivial. First I started to look for a Russian brokerage company where I could trade in the automatic mode. This is where I came across MetaTrader 4. Well, I liked this platform.

I started learning MQL4, but in some time I understood that within a short period of time it would be impossible for me to learn this language so that I would be able to create complex programs professionally. So, I decided to stay an amateur in one field and become a professional in another one. And again I referred to developers, this time developers of MetaTrader 4. They recommended me to visit MQL4 Forum. And in the forum I found many interesting programmers who then became my friends. As a result, I know MQL4 on the amateur level. But if a task is set and stated correctly, a good programmer will handle it quicker than I can and without many errors. I consider this quite successful for the creation of Expert Advisors. Moreover, by nature I am more a developer, not a programmer. Generally I think that a trader and a programmer are different jobs. They can be combined in one person, but this condition is not obligatory. The most important thing is a correctly set task.

Together with Roman Kramar (bstone) we made a professional binding MetaTrader 4-NeuroShell MTFeed, which fully synchronizes these two programs not only in ticks and bars, but also in time which is very important for some indicators. Besides, Yuriy Zaytsev (YuraZ), Viktor Nikolaev (Vinin), Igor Kolmogorov and again Roman Kramar (bstone) helped me to write Expert Advisors. I have several of them and I use them depending on market conditions and working symbols. So, with the help of programmers I translated my trading models from NeuroShell Trader to MQL4.

Generally, I think MQL4 is, on the one hand, quite easy and, on the other hand, very efficient for the creation of profitable Trading Systems. I am not a supporter of strong transformations of a source time raw of Close or Open, because at any signal transformations nonlinear distortions occur that may result in the loss of information value. And at some moment of time one more transformation a source signal can be completely lost - it will be unnoticeable against the background of such distortions which will naturally give a false picture of market to your Trading System. And MQL4 offers all the means for creating a profitable TS.

Let us talk about the Expert Advisor contesting in the Championship. It trades two currency pairs - EURUSD and GBPUSD. However, judging by comments to orders, trading is based on three strategies. Why is it so? How many trading strategies did you consider for the participation in the Championship?

Initially I planned to make three TSs based on three currency pairs - EURUSD, GBPUSD and USDCHF, to show the best of my Expert Advisor, so to say. When preparing to the Championship I communicated actively in the forum and had consultations with my friend trader Mikhail Valko, financial mathematician, vice-president of Interfintrade Group who also uses neural networks, specifically NeuroShell Trader as one of tools for trading in financial markets. Finally I have concluded that the Championship conditions are very harsh and if one wants to win one should use the most volatile pairs.

So, a Trading System on USDCHF was removed because of the low volatility and replaced by a TS on EURUSD. Well, a TS on EURUSD makes more profit than a TS on USDCHF. Actually, the main difficulty of the Championship for me is that an Expert Advisor should operate for 3 months with a growing Equity without changing of parameters. In live trading I change EA parameters from once a week to once for two months. It means I optimize (train) my strategy constantly, almost every day. Together with that I watch 10-20 TSs with various parameters on a real account. As soon as I understand that any of these TSs starts working better than the one that I use for trading, I change the parameters of my TS.

Fig. 1.

As for this EA, it should be noted that a TS on GBPUSD has not changed and I trade on a real account with the same parameters, I have not found better ones yet. You can see the operation of this system in Fig. 1. The green part (encircled) is OOS (Out Of Sample – outside the optimization interval). Though, TSs on EURUSD were changed a little. After the 23rd of September I found parameters for systems much better than those used in the contesting Expert Advisor. I must admit, these 2 TSs on EURUSD were not very successful at the beginning of the Championship - drawdown was larger than I had planned.

Fig. 2.

But this was difficult to predict before the 19th of September (time of EA submission), though now I understand it was possible. But as I have little experience in it (I hadn't earlier had the necessity of a long-term operation without reoptimization), I made a mistake. This was the influence of the increased volatility of markets, my EA started to trade too often. Besides I was wrong choosing StopLoss and TakeProfit - I had not expected such a high volatility. For the whole 2008 year SL or TP had never triggered. And in October this happened several times. Fig. 2 displays Trading Systems for EURUSD with new parameters. Encircled is the place when I understood that these TSs started working better than those submitted to the Championship. The green part is OOS. But I still hope that market will calm down, volatility will diminish and TSs on EURUSD will show their best.

All my TSs are turning ones, they are always in the market, always trading. TP and SL are placed only just in case. When they hit in the Championship, I had additional loss trades and more losses. This is connected with the fact that at the increased volatility TP levels were hit at the very bottom of falling; and as the falling was strong, my TS gave a signal to reopen in the same direction. A position was reopened and on the rollback I got StopLoss or at a turnover I had losses, though I should get just less profit, if TP were not hit and at a quite aggressive MM lot increased at reopening. And though in points it was less, in money terms it was more. This can be found in my statistics - not a high percentage of profitable short trades. This is all because of hit TakeProfit.

According to statistical report of the first month of the Championship, the currency pair EURUSD is the most popular among Participants. The second popular is GBPUSD. Why are they so popular, in your opinion?

In my opinion, all this is because of a certain stereotype and some passivity of traders. When one starts trading, one first tries EURUSD as the most popular pair. And then it is difficult to quit it because one gets used to it. Though, I think that one can trade successfully other pairs as well. That is why I wanted to make the third TS in my Expert Advisor for USDCHF. By the way, statistical data show that the most profitable pair is GBPJPY, and EURJPY is quite good.

Fig. 3.

Well, my first steps in the automated trading are connected with trading main cross-rates EURGBP, GBPJPY, EURJPY and others. Using neuronets and direct rates they consist of, one can make very profitable trading models. It is a known fact that crossrates are easier to predict than direct rates. In fig. 3 you can see one of them on GBPJPY, the green part is OOS. Further, searching for systems for automated trading in financial markets I developed a lot of working systems for the main most liquid securities. For some time I traded MICEX stocks, but in a certain moment the decreased volatility and increased number of gaps connected with a great number of aftermarker news made me return to Forex. My largest achievement in the stock market was when trading Sberbank stocks I entered a short position at the price of 110 rubles. This concurred with the beginning of the crisis and global fall of stocks. And I exited the position at 60 rubles. But exactly at that moment volatility decreased and market became of a very chaotic nature. At this moment my colleague Mikhail Valko and me are developing the automated trading of derivatives on European Exchanges.

You also prefer EURUSD. Why are two trading strategies working on EURUSD and one on GBPUSD? What is it connected with?

When selecting a strategy I conducted a very deep analysis of various econometric methods of estimating the future behavior of trading models based on preceding results. When building a system of decision making based on one or several stochastic modes it was necessary to find a balance between the complexity and effectiveness of the system. Initially I planned trading three pairs: USDCHF, EURUSD and GBPUSD. Testing showed that trading models on USDCHF gave objectively lower equity increase because of the volatility lower than of euro and pound. Out of all models on GBPUSD only two were left - those that could give substantial equity increase.

Placing two almost identical trading models on GBPUSD was not rational because I wanted to avoid double drawdown and maximally diversify the profile. On EURUSD I found two models with various parameters that successfully hedged each other and besides hedged well with GBPUSD. The purpose of building a profile from three TSs was having minimal drawdown with a smooth growth of equity. That is why, having chosen exactly these three systems and analyzed their combined behavior I concluded that I should use them. And you see, I was right - hedging helped me very much.

Your Expert Advisor executed 28 trades on GBPUSD and made 2205 points of profit not accounting for swaps (mathematical expectation = 78.8 points). It executed 105 trades on EURUSD and earned 2703 points not accounting for swaps (mathematical expectation = 26.03 points). And on EURUSD "Third TS" (as written in order comments) executed 61 trades with the total profit of 1411 points (ME = 23.1 points), while the "First TS" strategy closed 44 positions with the profit of 1322 points (ME = 30.1). I.e. we see that all the three strategies are profitable on both currency pairs! How did you manage to achieve this? Were results the same during the preliminary testing?

When planning the future efficiency of the system, choosing optimal parameters for risk management, the most important thing was to make a correct forecast of the week volatility of three currency pairs. Together with Mikhail Valko we calculated the most probable future daily and weekly volatility and dynamics of the major currency pairs based on advanced models of the quantitative and econometric analysis. Having formed such forecasting, having conducted simulation of the behavior of the trading systems profile in such markets and stress tests, and having estimated the most probable distributions of profit, we selected optimal parameters for risk management and once again approved the profile structure. In fig. 4 (EURUSD) and 5 (GBPUSD) you can see the analysis of a strategy profitability using means of MatLab.

Fig. 4.

Fig. 5.

When estimating the future effectiveness of systems we used latest achievements on this sphere. The complex analysis was conducted using proprietary risk management systems developed in Matlab, SPlus and Eviews. If not for the TP hitting, results would be even better. These are pictures of TS Equity for EURUSD and GBPUSD in MetaTrader 4 before the Championship. They show Equity in the period of optimization and OOS. Fig. 6 - EURUSD, 7 - GBPUSD.

Fig.6.

Fig. 7.

The most important thing is to have in the period of optimization and OOS maximal profit and smooth growth of equity without sharp rises and falls. Both with fixed lot and with MM. Generally I think that the optimization is an art, on which success in autotrading depends; as well as the quantitative estimation of future profit probability based on results of previous trading. Fig. 8 shows an almost ideal performance.

Fig. 8.

For the sake of security you have introduced levels SL/ TP. Do parameters of SL and TP placing result from the EA optimization in MetaTrader 4 tester? Or are there any other reasons?

Yes, they are to some extent results of optimization, they were chosen maximal in order not to hinder Trading Strategies. But at the same time they should perform the protective function for MM so that loss does not exceed certain values in various emergency cases.

The distance of SL and TP values are absolutely equal in points (for EURUSD SL = TP = 300 points, for GBPUSD SL = TP = 500 points). Does it mean that the EA strategies are not trend-following, that it does not matter where the price will go after opening?

That is true, TSs that I use are neither trend nor flat ones. Though, I should admit that they may experience some difficulties in flat - unexpected drawdown can occur. Everything depends on the kind of flat, you know they can be different. TSs go in the direction where the market goes. Actually, I use regularities that are so numerous in the Forex market. They can be found both using neural networks and without them. These regularities are not evident, i.e. they are hard to see or detect by analyses. Actually they are probabilities where a price can go - up or down in this market situation.

It should be mentioned that the Championship is conducted in the severe conditions of a crisis when volatility of all currency increased many times and we see unprecedented growth of the US dollar. That is why various Trading Strategies are tested in extreme conditions. For some of them higher volatility is favorable, i.e. they have much higher profit than they would have in a usual market. For other TSs on the contrary - higher volatility results in false market entering signals and they get much larger loss than in a non-crisis market. This can be seen in my Expert Advisor - the two TSs trading EURUSD. One of them is very sensitive to the market and higher volatility increased the number of trades - a lot of false market entering signals appear which increases the drawdown of this TS. While the second TS does not have such problems.

Your Expert Advisor executed 28 trades on GBPUSD and made 2205 points of profit not accounting for swaps (mathematical expectation = 78.8 points). It executed 105 trades on EURUSD and earned 2703 points not accounting for swaps (mathematical expectation = 26.03 points). On the other hand, analyzing GBPUSD trades of your EA on history, we can see that one short position is closed and another one is opened. And this can happen several times in succession. This is an unusual strategy. Could you please explain how it happens?

This happens when TP is hit. The TS tactic is simple: if TP is hit, the TS waits for the next signal. And if the signal is in the same direction, a position is opened in the same direction. The only moment is that at TP triggering a rollback may occur and a trade can be closed with minus at a turnover. And if the aggressive MM is used, the loss can be larger than profit. So, actually TS is not designed for TP hitting - I described this situation earlier.

Now let us talk about neural networks. What types of neuronets do you use? How are neuronets trained?

There is an opinion that all types of neural networks can work in financial markets. But actually it is far from true. Financial markets are very noisy and unsteady. That is why not all neuronets can correctly and adequately perceive these data. Depending on the currency pair or stocks, I like to use Adaptive Probabilistic Neural Network, Recurrent network, Jump Connection network and some other networks. As for the training process - the answer is difficult, actually there is no a unique answer. The correctness of training depends on many parameters.

As a neural network is a strongest non-linear transform, if training parameters are selected incorrectly, the so called network over-training occurs. It means, when a network learns market rules on history very well, this hinders its profitable operation in future. One of the main parameters is the period of training. If it is too small, the network retraining will happen too quickly, and if the period is too large, the network will not find necessary regularities. That is why one should be very attentive and careful choosing the training period.

The second parameter is entries. This is one of the most important parameters. If entries do not give necessary information to a network, its training is senseless. There is an opinion that the more entries, the quicker retraining occurs, and if entries are not informative, a network will not be trained.

And the third parameter is the genetic optimizer. Its use does not always give a wanted result. And finding a correct result depends on the parameters changing range and step of each variable searching. If the search range is too large, a genetic optimizer can fail to find a necessary value, i.e. it will enter a local minimum that it will not then escape. This is one of the variants of retraining. Generally, there are more questions than correct answers.

Did you first train your neural network to find optimal position entering signals and then added search for exit strategies or did it happen in another way?

I like turnover strategies. I.e. entry is an exit from a previous entry. I think if favorable market entries are found, for example buy, they will be favorable exits from the previous sell, and vice versa. Well, favorable entries are not numerous. That is why I train my network to search for entries only. Usually I use one neuronet for long entries and another one for short entries. But sometimes I use one for both. By the way, I can develop a Trading Strategy without neural networks. And it will work. A neural network is not a grail, not a panacea. This is gust a tool, and one must know how to use it. There are a lot of other tools that work successfully in financial market not worse, or maybe even better than neuronets. Well, generally the task of making a profitable trading strategy in the Forex market, of finding market regularities is solved for me. Not fully solved is the question of choosing the most optimal TS parameters for future trading (best profitability) after its optimization.

Last year the winner of the Championship was Better with an Expert Advisor based on neural networks. Were his ideas interesting to you? Did you learn them?

Of course, they were very interesting to me. Generally, the whole Championship statistics are interesting and give a lot of useful information. As there are no statistics of profitable Expert Advisors and strategies at all, one cannot consult or get a correct advice, statistics that you collect in the Championships is very helpful and informative. And it gave me much, it helped me understand that I am moving in the right direction. Though I didn't participate in the previous-year Championship, I was following it very attentively reading all interviews and articles.

Who, in your opinion, has good chances this year? Do you think multi-currency Expert Advisors have higher chances to win the contest?

Multi-currency EAs have one large advantage - if strategies are chosen correctly and the MM is optimized, hedging can be of great help to an Expert Advisor. This can be seen on the example of my EA in this Championship. When TSs on EURUSD started working not so well, the TS on GBPUSD saved the Expert Advisor; and when all the three TSs started working, equity grew up. As for contesting Expert Advisor, I like those of Cronex and Liliput for their advanced algorithms. The Participant YuraZ has a good EA based on divergences, with which I worked some time and which I know very well, but it seems he overcomplicated with optimization. The Participant Pushistik had a good rise, but, in my opinion, she made a mistake with MM. If talking about Gorez, it is hard to judge about an EA without long positions. This is like a cut EA. Of course, we can explain it by the fact that a strong downtrend was predicted that we see now on GBPJPY, but such trends are very rare, that is why it is unreasonable to be governed only by them, because it is not clear how it will behave in flat when both buy and sell will work. The more so, there were places for long positions, but they were successfully missed.

I cannot tell anything about the group of Expert Advisors called scalpers, or pipsers who are in the top ten (strelec, Prizmal and abeiks), because this is not my style of working. I think such method of trading is not always correct, because not all brokerage companies agree to it and pay profit from such a method. Though, it seems they are ok by the Championship rules, but the question of real accounts is open for them. On the other hand, having a normal (non-pipsing) TS one can use near targets (4-5 points) as a hedging TS. And in the conditions of the Championship, with high market volatility, they have much higher chances to win, as trading conditions have not changed. Well, there are a lot of interesting Expert Advisors, I cannot recall all of them. In my opinion, in the level of Expert Advisors, this Championship excels the previous one.

Thank you, Leonid, for the interesting interview. We wish you good luck!