First peek at the simulated trading results in the simultaneous multi-market mode.

User Rating: / 0
PoorBest 
Written by Forex Automaton   
Monday, 03 August 2009 18:04

The first peek at the simulated trading system performance results, obtained in the multi-market mode of operation, reveals some exciting early results but also highlights the amount of work ahead. These results are obtained with the forecasting algorithm handling the different markets jointly, attempting to learn and use cross-market patterns. The main question is: what do we gain from the combined multi-market analysis? Theoretically, the benefits are to be expected in the area of pattern recognition (where new classes of patterns, those involving different markets, may appear and be detected) as well as in the area of money management (cf. asset diversification along the lines of the modern portfolio theory). We are not ready to address this question, as it turns out that having switched to the combination of the markets, we have landed in the trading regime which is different from that seen for the nominally equivalent parameter values within any single market. The next task will be to sort out the trivial and not-so-trivial differences.

1. The basics

The summary report on the single-market trading system optimization, posted on July 21, 2009, ended with the words:

"For the immediate future, the plan is to switch to the optimization of the multi-market version of the system. This will proceed in two distinct steps, Step 2 and Step 3 respectively (this one being Step 1).

In Step 2, the money management component of the algorithm will become aware of the concurrent trade ideas in the different markets and will learn to treat them properly. The markets themselves at this point will be forecast independently, with the "consensus" parameter settings just obtained. The thing to check is the reduced "risk" or fluctuation in returns. It is expected to be reduced since the concurrent trades are not 100% correlated.

In Step 3, the forecasting algorithm itself will begin to handle the different markets jointly, effectively learning to utilize cross-market patterns. This is a step into uncharted territory -- not because nobody has done it, but because those who have done it are understandably reluctant to share the charts. The result of the Step 2 alone may already constitute a useful product."

 

A technical detail of note is the fact that transition from Step 1 to Step 3 requires no new code development (however strange this may sound), whereas Step 2 does require new code. The code for Step 2 is currently being written. Meanwhile, the Step 1/3 code has been used in our usual Monte Carlo back-testing simulations, this time for the multiple markets: EUR/USD, USD/JPY, GBP/USD, USD/CAD, USD/CHF, AUD/USD.

The basic framework remains the same: a run of the program included simulations of trading histories of 2592 independent "virtual traders" (forex robots), each of them being an incarnation of the same algorithm, differing by the setting of the adjustable knobs. As in all other studies published so far, the trading is performed on one-decision-a-day time scale, with 1:100 leverage, risking no more than 10% of the trading capital at any point in time. This report uses the day scale forex data covering the time interval from August 20, 2002 to March 23, 2009, with the actual trading starting in March 2006 (when the initial "training" of the system was completed). The key concepts of conditional projection distributions and profile histograms have been explained before. The trading system control parameters remain as previously defined.

 

2. "Minimum bias" look at the data

The term "minimum bias", borrowed from the vocabulary of physicists hunting "the God particle" and the like, denotes (in the original context) events that pass the minimum criteria of usefulness and purity (no garbage), but otherwise embody minimum effort to enrich the data with the objects one is hunting. In our case, the object being hunted is the successful trading algorithm.

Effect of stop-loss on the trading returns.

Fig.2.1. A profile histogram showing dependence of the annualized return (measured directly by comparing equity at the beginning and end points of trading) on the stop-loss placement. The unit of return is 100% (100%=1).

Effect of the entry parameter on the trading returns.

Fig.2.2. A profile histogram showing dependence of the annualized return (measured directly by comparing equity at the beginning and end points of trading) on the trade entry parameter. Larger values of the parameter correspond to more "conservatism" in the selection of trade ideas to execute. The unit of return is 100% (100%=1).

Effect of the forecasting parameter on the trading returns.

Fig.2.3. A profile histogram showing dependence of the annualized return (measured directly by comparing equity at the beginning and end points of trading) on the Fred parameter that controls the forecasting. The unit of return is 100% (100%=1).

Effect of the stop-loss parameter on the mean duration of a trading position.

Fig.2.4. A profile histogram showing dependence of the mean duration of a trading position (in days) on the stop-loss parameter.

Given that the figures represent minimum bias data, one should not be too discouraged to see the data to average themselves to yield negative returns, as they do in the figures.

The Fred distribution (or, more rigorously speaking, distribution of annualized return on Fred), Fig.2.3, looks much more dramatic now, with the peak of high returns relatively sharper. Curiously, the main multi-market peak is roughly at the same place where we placed it in the course of the "consensus optimum" study for the single (isolated) markets.

Fig.2.4 should be compared with Fig.2.1 of the Stop-loss section of the single-market summary report. Such a comparison indicates that despite the similarity of input, the life time of a trade is considerably reduced in the multi-market mode of operation. Why this is so remains to be understood. The first guess is that some positions end up being reduced because of competitive trade ideas from the other markets, with better return/risk profile. Whether this process is optimal, remains to be understood. Currently this is done the way described in the section Portfolio Allocation of the Setting up an algo trading system test bench post.

3. Approaching the multivariate optimum

The following figures are, like the previous ones, profile histograms, projecting the distribution onto a two-dimensional plane of variables whose inter-dependence is being investigated. The difference is what subset of the distribution is now included: the subset is limited to what is believed to be the neighbourhood of the optimum. Technically, the neighbourhood is defined by applying cuts to the algorithmic parameters, some of them or all but one -- the one being selected for the horizontal axis of the plot. The vertical axis then presents the figure of merit under study, typically some measure of return or risk.

Selecting good ranges for the stop-loss and Fred parameters is straightforward on the basis of the previous section. These are used below to clarify the dependences in the vicinity of the global optimum.

Effect of the entry parameter on the trading returns.

Fig.3.1. A profile histogram showing dependence of the annualized return (measured directly by comparing equity at the beginning and end points of trading) on the trade entry parameter. Larger values of the parameter correspond to more "conservatism" in the selection of trade ideas to execute. The unit of return is 100% (100%=1).

Fig.3.1 leads to very different conclusions than Fig.2.2: it's now clear that positive returns are possible and we are far from having reached high enough value of ENTRY: there is no indication of saturation or maximum of return being reached as ENTRY is being increased.

Effect of the entry parameter on the trading returns.

Fig.3.2. A profile histogram showing dependence of the annualized return (measured directly by comparing equity at the beginning and end points of trading) on the trade exit parameter. Larger values of the parameter correspond to the tendency of "getting married to a trade". The unit of return is 100% (100%=1).

The exit parameter, on the contrary, seems to matter very little in the vicinity of the global optimum.

Effect of the entry parameter on the trading returns.

Fig.3.3. A profile histogram showing dependence of the annualized return (measured directly by comparing equity at the beginning and end points of trading) on the forecasting parameter Fred. The unit of return is 100% (100%=1).

The ENTRY, EXITT and STOPLOSS selection change the shape of the Fred dependence of returns very little, their main effect is the overall upward shift of the returns: compare Fig.3.3 with Fig.2.3.

4. The winners

Annualized return distribution, forecasting parameter selection on and off

Fig.4.1. Distributions of the annualized return in the course of the trading system parameter optimization. Black histograms: all parameters except for Fred are fixed at the best ranges, no selection on Fred. Yellow: same as the black in the respective panel, with an extra condition on Fred. Normalization is arbitrary. The unit of return is 100%.

Figure 4.1 follows the format that is likely to be familiar from the earlier reports in this series. Due to the sharpness of the Fred peak, seen in Fig.3.3, the Fred selection gives a huge boost to the annualized return: there is no doubts regarding clear inconsistency of yellow and black distributions. A not-so-good news is that Fred seems crucial now: the black distribution, the one accumulating the money management wisdom, is clearly not the one to work with. There is a simple explanation: as Fig.3.1 indicates, we are far away from the best money management scenario -- the optimum is likely to be located at much higher values of the trade entry threshold. These would correspond to more "conservatism" in the seleciton of trade ideas. The thing to do is therefore to continue simulations, extending the trade entry range.

Bookmark with:

Deli.cio.us    Digg    reddit    Facebook    StumbleUpon    Newsvine
Last Updated ( Monday, 04 January 2010 12:30 )