Click on the above image to get a larger picture (1600x1200)
[Printed using Portfolio Maximizer by TradeStation Technologies and RINA systems ]
DMI(5) difference results. No fuzzy logic. 10 min 1998 SP500 data
| Ideas have
floated around that the classical indicators like RSI, %K,%D stochastics, DMI ...
COULD be nothing different from poor tools for the average trader, and that, if they work,
it's only by chance. (so, they were nicknamed "Gambling Indicators"). This statement is certainly true at a first glance, but it has nothing to do with the indicators themselves, rather than with the way they are commonly used. Most
of classical indicators have been developed with a simple idea in mind: They
should show to the trader a better reading of price market action, and therefore allow
writing simple trading rules because writing trading rules is easier with an easy to read
indicator. |
System
development with Gambling Indicators: The basis remains the same: one or two simple indicators and a few simple rules are the rule of thumb of most trading systems. The reason is simple: A programmer cannot easily devise a more complicated system without spending a great deal of time on it, and generally speaking, the results produced when testing on unseen data is often disappointing. After some months or even years of research, anyone in this respect should be disappointing, or at least have lost most of the enthusiasm of the early days. A fix has been suggested to this situation: Hill climbing optimization: As we cannot easily adapt the rules' complexity to the market complexity, let's adapt the indicators to the rules set ... This is the first big mistake to the robust system development. At a minimum it can show if there is a stability plateau in the indicator's parameters set for the considered set of rules. But what if the rules were not the best suited? No valid answer given here because we were not able to find the best rules (otherwise, no need to optimize the indicators!) These do not solve anything concerning the knowledge supposed to be stored in the system code, then no guarantee could be given regarding its future behavior. |
| With the arrival of
modern computers, it has been easier to develop Neural Networks algorithms. Neural Networks are powerful pattern recognition and prediction tools, and the idea that quickly appeared was Our brain has neurons / It can learn / some traders succeeded in learning to trade Neural Nets have artificial neurons / they can learn./ They have been proven to be useful in many domains. ==> A Neural Network fed with all the necessary indicators will learn how to trade the market, et voila. Here is the second mistake! Experience has shown that if it had been not impossible, the results would not often have been at the level expected by the average AI computerized trader. As a proof of this, suffice to see the number of ads concerning Neural Networks applied to trading in the early 1990 and the very few now available to traders. In the meantime, Neural Nets technologies have been enhanced, computer speed now available is 20 times or more what it was in the early 90, cheaper, but the results concerning NN trading are still somewhat sparse. back to top |
For ending, during these last 3 years, most
researchers agreed :"Now , it's time to go back to the roots and simple
proven things". An answer in itself to the partial failure of general AI techniques in the trading area using indicators No magic in it. This does not mean that it's impossible to succeed in Neural Nets.
Neural Networks are better in pattern recognition from raw market
data that in dealing with indicators, that are themselves an interpretation
of the raw data. The problem comes from the fact that the Neural
Network is not very well suited to understand indicators like the trader can.
Their kind of intelligence (is this word well suited ?) is not the same than the human
one. At least it's far more mechanical and less subtle, or exactly, less intuitive.
|
| We are not able to
deal with complex relationships, and the Neural Networks are, but their
"thinking" is not similar to ours:A Neural Network will find complex
relationships by determining the correct weights that figure artificial synapses, and the
control of their internal logic is beyond the scope of the average Neural Network user. A Neural Network can find very good trading systems on training data, less on unseen data, and the reason is simple: Their learning is made by trial and error, without any control of the logic that is produced to combine the indicators. Their goal is often to minimize the prediction error, regardless of the method used to develop the logic. Better is to make it learn from an evaluation function (like net profit). You can also use alternate methods for training (Genetic Algorithms, Simulated Annealing...), but the problem remains the same: A Neural Network will have the internal structure that fit its needs: To learn at any price the training set. A Neural Network is able to discard some useless inputs (indicator), but you will never know why it has been discarded (in fact because the AI predictor did not use it) . back to top |
However, when feeding the Neural Network with this
indicator, you're were considering it useful. This is somewhat disappointing, and finally, we always displace the problem to another one. I only spoke about NN because they are widely known in this domain,
but I could have done the same with other techniques like machine learning rules'
bases, genetic scorecards, genetically evolved Neural Networks, genetic rule bases. Indicators are devised by human who project their perception of the phenomenon on the way they devise the indicator. So what could better use a human way of thinking? An Artificial Intelligence structure that mimics close the human way. This technique exists and is known as "fuzzy logic". |
Technical
analysis is Because the perfect rules that may drive any
market are not known (and never will be as for physical and chemistry) the task of the
technical analyst is to find an interpretative way to discover the beginning and the end
of a trend as soon as it seems to appear. |
Any technical analyst that tried a system based
on RSI and that tried to optimize the above levels have been told those sometimes 30/70
level works, sometimes 25/65, sometimes other values. More, for a given value set, it’s easy to see that a microscopic variation of the RSI level above or below the OB/OS level would have produced a better result. In fact, too much precision when using an indicator compared to a given level could bring very unstable results. Above frontiers (30 and 70) are defining crisp values of the RSI domains. 30 is 30, not 30.1. Fuzzyfication of the inputs allow saying that " around" 30 is the Oversold frontier, and "around" 70 the overbought. In the overlapping zone (that can be seen as centered on 30 or 70), the RSI will be seen as "partially neutral and partially oversold" when around 30, "Partially neutral and partially overbought when around 70. The word partially is expressed by the membership functions used in fuzzy logic, but I want to remain simple. None will say that if RSI=70.1, the market oversold and neutral if RSI =69.9. Nonsense certainly? Yes. |
| What would you think
otherwise? That the first thing to do is to see how the trend is ( flat, moderated up,
uptrending, and so on) before declaring the RSI status. Do you know what you are now doing? You are doing fuzzy logic. Your logic is fuzzy in essence, as technical analysis is, as human thinking is. You just spoke about trend direction (up or down) in terms that were fuzzy. You spoke on that to qualify the RSI status that you feel also fuzzy. Your brain has built intuitive assumptions between the RSI and the trend strength that are both fuzzy. You are doing fuzzy logic without knowing that you were doing it, and attempted to build fuzzy rules to better understand the RSI signal. Fuzzy Logic deals with this apparent uncertainty in a very mathematical way of solving the problem of interpreting the indicators. In this sense, the name is misleading. Fuzzy Logic deals with fuzzy sets, but the math involved in it are anything you want but not fuzzy in essence. back to top |
Now, you should better understand our idea of solving
the difficult problem of Technical Analysis trading systems: We need to build fuzzy rules to interpret the indicators. The rules will be difficult to design to build and to test as they can be numerous. More we are unable to do this task by hand programming. Computerized techniques can be devised to build the fuzzy rules. We do not want to waste our time in things that are beyond our own patience, so the computer must do all the tedious work for us (building and testing the fuzzy system) Neural Networks are powerful tools, but their use is not well suited to interpret the raw indicators. We will use them to enhance the fuzzy rules with some neural network power, that will add neural knowledge to the fuzzy system: This association is called neurofuzzy logic. It has been devised to minimize the drawback of both methods (fuzzy logic and neural networks). ... And it's a very powerful one as we will now demonstrate by the example: |
| We offer as an
example to ONLY use TWO widely
known indicators as inputs to the neurofuzzy logic trading system:
|
As we are not so stupid, when choosing the indicators
to be use as inputs to the neurofuzzy system, we keep in mind a logical idea that is The DMI difference gives the buy and sell answer (delayed), but we want a subtler interpretation by using the ADX information. The system should be able to build different rules according to the ADX status.At this moment, we do not know how many rules will be necessary, but again, we don't care: The task is devoted to the AI software. We only help it by feeding it with information that seems sounding relevant to us. A closer examination of the indicators shows that two pieces of important information are lacking: We must tell the system if the ADX is rising or falling, and if the DMI difference is rising or falling too. |
So, we will use FOUR inputs to the system:
ADX(14)
DMI(5) difference
ADX(14) -ADX(14)[1] (first derivative)
DMI(5) difference - DMI(5) difference [1] (first derivative).
| Now, we need to know
what the lookback period for DMI and ADX calculation is. Here again, a little logical thinking is necessary: We will feed the systems with the ideal buy sell signals. There is dozen of possibilities of defining what the interesting swing amplitude is to make learn to the system. It's a matter of choice that depends on the trade frequency expected the average trade. As this is calculated with a template indicator in TradeStation, we only have to choose lookback periods whose information is consistent with the major moves that we intend to learn. For example, a DMI(50) difference will certainly be the best suited value if the average duration of the moves is for example 5 bars. DMI( 5) or a little more is certainly better. Same apply with the ADX. As our purpose is only to help the software by using a useful indicator, we do not try to optimize the lookback periods of the considered indicators. We only consider them measurement tools that have their own pros and cons, but this value will remain then fixed. back to top |
To make things clear, we apply a well-known
measurement principle used in physics: The measurement instrument there the indicator)
must be adapted to the magnitude of what you want to describe with it (here the trend legs
of the ideal buy sell sequences). To remain more "down Earth", if you want to measure the surface of a room, you may do it with several tools: A surveyor's chain, a 5 m measuring tape, a micrometer caliper... My guess is that you will choose a measuring tape, that is the best suited too, unless the room has the size of a stadium or of a match box. The same principle applies to technical indicators, if you want a correct result. Here we will remain with the default 14 bars lookback period for the ADX,but 5 bars for the DMI difference, only to take advantage of short terms movements. All
of the tests below will be performed on the SP500
futures 10 minutes bars. |
| Before starting, lets
do a comparison under the same conditions I have written this very simple example that does not make use of the ADX(14), but only the DMI(5) difference: TradeStation code is as follows: inputs:dm(5); |
But stop and sell stops have been
added to filter out. It's like adding a small channel breakout to the DMI.difference setup
signal. Let's admit that with a huge research, we could have been able to devise an Easy Language code able to do 2 or 3 times better than this one (but with the same indicators and no optimization of the lookback period). |
Result are below (TradeStation 4 Performance Summary):
DMI SP_9698.2-10 min 03/02/98 - 09/03/98
Performance Summary: All Trades
Total net profit $ 24450.00
Open position P/L $ 250.00
Gross profit $ 473600.00 Gross loss
$-449150.00
Total # of trades 450 Percent profitable
33%
Number winning trades 147 Number losing trades
303
Largest winning trade $ 19850.00 Largest losing trade
$ -6700.00
Average winning trade $ 3221.77 Average losing trade
$ -1482.34
Ratio avg win/avg loss 2.17 Avg trade(win &
loss) $ 54.33
Max consec. winners 4 Max consec. losers
10
Avg # bars in winners 22 Avg # bars in losers
7
Max intraday drawdown
$ -32200.00
Profit factor 1.05 Max #
contracts held 1
Account size required $ 43200.00 Return on account
57%
back to top
Let's have a quick look at the Equity curve produced
Click on the above image to get a larger
picture (1600x1200)
[Printed using Portfolio Maximizer by TradeStation Technologies
and RINA systems ]
DMI(5) difference results. No fuzzy logic. 10 min 1998 SP500 data
Somewhat seriously jagged, but climbing however
Note the serious Max drawdown (higher than net profit),
the low average trade.
Operating mode:
We just need to build an ASCII file in TradeStation with the four
indicators above,
Open the Assistant SAFIR-X software,
load the ASCII file data
and click on the EXPLORE button.
Now , it's time to go drinking some coffee, and be back in less than 15 minutes...if you
are using a Pentium II 300Mhz.
Here is what appears on your screen when you are back:
Click on the above image to get a larger picture
(1600x1200)
Result of DMI(5) ADX(14) on SP500 980302-980903 10 min
data
back to top
Detail's explanations of the various above
windows are summarized below.
| ALL available solutions tree. Here Solution #11was selected (1) |
All
solution results (tied to the tree). Recallable.Sortable by columns (2) |
Costs, settings... Detailed Performance summary (5) |
||
| Details on the neurofuzzy rules for the considered selected solution (#11)Sortable by columns (3) | Graphic
display of the fuzzy sets for the considered highlighted rule (#227 among 405). Active fuzzy sets for this rule are green (4) |
|||
Journal file. (8) Shows here the Target Ideal Performance Summary (displayed here, as well as ALL the operations done during the EXPLORE automated search) |
Detailed Equity Curve and trading signals (7) |
|||
| recall
previous solution (graphic) zoom data window |
TRAIN Red=sell Blue=buy overlaid on raw train data Equity=Yellow line |
TEST Red=sell Blue=buy overlaid on raw test data Equity=White line |
||
Indicators display |
||||
The Assistant Version2 (Windows explanations)
| The system has been
trained on the first half of the data file (the same SP500 data than before) and tested on
the second half (that is the test data, unseen to the rules when learning). Results are far different than what we get with the simple reference DMI5 system: $ 125590.00 net profit on test data instead of only $ 24 445.00 for all the 1998 file Average trade $ 1773.94 on test data instead of only $ 54.33 We do not even want to compare the reference DMI system to the whole data (238350.00) that contains Train data already seen by the system ... This result was achieved only with TWO classical indicators, but exploited to the maximum possible of precision. Would you have been able to write a trading system featuring 169 (fuzzy) rules in less than 10 minutes and to test it ? back to top |
Certainly not. 15 days , maybe, but I bet that most of
the trading system developer (including me first) would give up far before. By observing the results of all the solution stored in the window (2), you may see that most of the 10 retained solution were quite good. We must notice that the uptrend has
not been correctly understood by the system, so we will consider this fact when later
improving the basic system |
Click on the above image to get a larger
picture (1600x1200)
Examining the retained solutions (DMI5 and ADX14).
We have move the detailed view (7) to
the left. Now this window the features solution #8,
and we can see the original Equity curve of the best solution (#9),
still displayed below the Performance Summary for an easy comparison
Would you like to see the Performance Summary for solution #8?
Easy to do: Only double click on this solution is the Solution tree window (1).
I am not showing this now, because we have better things to do.
The Performance summary for any solution is also stored in the Journal File (8)
back to top
Now, it's time to do...
Let's take THE SP500 CONTRACT with the same 10 min. bars and of
course the same DMI(5) difference and ADX(14).
Testing on unseen data is even easier:
You only need a new ASCII file (similar to the file containing 1998 data)
that you will load into the assistant (here the SP97 TSD ASCII file).
Click on CHECK
Then it will test not only on the current solution (#9) but all the available 10
stored solutions.
Doing so, you can easily see if another solution had been better.
Here are the results for one of them:
Click on the above image to get a larger
picture (1600x1200)
Result on SP500 961030-970610 10 min unseen data (DMI5 and ADX14).
Conclusion:
| The above result shows that
the trivial DMI(5) / ADX(14) the neurofuzzy system can work on unseen data with
quite acceptable results.This is not still the case with such simple systems
because some important information is being missing (see for example the first half of the above equity curve: It' s obvious that the system has not well understood the major uptrend of the first half of SMP7). The explanation is simple: Two indicators is maybe somewhat dangerous to claim that they will explain most of the possible moves. However, you can see that we have pushed the limits to the breaking point: Few indicators and biased uptrend data during training ... for an apparently acceptable result. back to top |
We want better now. The system is however too simple to get consistent results. The twin indicator system was used here for demonstration purposes and to accept that there are better ways to find a solution even with a few indicators like these one. Now we want a robust system able to work with more confidence. Miracles do not exist in trading systems, and we may devise a better one by using some others additional indicators to enhance the overall stability and performance. Now; it's time to do a more serious work. See the next step coming... |
Frankly, we cannot succeed in any good accuracy by using only two
indicators.
The market is more complex than this and cannot be reduced so easily to four numbers
(including the first differences).
Let's go back
to our original idea, We need to measure the
trend quality that was devoted to the ADX(14).indicator. We also need to have an idea of the trend direction (ADX
gives no clue to this), and we used the DMI(5) difference, that does a short term tend
measurement. Due to the fact that our lookback period was maybe too short ( and
more, we have increased the ADX lookback) we will chose the following indicators: |
We do not want to increase too much the DMI lookback,
because we need an information on the short term and trading range. As we have no overbought oversold oscillator, DMIplus(7) and DMIminus(7) will carry the short term and trading range information. We previous used the first difference for ADX and DMI difference. One must keep the ADX first difference , because it gives invaluable information in case of Trading to Trend or Trend to Trading phase. To decrease the sensitivity, lets take a 3 bar momentum of the new ADX. Finally, we discard the DMI first difference (DMIplus and DMIminus may bring a similar information, and we want to stay with the same number of inputs (4 in this case) for a valid comparison. We only discussed how to improve the indicators according to the weaknesses observed with our basic model. To summary, |
| As
a rule of thumb, help the neurofuzzy system with logical indicators, then it will help you
to the best that it can. Avoid redundancy otherwise you will confuse it as it would
confuse you. Treat it like if it had been a technical analyst ready to code a system to which you provide a basic idea and the indicators that you feel meaningful to solve. That's all, and that's enough. To remain simple and to avoid longer training, we will remove the previous first differences that are important too. But we have here two systems to compare, with the same number of inputs (e.g., 4). In the second case, we have pushed the thought a little bit more in depth to choose the neurofuzzy inputs. As a hint, and to determine if the indicators may be correct, we recommend to plot them on the same chart where the ideal buy sell signal is also displayed. back to top |
By eyeballing, you must verify that most of the considered zig zag moves are reflected by
an obvious change of at last one indicator. See the TradeStation chart below that was used to compute the ASCII file in use with the Safir-X Assistant:It should be obvious to anyone that the small moves are producing DMI's swings, some saturation of the DMI's., big moves show a rising ADX A general noise seems to appear on indicators. It's noise, but not only noise: It's the kind of information that the neurofuzzy logic system will exploit better than any human being. As we will use the first half of the file, you may notice that we have chosen a better training period showing uptrend and more downtrend, unlike in our first example. |
Click on the above image to get a larger
picture (1600x1200)
TradeStation Chart on SP500 10 min data (970313 to 970606)
Indicators : ADX(25), DMIplus(7), DMIminus(7), ADX(25) -ADX(25)[3] .Zig
Zag ideal moves (Cyan)
The same process than with the previous system:
:After loading the SP500 ASCII file (10 min data from 961021 to 970606)
and launching the EXPLORE command and drinking another cup of coffee, the results
are here:
Click on the above image to get a larger picture (1600x1200)
Training and testing on SP500 10 min data (961021 to 970606)
Indicators : ADX(25),DMIplus(7), DMIminus(7),ADX(25) -ADX(25)[3] .
A quick examination of the equity curve shows that the system is
correct on test data
($36100, e.g., a 137 % ROMID in a 2 months period with a $ 270 average trade, not so
bad).
The Equity Curve continues to climb (white line) on the test data (second half of
the file).
Several solutions were found profitable #10,11,19,24...
(best one is #17,displayed here, but remember that they contain training data).
back to top
Let's seen how it is now behaving on unseen data :
We will do more than testing on SP500 data. We chose to test on a different
market!
You only have to load the ASCII file related to this period.
Then click on CHECK.
The entire solution stored are evaluated using these unseen data
(wait 2 minutes, in real world):
Click on the above image to get a larger
picture (1600x1200)
Testing on CAC 10 min unseen data (980101 to 980903)
Indicators : ADX(25),DMIplus(7), DMIminus(7),ADX(25) -ADX(25)[3] .
The best
solution is not #17 (normal, see above), but #3, followed by
#14,24,16,.... Now, you may want to examine the behavior of the other solutions on
the unseen data. |
The Performance summary can be stored in the Journal
File for all the solutions. You can even test now using a different Performance Summary field (Highlighted in red, window 5). Solution #19 is displayed for comparison in the details window. |
Same process.
We still have the confirmation that solution #19 was not the best one (see Details)
Solution #3 is also in a good place, but solution #24 and
#14 too.
This is the confirmation that several of our solutions are able to generalize.
Click on the above image to get a larger
picture (1600x1200)
Testing on CAC40 97 10 min unseen data (970101 to 971231)
Indicators : ADX(25),DMIplus(7), DMIminus(7),ADX(25) -ADX(25)[3] .
Let's stick again with the solution #3
and run again the Check mode on the corresponding ASCII file.
5 minutes later, we get all the results for all solutions:
Click on the above image to get a larger
picture (1600x1200)
Testing on SP500 10 min unseen data (980101 to 980903)
Indicators : ADX(25),DMIplus(7), DMIminus(7),ADX(25) -ADX(25)[3] .
The best solution is now #3.
Column1 [Test Crit] in the sub window 2 shows
that the best solutions are #3,20,2,16,24,17,14
We may notice that #3,14,24, were already found as
the best solutions in the previous test on unseen data,
so, it demonstrates their robustness.
If you forgot it, do not worry, so the results are stored in the Journal File.
We definitively save Sol #3 into a FZB file ,
ready to run in TradeStation from a template system.
This model was trained on 4892 bars of SP500 10 min intraday
data,
tested on unseen data on a broader scale and even different market.
A robust model must work on different data, different timeframes.
We propose to run solution #3 on...daily data
(US commodities daily Continuous contracts)!
We will also compare the results with the DMI(7) crossover,
directly in TradeStation.
value1=dmiplus(7)-DMIminus(7);
if value1>0 then buy at h stop;
if value1<0 then sell at l stop;
Various commodities are presented here
back to top
COTTON, ORANGE JUICE, WHEAT, BRITISH POUND.
The dual screen displays
The Neurofuzzy system on top window
(our very same previous solution #3 trained on the SP500 10 min file).
Equity curve displayed in blue color.
The DMI(7) crossover system on bottom window.
Equity curve displayed in yellow color.
Tested on Continuous contracts provided by CSI. Ends 1994 03 31 .
Over 20 years of data for each commodity.
No transaction costs for both systems.
Please notice the Equity Curve slope,
and the buy sell arrow's location.
Then you will really understand what Safir-X was able to do...with so few information!
Click on the above image to get a larger
picture (1600x1200)
Printed usingTradeSation4 Charting on COTTON daily data
TOP Neurofuzzy system / Bottom
DMI(7) crossover
Click on the above image to get a larger
picture (1600x1200)
Printed usingTradeSation4 Charting on ORANGE JUICE daily data
TOP Neurofuzzy system / Bottom
DMI(7) crossover
Click on the above image to get a larger
picture (1600x1200)
Printed usingTradeSation4 Charting on WHEAT daily data
TOP Neurofuzzy system / Bottom
DMI(7) crossover
Click on the above image to get a larger
picture (1600x1200)
Printed usingTradeSation4 Charting on BRITISH POUND daily data
TOP Neurofuzzy system / Bottom
DMI(7) crossover
back to top
To summary:
We trained the model on SP500 10 min data (961021 to 970606)
using 4892 bars for training with only 4 indicators:
(ADX(25), DMIplus(7), DMIminus(7), and the ADX(25) variation over 3 bars.
It worked on 100 years of daily data and 3 years
of 10 min intraday data,
different markets, different timeframes, and outperformed from far
the classical system using the very same directional indicators
breakout system, with fewer trades and steeper equity curves...
First, the above results prove that it was not gambling at
all,
if we were able to discover the proper rules.
This is what Safir-X Assistant has done for you,
in less than 30 minutes. All automated...
Unfortunately to our comprehension, they are numerous and not easy to understand
You may wish to know more, and we can even go a step further
( the fuzzy rules are disclosed in a special text file if you run the ANALYZE command
in the file menu).
You may want to understand, so we provide the necessary tools.
Here is a partial example of the detailed information
that you may want to get...
back to top
------------------------SUMMARY--------------------
FUZZY DESCRIPTION : new.906252923
DATA from : C:\SirTest\Spom2.tsd
FUZZY VARIABLES :
| Ind.1 : | <-82.970 | 71.010>; | Nb F.S. : | 3 |
| Ind.2 : | <0.000 | 61.150>; | Nb F.S. : | 9 |
| Ind.3 : | <-64.920 | 79.510>; | Nb F.S. : | 5 |
| Ind.4 : | <-3.090 | 4.530>; | Nb F.S. : | 3 |
RESULT OF DATA ANALYSIS (TRAINING PART)
| RULE | Fuzzy Sets # | P. | # Mean | ConcV | # Wrong |
| 31 | -0-2-0-1 | 1 | -1.000 | -0.273 | 0 (0 %, 100.0 %) |
| 34 | -0-2-1-1 | 1 | -1.000 | -1.045 | 0 (0 %, 100.0 %) |
| 37 | -0-2-2-1 | 1 | -1.000 | -1.098 | 0 (0 %, 100.0 %) |
| 46 | -0-3-0-1 | 1 | -1.000 | -0.243 | 0 (0 %, 100.0 %) |
| 49 | -0-3-1-1 | 2 | -1.000 | -0.669 | 0 (0 %, 100.0 %) |
| 50 | -0-3-1-2 | 1 | 1.000 | -0.140 | 1 (100 %, 99.5 %) |
| 52 | -0-3-2-1 | 2 | -1.000 | -1.168 | 0 (0 %, 100.0 %) |
| 61 | -0-4-0-1 | 1 | 1.000 | 0.305 | 1 (100 %, 99.6 %) |
| 62 | -0-4-0-2 | 1 | 1.000 | 0.263 | 0 (0 %, 100.0 %) |
| 64 | -0-4-1-1 | 4 | -1.000 | -1.011 | 0 (0 %, 100.0 %) |
| 65 | -0-4-1-2 | 2 | -1.000 | -0.393 | 0 (0 %, 100.0 %) |
| 67 | -0-4-2-1 | 4 | -1.000 | -1.108 | 0 (0 %, 100.0 %) |
and so on...for all the 397 rules in this model
back to top
RESULT OF DATA ANALYSIS (TEST PART)
| RULE | Fuzzy Sets # | P. | # Mean | ConcV | # Wrong |
| 19 | -0-1-1-1 | 1 | -1.000 | -0.746 | 0 (0 %, 100.0 %) |
| 31 | -0-2-0-1 | 1 | -1.000 | -0.273 | 0 (0 %, 100.0 %) |
| 32 | -0-2-0-2 | 1 | -1.000 | -0.085 | 0 (0 %, 100.0 %) |
| 34 | -0-2-1-1 | 2 | -1.000 | -1.045 | 0 (0 %, 100.0 %) |
| 37 | -0-2-2-1 | 3 | -0.333 | -1.098 | 1 (33 %, 99.5 %) |
| 38 | -0-2-2-2 | 1 | 1.000 | -0.165 | 1 (100 %, 100.2 %) |
| 47 | -0-3-0-2 | 1 | -1.000 | -0.070 | 0 (0 %, 100.0 %) |
| 49 | -0-3-1-1 | 1 | 1.000 | -0.669 | 1 (100 %, 99.0 %) |
| 52 | -0-3-2-1 | 1 | -1.000 | -1.168 | 0 (0 %, 100.0 %) |
| 53 | -0-3-2-2 | 1 | -1.000 | -0.365 | 0 (0 %, 100.0 %) |
| 68 | -0-4-2-2 | 2 | -1.000 | -0.411 | 0 (0 %, 100.0 %) |
| 77 | -0-5-0-2 | 1 | -1.000 | -0.008 | 0 (0 %, 100.0 %) |
| 79 | -0-5-1-1 | 8 | -0.500 | -0.900 | 2 (25 %, 99.8 %) |
and so on...for all the 397 rules in this neurofuzzy model.
Fuzzy rules are also disclosed, for each rule available and for each pattern that occurred for this
rule
Fuzzy sets boundaries are disclosed too.
Internal statistics also available (shows
the patterns e.g. bars where the rule was active, patterns in error).
I only copy three examples (among 397 available) for
this solution:
back to top
------------------------DETAILS (TRAINING) --------------------
Rule ----- 31 (nb
patterns = 1) # wrong answers 0 (0 %)
IF (Ind.1 = F.S.0 AND Ind.2 = F.S.2 AND Ind.3 = F.S.0 AND Ind.4 = F.S.1 ) THEN
Sig1 = -0.2728;
<-159.960 | -5.980>;
<11.943 | 20.184>; <-137.135 | -19.786>; <-3.090 | 4.530>;
Pat# Exp. Comp. Act. Err
1430 -53.629 17.747
-51.599 1.765 -1 -0.371 0.27 0 0.63
--------------------- Mean value : -1.000 Min value : -1.000 Max value : -1.000 Conclusion
value : -0.273
Rule ----- 171 (nb
patterns = 16) # wrong answers 7 (44 %)
IF (Ind.1 = F.S.1 AND Ind.2 = F.S.2 AND
Ind.3 = F.S.2 AND Ind.4 = F.S.0 ) THEN Sig1 = 0.3243;
<-82.970 | 71.010>;
<11.943 | 20.184>; <-19.786 | 34.376>; <-6.900 | 0.720>;
Pat# Exp. Comp. Act. Err
186 -0.964 17.323 -3.409
-1.266 -1 -0.204 0.29 0 0.80
1096 -11.120 17.903 3.003 -1.260 -1 -0.040 0.32 0 0.96
1293 6.201 18.102 9.173 -1.344 1 0.290 0.31 0 0.71
2176 -1.486 17.785 1.375 -1.323 -1 -0.013 0.33 0 0.99
192 16.218 17.985 5.874 -1.342 1 0.319 0.28 0 0.68
2193 15.607 16.726 -0.611 -1.259 1 0.035 0.24 0 0.96
2409 -3.418 18.368 0.686 -1.221 1 0.049 0.24 0 0.95
2460 12.210 16.640 8.399 -1.187 1 0.138 0.32 0 0.86
2511 11.368 16.495 17.632 -1.257 1 0.148 0.21 0 0.85
1376 9.088 17.668 13.571 -1.329 -1 0.236 0.29 1 1.24
2022 -10.999 17.546 -1.087 -1.200 1 -0.214 0.30 1 1.21
050 14.053 18.345 -1.776 -1.309 -1 0.127 0.17 1 1.13
2065 15.401 16.831 -2.123 -1.260 -1 0.008 0.22 1 1.01
2288 3.260 17.128 -6.032 -1.270 1 -0.216 0.23 1 1.22
2422 -6.950 18.217 13.441 -1.333 -1 0.233 0.28 1 1.23
2436 18.142 17.724 -4.127 -1.189 -1 0.055 0.17 1 1.06
--------------------- Mean value : 0.000 Min value : -1.000 Max value : 1.000 Conclusion
value : 0.324
These mean that rule #31 was
seldom used (one bar only), rule #171 has
been made active 16 times during training and produced a 56% rate of the good answer (44%
in error).The error occurred on patterns in italic.
The same results are available for test data.
You can examine anything wrong using this method, compare the behavior of any
rule on any pattern between train and test data, but it is not mandatory. This is an
ultimate refinement available to those who want to verify what is really inside the black
box and understand all of its subtle behavior.
We now want to stop the discourse and let you experiment by
yourself with the free evaluation version.
What could be done is
Add a couple of indicators (a first difference on DMI or K%, on ADX, a volatility indicator, a volume based indicator).Usually around 5 inputs is a fair number. No need to use 14 indicators that are partially redundant.
Use the TRAIN command than can enhance a solution by refining the neurofuzzy rules.
Try to train with a different Evaluation function from the performance Summary. Chose the one that could highlight the weakness of the Equity curve.
You may consider using several neurofuzzy systems into a big one (feasible in TradeStation only). Use a multiple vote in this case.
You can retrain the neurofuzzy system within TradeStation under any condition that you are able to write in Easy Language. This is a time consuming task, but very interesting and promising. Being done after a fairly good system was found, anyway.
You may train different neurofuzzy systems and switch them in TradeStation.
Possibilities are endless and only limited to your imagination...
Use different indicators. The examples given here are not the best available. They were used because anyone knows them and had used them in the past with probably fewer successes.
Not at all!
Here are the details of the whole operation
| Task# Time to completion on a Pentium II 300Mhz |
Difficulty level: 0= obvious to anyone. 5= Obvious to Einstein. |
TASK (OPERATION) DETAILS |
| 1 2 min |
1 | We have
built a training ASCII file (*.TSD) |
| 2 15 s |
0 | We have
opened The Assistant version 2 |
| 3 15 min+coffee |
0 | A few clicks later (LOAD,..EXPLORE), the Assistant was doing its automated search, stored the results and produced the comprehensive reports that you have already seen in the various screen shots above. |
| 4 1 min |
1 | We have made a
new ASCII TSD file |
| 5
min back to top |
1 | We used
the same template to build any TSD ASCII file for use in the
SAFIR-X Assistant for Expert Traders. Only the indicators and filename were modified. |
| 5 1 min |
0 | We
loaded the new ASCII TSD file in the Assistant |
| 6 10 to 20min+coffee |
1 to 3 ( I Can't say for you!) |
...click
on CHECK, and all the solutions were tested, stored, displayed on
unseen data. |
| 7 5 min |
1 | Running
the neurofuzzy system in TradeStation |
Like any other trading system...You get buy sell arrows, Performance Summary,
Active Orders, like if it had been a moving average crossover system.
The results are the same than if you were testing in the Assistant:
For ending, here is a comparative screen copy of another system built with the Assistant
running in TradeStation 4 .
(With the above templates using another set of
"gambling indicators").
SAFIR-X display
Indic1(dmiplus(14)), Indic2(DMIminus(14)), Indic3(slowd(20)), Indic4(rsi(c,5));
SP500 961202-970613 10 min
Training was made on the first half of the file.Buy Sell signals display as paintbars on price data
back to top
click on the above image to get a larger
picture
TradeStation 4 display (same neurofuzzy system, same data, same
costs)
Indic1(dmiplus(14)), Indic2(DMIminus(14)), Indic3(slowd(20)),
Indic4(rsi(c,5));
You should be able to recognize the last third of
the Equity Curve(white) shown in the previous SAFIR-X picture,
( TradeStation do not display all the Equity Curve)
The above examples only represents least
performance that SAFIR-X version 2 should achieve using
such trivial indicators, provided that you select them with some basic logic
sense. Better is not so difficult with it...
back to top
Safir-X, The Assistant for Expert Traders has been developed by JewelSoft using proprietary neurofuzzy logic techniques and original search algorithms especially adapted and created to the very specific problem of automated trading systems building.
Sirtrade 97, SirTrade98 have been developed by JewelSoft and SirTrade Intl.
Pierre Orphelin (Sirtrade Intl.) was acting as a consultant and technical analyst expert / advisor in both cases.
An evaluation version2 is available on our web site too.
Any version 1.11a and below owner qualify for a free upgrade to version 2 as soon as available.
The Assistant for Expert Traders version 2 FZB files will be TradeStation 2000i compatible.
International Distribution:
Sirtrade International.
Exclusive resellers (national) are welcome.
Test, screenshots and calculations were performed on the SAFIR-X Assistant for Expert Traders beta version 2.06d by JewelSoft and TradeStation 4 Build 21 by TradeStation Technologies, Inc.
Web page, tests and comments, © Pierre Orphelin, SirTrade International -1998. All rights reserved.
