Systematic Portfolio Diversification - Data Mining Concept

These explanations are super awesome A. Thanks muchly

What type of order does the robot uses?
Stop, limit or market order?

Entries are based on market order with the EA generator I am currently using...but this project is not restricted to a particular EA configuration. All EA's welcome in other words provided there is a Monte Carlo solution for it. The reason I am restricted to a particular approach (EA Studio) is due to the Monte Carlo simulator that it comes with which is a pretty good tool for our uses.

Finally!!!!!!!!.....Please Update Your Tools Guys

Attached Image

The tools are now complete and we are using Version 1.1

Please re-upload the database if you are not using this version number.

I am going to chill tonight....so will make a start tomorrow.

I will be using these tools with you and we will be starting from scratch so we can all keep on the same page. Remember that this thread is not restricted to this project alone so if you have good food for thought that relates to this project and the topic of data mining.....then please share. I really appreciate the input so far and am learning heaps from your contributions so please keep doing what you are doing.

Cheers

C

why don't you try it by yourself my friend. and you should have known

oh right, didn't notice there're attachments.

Hi alphadude,

Firstly thanks a lot for taking the time to describe these processes - it is very interesting and informative. I have done loads of such research myself over the years, and I hereby promise to document and describe some of my own research here.

That said, I'm keen to get the specifics of parameterless optimization clear in my head (and apologies if I'm being dense - I write my own software in c++ to do this stuff, so there may be assumed knowledge of some platform that I'm missing - and did I mention I may well be dense! :-)

1 Am I right in thinking that you are distinguishing the algorithmic structure and logic of a strategy from the particular parameters specifying it's inputs? For example, if I optimized the lengths of 2 moving averages in an MA crossover strategy, that would be parameterless optimization - whereas if I added a third moving average into the mix, that would not?

2. You say that parameter optimization has nothing to do with system profitability, and that you are choosing parameter values that "make sense". I'm left wondering what criterion you are using to choose your parameter values if not system profitability - or have I got the wrong end of the stick and it's pure human judgement?

Once again, thanks for your patience and for the book recommendation. I'm reading it atm., and it's very interesting indeed.

Best regards

Matt

All Good Guys

Put your helmets on.....we are about to go mining :-)



The workflow process has been adjusted since the beginning of this thread and will take the following form.

A) Data Mining (Robot generation) using available data up to 31/07/2016.

I will describe the broad parameter constraints that we use to restrict our universe of options to a defined logic that is divergent in nature or at least forces a divergent outcome and show what settings are used on EA studio to do this.

Data Mining will be undertaken on whatever data we can get our hands on for a particular instrument and timeframe that can be applied to my robot generator. To commence this exercise we will be restricting our solutions to D1 and H4 given the availability of long term data that captures a range of market conditions.

We only test up to 31/7/2016 as we leave data untouched for OOS testing up to current day. The idea behind this philosophy is that unlike alternative methods looking for what is working in recent conditions, the conditions we are looking for are extraordinary conditions of divergence where we reach for the moon. We need to ensure that these conditions are present in the IS sample range to ensure our bots do what they are meant to do during these times. At all other times, this is where our focus is applied to uncorrelated collections. We do not pay any head to the profit performance suggested by the solution over these periods as they are likely to be curve fit results. We actually have a way to discriminate what is curve fit and what is not in this process using visual mapping methods so stay tuned.

The Robot Generator we will be using is EA Studio. The bottom line is that it really doesn't matter what tool you use here...provided it is quick and can generate well written standardised code that we can later tweak if we need to. Most of our testing actually will be undertaken outside this tool as you will find that nearly all generators on the market bias results to either convergent solutions or curve fit solutions.

We can utilise EA's from any source however and test with our own method.....however it just will not include the Monte Carlo simulator unless we can find a stand alone solution to do what I would like it to do. So at least while we start, let's keep it simple and use the outputs of EA studio. When we refine this process to something we are happy with, then we can open the flood gates to other EA generated solutions.

During the data mining exercise we actually do not use data from our platform where these EA's will ultimately be trading live. We undertake this test as a later exercise to ensure that the EA's are not specific to a brokers data set. We will actually be using FXOpen data for our testing as it operates of the same time GMT+2 as Pepperstone which will be the broker I will be running this live on. Given that the robots generated by EA Studio operate on the open of the bar only (to fasten EA generation), having the same Open on D1 is an important factor in the EA generation phase.

We however still test on Dukascopy tick data as a necessary counterbalancing check......so in this exercise the one EA that is generated is tested on FXOpen, Dukascopy and Pepperstone data.

FXOpen (open price data) - In sample testing from whatever is available to 31/7/2016;
Dukascopy Tick Data for any available data before FXOpen IS data, IS data to compare with FXOpen plus OOS data from 1/8/2016 to current day;
Pepperstone open price data - from whatever is available up to current day

Monte Carlo
This process will produce a collection of EA's for a single instrument and timeframe where we have included a strident Monte Carlo validation method to hopefully ensure that of the options created in the collection, there is a good chance that the collection will comprise both curve fit solutions as well as some gold nuggets. It is the gold nugget that we want to be able to discriminate so it then passes to the next phase of testing outside EA Studio and in our own hands using the methods that have been homegrown using the tools I have provided on the commencement thread.

Validation Testing

B) Mapping to the Condition - This is where we test the concentration of gold in the samples generated and create the pile of profitable tailings for further processing.

1. We map the Monte Carlo array to the market condition. Result will be a yes (continue) or no ( eliminate). Reasons will be provided and discussed.
2. Using Dukascopy Tick Data. We map the equity curve to the market condition and define IS periods and OOS periods and look for evidence of Divergent Behaviour plus evidence of curve fitting. Result will either be a yes (continue) or no (eliminate). Reasons will be provided and discussed.
3. Using Pepperstone Data we map the equity curve to the market condition as above for Dukascopy. Result will be a yes (continue) or no (Eliminate). Reasons will be provided and discussed.

C) Analysis (Crunching Phase) where we use benchmarks and our analytical spreadsheet as a basis to eliminate the pyrite and fools gold from the Gold. We go a bit geek here :-)

D) Incubation Mode - Where we put the robots on demo for at least 6 months to ensure they are free of execution errors, continue to perform as planned out of sample, and remove any slag (underperformers) that take results below required benchmarks.

E) Live Mode - This is where we put our capital on the line and invest in Fort Knox and continuously monitor performance. This process involves rotating out poor performers and adding new performers to the ranks from this continuous workflow process.

So tomorrow....I will describe the Robot generation process using EA studio. I have set up a run tonight which will push out strats for tomorrow so I will desceibe the details when we have some initial collections to work with.

We will also be removing our test data from the database and replacing this with the results of this exercise as we go. We are starting from scratch guys so no one will be left behind.

Cheers

C

hi Matt,

Good questions you are raising; and the answer lies within them

To make the matter simpler to understand; let's divide "system designers" onto two categories:

1. Category 1: system design based on some existing strategy fond on forums, given by friends, read in books, etc; but you do not understand where the edge is coming from.
This is perhaps the most common scenario used by various traders.
They get an idea of a system from the forums/books/friends; and there you go; optimize it to death till it produces a good parameter combination with profitable results.

Basically you are searching many unknowns out there. You are optimizing a system parameters; but you do not understand why certain values are profitable.

2. Category 2: system design that starts from a known edge in the market; or perhaps from observations.
given your example above; for a dual moving average crossover system; what is your understanding of that system; why would you think that it can be profitable? in other words; what is the edge?

if you understand the edge; then you can easily do statistical analysis of every single bar in historical data; to see how the market behave.

for example; in your dual moving average crossover system (let's assume it is a trend following strategy); you would understand the edge as; whenver the two moving averages crossover; we should go long; and we should exit position when they cross back (or any other trigger).

Now you should do a statistical study as following:
1. Assume transaction cost is 0.0 (no commission, and no spread). This is essential; and this is berhaps a *secret* am revealing here
the price to use for entr/exit is the mid price (middle between bid and ask)

If a system is not profitable (without transaction cost) then it will definitely not be profitable with it.
If a system is profitable without transaction cost; then you may have an edge. You can add filters to overcome transaction cost later.
why would you spend your valuable time; on a strategy that even with no transaction cost; is not profitabe!

2. For every single bar after dual ma crossed over; assume you enter at the close price of that bar; and exit when the moving aveages crossed back.

3. Calcualte the average PNL of all bars.

4. If the average pnl is positive; then you have discovered an edge.

5. You can further do stability analysis; but dividing the pnl onto yearly; if the average pnl is similar for every year; then your result are stable; or in other words; the edge is stable.

6. Now that you beleive you have an edge; you can start devising your strategy. Like adding some filters; including transaction costs, etc.

7. How do you decide on the period of the moving averages?
this is where parametless params comes to place.

There is no easy answer in this specific example. Some math/statistic scientists spent years in this area with no success.
see for example: Stationarity tests of timeseries

You could however; do optimization to find out if the timeseries of the dual ma crossover; is stationary or not. By testing various values for moving averages.

Also the timeframe is important; would you do the study on daily timeframe, hourly, 15min, 1m, 1s, ticks?
We know from observations; that the shorter timeframe; the noisier the market is. so perhaps you can choose higher timeframe for momentum strategy; and shorter timeframe for mean reversion.

==========
Another way to put it; you need to test your edge; before you even design your strategy.
If the edge exists; and proven by your study; then the next step is how to turn it onto a profitable strategy.
Doing optimization blindly will not get you there.


That is why I personally do not use any machine learning algos, or neural networks in my trading. I have ventured onto that area in the past; and concluded that there is nothing better than starting from market observations, from having an stable edge. A la Crable style.

hope this helps.

btw; I am a java developer (also c++ developer); but I use java in all my work. I would strongly suggest you jump to java because you will save heaps of time in system development.

I would highly recommend reading Ernest Chan books; and perhaps his blog as well. He talks about parametless systems in detail.
His process of designing systems; is very methodical.

I am at a point where; if you describe a strategy to me in few words; and what market you are trading it, and timeframe; I can tell you straight away if it works or not. Simply by having tried many experiments; on many markets; I get some understanding of the mechanics of markets.

Sigh...I am trying hard to stay away from this thread. :-(

C'mon....you know you should be in here :-)

Just another pearler A :-) I am going to have to stop repeating myself.

Good Tips RR. We are taking care of some of them. :-)

Hmm... so if all my HTF and MTF setups are tradable on M1, am I trading noise and against algo? Not sure I would agree ....what we see on monthly shows up on M1 and what we see on M1 also contribute to monthly setup.

You are all looking a bit clean. It's time to go down t'ut Pit and see how the plant operates :-)



EA Generator Description (Robot Factory)
While there are a number of fast EA generators available on the market, the approach I am using uses EA Studio. It has it's shortcomings like other systems but has some significant pluses as well, namely the ability to generate huge collections of strats which can be seamlessly generated for MT4 or MT5 with minimal execution issues. We have one technical issue on the broker open condition we need to look at when we put it in the walk-forward environment, but that is easily fixed with a code tweak.

EA Studio operates in the cloud and is browser based. You can have up to 4 sessions running per PC but with more running you may notice a deterioration in processing speed.

Ideally if you are running long sessions, then put in on your VPS as the system itself runs 'similar strategy' and 'correlation tests' with the solutions it generates to keep the collections of solutions generated 'fairly' non-correlated. We ensure this is not the case in later treatment...but it is a handy process while inside the generator as it broadens the outputs generated across your design logic universe.

We run EA studio in 'Reactor Mode' whereby there is a workflow applied to the random strategies that are generated.



The right hand side of the screen shot above displays the workflow we will be using. More options are available if you like such as 'Walk Forward Validation' within EA studio, however we do this off the platform using out of sample data (OOS).

Generator
You can see from the right hand side workflow that the 'generator' creates random strategies that lie within the design logic to define for the generation and then undertakes validation according to your defined rules as a first pass. In the example above the generator produced 246768 strategies in about 2 hours (for EURUSD D1 for the timeframe 2/1/2007 to 29/7/2016 which is 1% of time assigned to the workflow. Remember that you can have 4 of these running simultaneouly so the mind will boggle regarding the permutations being considered.

The speed of the testing is due to it's operation on Open bar condition only. This has its pluses and minuses....however as a later treatment stage, we test under tick data to overcome some of the downside of open bar design.

Of the 246768 strategies generated 332 strategies or 0.13% of the strategies passed the first validation test. We will discuss the validation criteria I use later.

Optimisation
The 332 strategies then move to the next workflow process of optimisation. We loosely optimise the result as the correlation and similar strategy rules eliminate similar strategies and the optimiser configures the strategy with the optimal variables in accordance with the data sample it is working on. Unfortunately this is a necessity given that you are using a 'random generator' constrained by your logic universe. You want to at least ensure that the solution has been honed to the optimal condition BUT it will by definition be curve fitted....so beware.

Within this end solution the result will either be totally curve fit which we can detect using later treatment methods that we will discuss, or it may contain a kernel of gold (an auto-correlated non random solution') within the noise of curve fitting.

The 332 strategies that flow into the optimiser are then optimally configured to produce 332 possibilities that may include a gold seam.

Monte Carlo Validation
Then onto the MonteCarlo Validation which is our first attempt to strip out the noise (curve fitting) from the solution. There are many different types of Montecarlo treatment which I will demonstrate in a later post, but the one we will be focusing on as a more rigorous validation method is the 'Randomisation of Strategy Parameters' to see the sensitivity of variables contained within the strategy and thereby make a judgement of the robustness of the variables chosen. We adopt the Perry Kaufman approach to validation here to test for strategy robustness.

Of the 332 strats that entered the Monte Carlo Validation phase, 5 passed the MC validation criteria (we will discuss the criteria later).

Multi-Market Validation
Given that we are looking for divergent solutions of 'fat tail like disruptions' which are not market specific (they can be present anywhere at any time), we want to ensure that the solution that has got this far can handle a far larger data set. Because we are data mining...data is just data. We are not looking for fundamental or technical reasons for why a strategy works or not with this approach, but simply letting the data speak for itself. In this sense as opposed to 'thinking in terms of different markets'...it really is just data configured in different ways.

So what this test really does (forgetting for the moment that we apply it to different instruments), is that we are just throwing additional data at the solution as a test for robustness and to validate that there may be a better chance than random that the strategy does perform during divergent market conditions.

Of the 5 strategies that went into the Multi-market test, 4 passed the test. (We will be describing the validation criteria of the test later). Note that it is now 5 minutes later and a 5th has passed the test. If I select the built MT4 or build MT5 EA above each strategy in the collection then in an instant an EA is generated which can be plonked onto you preferred broker MT4 or MT5 platform where you can take it for a roadtest in the backtest or demo/live walk forward environment.

That's when we start then applying our own home grown treatment of the tailings as we like to be a bit surer about the possibility of a gold strike and as much as we like the programmatic of the factory, we prefer treating the spoils in our own preferred way. One of the surprising benefits of this approach is that you start recognizing each strategy for what it is,........simply a means to an end. We don't form any lasting attachments to these creations (given their speed of generation) and our ability to make millions of them. This ensures that we have an underlying adaptable solution on our hands as we will not be emotionally attached to the strats and simply use performance criteria as a basis to keep them or not. We don't hold onto 'weak individual solutions'. Our entire focus is preferentially oriented to the portfolio level (which is coming) and these strategies can be thought of as individual return paths of a much bigger picture.



Think about these solutions not as end products but as a sample that may or may not contain a concentration of gold. We collate these collections of strats into a pile (gold prospects) which we then further treat off the platform to strip away the propensity for curve fit results and further concentrate the deposit of gold that may be contained therein.

So now we understand a bit about how the factory works we will start delving into the detail of the settings to define our logic build.

That's the first shift over. It's smoko time.



Cheers

C

Cheers RR...but there is a generalization in that statement. Those particular investors may have found that diversification failed. Another class of more highly diversified investor who chose CTA's as there vehicle were paid in spades because of correct methods of applied risk-weighted diversification.

Dumping your money aimlessly in different amounts across asset classes does not equate to a sound method of diversification.

C, Nice work; am going to try this studio.

a word of caution; you may get kicked out of VPS if they discover you are doing backtesting.

this is because it drains their main server CPU affecting other users.

It is all written in the small prints

Cheers A.

Definitely noted mate.

I had it going (2 instances) for a while in conjunction with a demo account on VPS. I therefore needed to ensure the process was minimal on system resources to not compromise trade execution on the demo and monitored system resources to ensure that max use kept to below levels of 50% usage. It worked for a few months without interfering in any way with the demo account pushing 40 strats. The moral of this story is that it is pretty lean on it's share of syatem resources. This was very surprising to me. I have got no idea how this is achieved but its cloud basis may be a major reason why?

It's Time to Look Under the Hood at some of the Detail and How to Configure for the Divergent Case

Attached Image (click to enlarge)

In essence the method of strategy construction by an EA generator like EA studio is that you set broad universal constraints within which the random strategy generating exercise is performed.

Given our knowledge of some of the principles of the Divergent case (being advocates of momentum/trend) such as; the principle of 'cutting losses short and letting profits run"; the family of traditional indicators used; and other broad rules such trailing your profits......you can enforce these design principles into the logic to restrict the random outcomes within a broad class of design principles.

In other words we want an all purpose car that is rugged and can navigate difficult terrains and therefore we want standard core principles embedded in the design that will result in a positive correlated case when markets diverge but we also need variation around this core design to give non-correlated cases when markets are not diverging.

The beauty of system design is that you can enforce the degree of correlation unlike say diversification into alternative asset classes and timeframes where correlations are a moving feast dependent on the markets whim.

For example:
An investment in Copper and an investment in EURUSD may be uncorrelated over certain conditions, but may become correlated during others. You will not be able to manage this degree of correlation. It is the market that decides. While in general it is wise to spread your capital allocation across current non-correlated asset classes, you need to constantly monitor market correlations and also examine market data to see where non-correlated assets become correlated...........On the other hand let's take an extreme example of an enforced anti-correlated design solution. You have heard it all before. It is a hedged trade. It is the design of the trade that guarantees the permanent hedge and the enforced correlation. With design you can enforce static correlations over time which is just what you need as the method to quell the drawdowns during extended periods of market convergence.

Anyway.....enough blarbing....let's get engine grease on ourselves and have a look at more of the detail of this fiendish contraption called EA Studio (which we will call our metal detector)....thank you Minelab.

To setup you the Reactor (aka workflow process), you start with the Tools Menu - Settings.



This screen is pretty self-explanatory but you can see the option for 'detect balance lines correlation' and 'detect strategies with similar rules'. This is a cool feature that assists in ensuring you are not generating lots of the same family of solutions with similar correlation.

You will also note the export settings option and import settings option. Like MT4 strategy tester, this creates set files of the entire workflow process and all settings configured so once you have set things up once and have saved it, you simply import it back in and change your instrument or timeframe to test on. Very handy feature and great time-saving device.

The set files created are stored in our dbase and categorized by family type so that when we want a logic build with a particular configuration, we can access this on the portfolio. I will set this up later as an example to show you. For example you could have a trend following setfile, a news event setfile, a a breakout trader setfile, a mean reverting set file....whatever floats your boat and you are confident that the applied logic meets the case.


Tools - Acceptance Criteria

This is where you set the initial validation criteria in the workflow that is the first cut applied to your randomly generated solution. You can do this for in sample (IS) and out of sample periods (OOS) or whatever floats your boat.



The range of different validation criteria is fairly comprehensive and you can configure one or as many as you like.



You can see that I am applying max ambiguous bars (which restricts the maximum number of invalid bars across the test range to 10 or less.

I also set low minimum counts of trades for validation as my divergent solutions are infrequent traders. Too high a number will bias the result towards convergence.

The Minimum Return/Drawdown ratio and the Max Equity drawdown allowance is used to simply ensure I do not receive disastrously volatile solutions. I actually want volatile equity curves as they are a necessary prerequisite for portfolio construction but there is a fine line between extreme and valuable to the portfolio. Extreme solutions bias the portfolio towards that extreme result. We want a nice blend.

Because I am testing over long timeframes I am being realistic in my expectations of minimum profit factor and max equity drawdown. If you set these too high, I'll warn you once./....you will get curve fit solutions only. Somewhere lurking within randomness lies the 'real value' and the history of trading is certain that this result is not an extreme value. There is only extreme profitability in 'lucky solutions', not 'realistic solutions'.

One of the reasons we need to define a least a very small profit factor is that the solution if found to have what it takes needs a slight positive edge for portfolio magnification purposes. Having such a low profit criterion at least guarantees a result that produces a wide range of both curve fit plus legitimate possible solutions.

We could use a swag of different validation criterion.....however for our purposes this is sufficient. It is the portfolio that does the heavy lifting not the individual strategy.

Tools - Available Indicators
This is where we define the range of possible indicators we will apply for the entry and exit condition. The list is large and it doess miss some that I would like...but beggars can't be choosers.

One of the benefits of this cloud based EA generator is that over time the functionality is improved by the programming team without any additional cost. There are some great projects coming that will be included in this software including some enhanced portfolio management functionality. You can have a look at the EA studio forum to get some ideas. Also for programmers familiar with the Programming language for the browser, a scripting language is available for users to create mods and new indicators etc.



Once again I define the choice of available indicators to those that I am familiar with an are possible solutions for divergent strategies. I simply turn off those that I have issues with or are utilised expressly for convergent outcomes.

Tools - Preset Indicators
There is also a very useful function that allows you to apply a preset condition to all strategies generated. I use this to restrict my trade solutions to those times when price has made an extreme excursion away from a past equilibrium. Traditionally I would use a Donchian here...however that is not available *sobs hysterically* so instead on D1 I use a 200 SMA as a preset condition to all strategies.




Tools - Data Horizon
Unfortunately the program has a restriction for 200000 bars of data. I therefore set it to this extreme value as I am looking for long data ranges over multiple market conditions. This constraint shortly will be released according to the developers where we can go to town on very long data samples.

Selecting the start date and End date limit is used to define your IS range or alternatively test a specific period of market data. Having

Tools - Data Import

This allows you to import data files for testing from alternate sources. There are scripts that convert the standards MT4 data into a format compatible with EA Studio. Provided the rules are consistent and in a format allowed by the script, you can import any data into the Robot Factory for testing and it is a very quick process provided the files have been script converted first.



For example I have a dedicated VPS assigned to capture market data from Pepperstone (Live Razor account) (Refer below). It is left on all the time to allow for progressive population.

Despite this backup and alternative dataset we could use, I use FXOPen data for all testing which is maintained by the EA platform provider....just another task I don't have to do myself.

Smoko Time again.

Attached Image

In summary this is very like prospecting for gold. The random generator alone would be like having to look for gold everywhere. Fortunately with a bit of logic applied we can restrict the search to divergent conditions (areas where gold is more likely to occur). We will be further restricting the possible search context further in my next post but you can see how we are getting to load bearing minerals that have a greater chance of containing gold. Of course when we turn our detectors on we also get a lot of other metal and general signal noise, but at least we have now defined a narrower context where our chances of finding gold have been dramatically increased already. There is more design work to do before we release the reactor.

Next stop when the Factory reopens we will be looking at the reactor settings.

Cheers

C

Let's get the Plant up and Running and Define the Reactor Settings

Attached Image

Now that we have defined a range of conditions within which the random generator will operate we have a final set of work processes that we need to define to fully automate the work undertaken during this phase. This data mining is back breaking I tell you. :-)

Having defined the settings applied on the Tools Menu, we now jump back to the Reactor Settings. At this stage we are not using the 'Generator' whjich is simply a subset of the work processes of the Reactor and is included in it's definition. The 'Validator' is something we won't be exploring currently but is a really great tool to import previously mined solutions back into the processing plant for either re-optimisation or to redeploy a successful previously mined strategy to another data source, instrument or timeframe. The process of validation re-formats the EA code to ensure it is standard to the generation suite and then you can undertake elements of the Reactor work flow to revalidate the earlier solution.



When you select Reactor, the following options appear. We have already discussed the workflow on the right hand side. Now we will focus on the options on the left hand side.



Historical Data

Attached Image (click to enlarge)

Here you select the broker data source. At the moment it includes 4 Broker Data sources that are maintained by the developer and also includes any Broker Data we have personally imported (previously described in Tools Menu).

The symbols available for testing are defined by the Brokers source data so varies dependent on Broker data supplied. Period is obvious.

In this test run I will be sampling EURUSD D1 data using FXOpen live data between the dates 2007-01-02 to 2018-08-30 which is derived from the data horizon settings previously discussed.

Strategy Properties


This option allows you to apply further constraint logic to your design and allows you to define Entry Lot size (Unfortunately at this stage you cannot apply % trade risk here....which is a real bummer....but they are working on it). You sometimes have to work with what you can get....so we will work with this limitation.

Stop Loss - Whether one is used or not and combinations of yes or no. In apply my 'cutting losses short' logic I always apply them.

Trailing - Whether a fixed or trailing stops or combinations are allowed. Once again my divergent logic dictates that I always apply a trail and never a simple fixed stop.

Min (pips) and Max (pips) defines the range in pip distance that it will look for an initial and trailing stop condition. I keep the minimum tight but let the distance wander to allow for variety for a divergent case.

Take Profit - If Take Profit is selected or as a combination further options are flagged to specify the conditions of the target condition....however given my divergent logic of letting profits run.....I do just that with all my solutions. I like to pray for infinite yield.


Generator Settings


Under this option you define the testing time you want to apply to the work processes. I keep this churning throughout the week and I am a slave to ensuring that something is crunching at all times. They key to this game is your capability of continuously mining the data and generating collections of strategies for further processing. Think of this EA Studio as the workhorse that allows to to focus on more important stuff at the portfolio level. It will work 24/7 for you. If you are not on VPS then periodically save your collections which you can do while the Reactor is still running.

The algorithmic search can be oriented towards a preferred performance metric. In this case I orientate the random bias towards return/drawdown ratio which improves my chance of obtaining a risk-weighted solution.

I unfortunately have trust issues when it comes to OOS as I want to ensure there is no internal data snooping going on inside a program I am not familiar with , so I ensure that I only conduct IS tests across a defined data set which in this case is for data up to 2 years from current. This window rolls forward as time progresses but 2 years for OOS and some OOS before the IS range is what I'm after for my specific testing method when we get off this platform. If you select OOS then you set the % OOS you wish to retain. This criteria is also retained for later optimisation of data and robustness tests....however we don't use this.

Once the Tools Menu, Historical data, Strategy Properties and Generator Settings are established this allows the work processes to get through the Generator sequence of tasks identified on the right hand side of the screen. This allows for random generation of strategies within a constrained divergent logic and preliminary validation of results to reduce the strategies generated to that set of strategies that meet validation criteria (previously discussed in Tools).


Optimize Strategies


This section of the work process optimises the validated prior solution 'weakly' to define a 'best fit' of parameters selected for the strategy. We could get extreme here and select up to 40 steps of whatever unit sizing increment you want. Unfortunately we do need to apply this feature weakly to improve our chances of having a valid result....but you need to recognise as discussed before that process will ensure there is a curve fit result produced that plots as the best outlier of your Monte Carlo distribution (coming up). Provided you accept this reality...we can treat the condition...but if you do not take this seriously....you are doomed Hanrohan :-(

Below is the application of these settings to a particular strategy example that has passed the previous stage. You can see how the parameters are selected and ranges applied for optimising.


Walk Forward Validation - I skip this process entirely as it is an outdated process from the past when we did not have better methods such as Monte Carlo at our disposal and a 2 year OOS untouched data sample to take the place of this process. It may be clever in it's rolling window method....but it doesn't suit what we are trying to achieve here. I've done the Bob Pardo course and I'm sorry.....the underlying logic does not make sense for divergence. It's time to move into the 21st century.....pfffffftttttt. Remember also that we cannot predict fat tails of divergence....so what is the logic in assuming that in a specially preserved portion of OOS data or a rolling window you are going to get this unpredictable event. If you use walk forward..then it's logic is tied to the single market condition and best left for convergent methods. This does not fit our purpose so we economically leave out a process that does not assist us in any measure.

Monte Carlo Validation
This is a summary description only. It is worthwhile focusing a bit on what we are trying to achieve here....so it will be a more detailed discussion in a future post. The number of single runs are defined here and the details of what is being tested is the subject of a future post....plus I'm getting exhausted. :-)



Multi-Market Validation


Simply select which additional markets you want the test to be conducted on over the same date range and then determine how many need to pass validation criteria (previously defined in Tools). In this case I am wanting at least 10 of 33 markets to pass the validation criteria for those strategies that get to this point of the process.

.....and there you have it.....now where is my beer?

C

PS The subject for the next few posts will be:
1. Monte Carlo Treatment;
2. MS Access Database Setup;
3. A return to EA Studio to look at processed results; and
4. Further refinement treatment off platform using our home grown tools of analysis and portfolio construction.

Stay tuned.