Systematic Portfolio Diversification - Data Mining Concept

I just checked EA Studio; the generator and backtester is running on their server. You are accessing their services through web browser; so you are ok with resources.

Cheers mate. Let their equipment work overtime I say. My equipment needs it's beauty sleep :-)

Very interesting video by Hedge Fund manager David Harding, on Trend Following Model on a diversified setting; pretty much similar to what Copernicus is trying to achieve.

Notice that David mentioned the secret of his success; is that individual strategies does not have a nice curve; but the portfolio does. Otherwise; many other traders would have taken that edge from the market.

Hi alphadude

Wow! That was a great elaboration on what you were saying. It's clear to me now, and I am in pretty much total agreement!

You make a good critique of "black box" strategy optimization. I have a background in machine learning (though via computer graphics rather than the financial area), and whenever I've tried to use a neural network for trading / forecasting it overfits atrociously. I think genetic programming (specifically cartesian genetic programming) is much better suited for this sort of task, as its results are always interpretable - but even having said that, its best use is as an "idea generator", rather than finished strategies.

However, even restricted to closed form optimisation of parameters, overfitting remains a constant adversary, so I'll share two techniques that I use in this battle. Having found an "optimum" set of parameters for some strategy, how can one judge the degree of overfitting? One test is to adjust the parameters a small distance away from the optimum values, and see how much the performance of the strategy degrades. Essentially measuring the "steepness" of the fitness landscape surrounding the optimum. If a small perturbation wrecks performance, this is a sign that the strategy is overfit, and is working by "chance".

Secondly, one can measure the percentage of somewhat successful parameter combinations in the search space. If the strategy has any sort of "edge" one should find a decent proportion of parameter combos that perform reasonably. This method seems closely related to your own process of tests you describe, for filtering out bogus strategies.

Anyway, thanks again for your patient descriptions. I shall now sit down, shut up (for a bit) and see what Copernicus' data mine produces! :-)

BTW Interesting you recommend Java over c++. I've programmed java myself (back in the applet days) and have been considering switching for a while, but need to look at what it brings to this area.

Best Regards

Matt

Yep mate....Harding is a great lead to follow. The key principal here is that divergence is non predictive whereas convergence relies on predictability. There is almost a dichotomy between the two. It really helps to think in terms of these two broad philosophies. In terms of the data mining technique itself that we will be applying, we step into techniques that we may not be familiar with such as mapping to market that allow us to examine the entire return path in the context of the market. Statistics tends to have an 'easy' bias towards linear thinking hence a tendency to be biased towards Gaussian outcomes as the statsitics we pay most heed to are the simplest forms of calculation. It is not that statistics are wrong....but rather in how we apply it for it is the simplest form of statistical model that we like to use in our interpretation of a reality given its ease in calculation and interpretation.

When it comes to 'hardness', we like to apply assumptions that have 'ease built into them' at a cost to reality. For example the force of Newtonian gravitation when applying a 2 mass problem (say the earth and the sun) is an easy calculation that we learnt at school which is a model that approximates gravitation in our earthly domain. However it is woefully inadequate to describe gravitation in a wider domain where we have 3 or more masses to describe and things move 'relationally fast' or where 'relational mass comparison' is extreme and fails as a descriptive model.

We therefore hold on to Newtons model to do most of the stuff involving gravitation. You don't find us using tensor mathematics for simple gravitational problems in 'earthly conditions'. The same occurs in most market conditions where we use 'Gaussian models' where calculations are easy using a bunch of linear mathematics. The reality however is that markets are far more non linear than we appreciate....and it is this non linearity where the Pareto Principle kicks in. This is where 'most of the money is made and lost'. We just fluff around in the less material linear stuff thinking that we are doing well....when in fact what matters most are the extreme non linear times. These are the events that shake out the species. We can all survive a degree of variability....but it is the extreme event were the wider system has the last laugh. In a battle zone it is often the narrowest of margins which decides whether you are alive or dead. It may be the bullet that grazes your head to appreciate the importance of the small perturbation in a model. We tend to attribute this to luck, when no such beast exists in a classical universe. It is all deterministic....but very....very complex. A rounding error based on a supposed assumption can have dramatic world shattering consequences when you think in a non linear way in a classical (not quantum) universe.

Fortunately at the visual level, our brains can detect non linear relationships of curves and wiggles hence are quite well equipped to deal with fat tails and their non-linear implications however apply a single statistical descriptor of an entire equity curve and you totally miss what the equity curve is telling you unless it is a perfectly smooth linear a perfectly smooth curve. Complex systems like nature itself is rarely euclidean. This is a rare condition. Complex systems rely on interrelationships and 1 to many relationships....and hence prefer non linearity.

If you are after predictive repeatability and prefer simple models full of linear assumptions you are convergent. If you are non predictive, appreciate system complexity and overall mathematical intractability and just defend your capital with divergent outcomes then you are divergent. As a divergent breed of thinker you wait until the market dictates when that event or events will happen. It forces you to diversify across instruments as this is the way to stay alive during the normal everyday churn. You simply rely on the principle that somewhere, someplace there may be an anomalous event that creates abnormal momentum surges in an efficient market context.

As retail traders we like to think we have an element of control in the market where we can predict outcomes based on past patterns....but there are not many stats you can find that are audited and validated that can demonstrate this over the long term. If you are not prepared to be divergent then you have to be supremely confident of your capabilities to adapt and read the market with precision. Yes you will perform spectacularly in a single market condition. In fact we have had one pretty much from 2010 to now across most asset classes as there have not been many fat tailed events with exceptions like the Swissy depeg, Brexit and the Trump election etc....but markets have been low vol and very predictable with QE driving the agenda....but you do not need to be a rocket scientist to understand that this illusory stability will last in perpetuity.

I have a long term agenda which makes me have to find a way with divergence which does not have the sweet smile of convergence all the time...but when conditions are ripe, history demonstrates that it dwarfs the alternatives. I definitely am a rainy day player. My positions are always viewed from the perspective of the long term overall result and I let the market do the talking....not me.

I personally feel there is an easier way and a less demanding way to play this game....that has a history of market behaviour (aka science) on its side.

Thanks DitD. Good food for thought in your post mate

So we have mined some of the rock in an area we think is gold bearing, let's conduct an assay of the sample using Monte Carlo with visual as opposed to statistical methods.

Attached Image

When the word Monte Carlo is thrust upon the uninformed bystander...they may shrug their shoulders and say...."I am not a Geek...what possible use is it for me? Well the bottom line is that you need to understand what it does to derive meaning from it....and it is certainly not a tool simply restricted to the Geeks amongst us.

In essence a Monte Carlo process is a means of reshuffling the output of a dataset in random ways to see if there is hidden meaning that emerges from the method of testing. It is a way to distill what is noise and what is real....or what is curve fit and what is not.....however while many may be stuck in the land of using single metrics to describe performance, they may not glimpse the power of the array. To understand what the array means.....you must assess the entire context of the array.

So.....here is how I use it to inform me about the sensitivity of the strategy to random changes to major assumptions used in the model.

Firstly I set the validation pass criteria of all runs conducted by the Monte Carlo method to a very high pass rate. This ensures that within the potential variability of outcomes produced by the array, that an edge is likely to be found in the strategy that is being tested (aka there is a high likelihood that gold exists in the sample). For example in a run of 500 arrays, I would set the validation criteria to 90% passing the validation criteria I have setup in the Tools menu.



This very high validation criteria sorts out heaps of chaff from the grain. Only say 5% of strategies that enter the validation phase of Monte Carlo might pass this strident test.

Now how do we get important info from these validated samples that have passed this strident test to assist us in our search for the real gold?

The outcome of any single strategy in the real world of no simulations produces a single return distribution (eg. an equity curve). One of the strategies that passed the test that we want to assay produced the following result after having been treated for Validation tests and Optimisation.



The single equity curve is one of many possible altered paths that occurs if there is a small perturbation in any of the market conditions that might influence the underlying strategy. Given that we have optimised our strategy, you can be guaranteed that it is the best path for the exact specification of historic data that has been applied to this test (the 'state' of the historic data). Now this arrangement of 'state for a single static solution' will not occur in the future. It may be a similar 'state' but there will be variation to it (maybe in the smallest degree), but enough to guarantee a different future return path.

The Monte Carlo array is a method to understand the degree of sensitivity that exists in a single return distribution by shuffling the underlying variables within broad constraints to observe the degree of 'sensitivity' the strategy has to future uncertainty.

There are many different methods of Monte Carlo shuffling, and EA studio gives quite a broad range that allows us to get some very useful info out of it.

So let's start with the single equity curve above and see where in plots in various Monte Carlo arrays....and understand what it is telling us about the characteristics of the strategy.

Shuffling History Data through Monte Carlo

Attached Image (click to enlarge)

We compare and contrast the actual single equity curve of the strategy against the Monte Carlo array to see where it plots and make some conclusions that may be right or wrong. As we progress with the process we will get more rights that wrongs that assists in confirming our assumptions.

The result of the actual equity curve produces an end result that plots at $15.6K commencing from $10K. If you shuffle the data history you will see that this equity curve result plots in the upper zones of the MC array. If the data history was simply rehuffled in segments (*fine print *assuming a bootstrapping technique that comprised segments to contain some degree of autocorrelation) the impact of this reshuffling demonstrates the sensitivity of the strategy to different auto-correlated arrangements at different times in the dataset.For a divergent trader, this is well understood as periods of divergence depend on particular anomalous conditions and lie anywhere in the data history. This variability of data history segments creates variability in profitable or unprofitable sequences of trades. We therefore expect variability from this data history reshuffling. It is not so nice for a convergent trader who relies on predictable repeatability in the market condition.

Now look where the equity curve plots against the most likely outcome if those past conditions were reshuffled. This is our first indication of the degree to which the equity curve of the single run is 'curve fit'. Now we expect following optimisation for a curve fit result to arise, so it is not a real problem provided we water down our expectations regarding the future profitability of the strategy. As opposed to producing a $5620 return for the period, let's use $4,550 as a more likely estimate provided those similar conditions of the past were presented to us in the future again.

Let's move on to get some further useful info using the Monte Carlo tool.


Randomising Spread through Monte Carlo
According to popular legend on this forum, the brokers are out to get you and the tactics they employ are enough to turn this game into a negative sum game. Let's look at what the data thinks of this notion through Monte Carlo. Let's see what happens when we randomise the variability of spread conditions on the overall equity curve outcome.



There you have it guys. Randomising spread is the least of our worries for this divergent strategy. There is no real signature curve fit element to this notion.

Moving on......

"Oh yeah Copernicus....but what about slippage....that is the biggest culprit".
"Is it?"..... is my reply? "Let's test it first as that might just be another assumption. You may find it is important for scapling....but not for this particular divergent approach?".




Randomising Position Entry through Monte Carlo
Now remember under a divergent philosophy we place a preset condition across all our strategies to only activate a strategy during extreme times. Let's test this assumption by randomly varying the position entry by removing this constraining condition.



Case solved Mr Holmes. Randomly varying the preset filter or the entry execution rule has a big impact on the outcome. This therefore infers that our assumption might be correct...that divergent strategies that trade all the time in any market condition fail over the long term. Filters to 'turn on' the strategy only during extreme market conditions perhaps has validity to its logical assumption.

Randomly Skipping Position Exit
Ok so we have trades running, what is the impact of randomly adjusting the exit condition within limits? Not much is the answer which is quite surporisjng bnut there is a caveat here. The MC randomises the variables that exist in the strategy that impact on the exit. One of the variables we have in every strategy is a trailing stop, but the preciseness of the trail does not appear to have a significant impact. WE would need to remove the trail to see if this assumption still held with MC.



This is confirmed by a random MC applied to close the position. You will see that this is where variability exists.....so exits are important rather than their precise positioning. If you occasionally miss some closes, you suffer....hence there is a degree of curve fit nature in the overall outcome as execution issues will occasionally apply.



...more to come

Intermission



C

Monte Carlo.....continued......

It's all just spaghetti man.........

Attached Image

Hopefully we now are starting to realise that non-quants can play this game of sample testing. The good old visual method and a bit of critical thinking about the nature of your strategy outclasses and outperforms the incorrectly applied simple maths that is typically thrown towards non-linear complex systems. The old adage of 'quality in and quality out' applies. Adopt a simplistic model ridden in inaccurate assumptions, and no matter how you apply the simple maths....it's meaning is not an accurate assumption of the underlying reality.

Now we are getting down to the important reasons for dispersion and curve fitting where there is heaps of info you can glean from the MC array and this lies in strategy variations themselves, namely if you vary your indicator parameters used or your backtest starting bar.

Randomise Indicator Parameters


This is the beast that you need to tame and holds important info about the robustness of the strategy and the curve fit nature of the resultant equity curve. Some of the posts have already touched on the importance of this test in terms of parameter stability in a landscape of variations.

Hidden within this morass is what curve fitting really means and the variability to the resulting return stream from it's presence. Ideally we want a strategy that is profitable whenever market conditions are divergent and also that does not lose too much dough when conditions are convergent (predictable).

You will note as you peer into the morass of spaghetti there are periods where all the lines move synchronously up and at other times, they splay (disperse). We get into this further down the track of this thread when we map this MC array to the market condition to ensure it is divergently positioned to capture the fat tails, but what we need to understand is this............

The host of variables in your strategy need to respond to an array of different market conditions during periods of convergence without detrimantally affecting performance too much. Having a stable landscape of different variations of parameter applied in your indicators tells you whether the solution is divergent in that all the variables irrespective of value do the same thing when markets are divergent..but all do a different thing when markets are convergent but not too much to cause mass dispersal of outcome.

If you find in the MC array this tendency for a coordinated movement in the same direction at the same time which maps to a divergent market state when you align your array with a chart of market data....then you have a divergent solution but robustness is about how well it does when conditions are not favourable and how much it is prone to dispersal (an array of different outcomes).

The variation in variables during divergence are all positively correlated. They all work in other words during different divergent states.....which is a good thing. So if you consider the MC array above, the times when you see this concertina of positively correlated movement, it tells you when the market is divergent. On the other hand the dispersal of return streams that give rise to performance variation occur during periods of convergence and your return paths can go any which way dependent on the variables chosen which is decided by luck alone.

A return distribution for a divergent solution that is highly profitable during periods of market convergence is either through luck or curve fitting. It is not an outcome of the divergent design of the strategy. This is where drawdowns occur within diversified divergent portfolios.

Now we actually do not care about the convergent mess of dispersion provided it is not too extreme in this array as we treat for robustness at the portfolio level through a principle of diluting dispersion of non correlated portions of the equity curve through superposition of different return stream profiles.

Look however at the dispersal between the end states of what can be attributed to realistic divergent movement versus convergent cases. This disparity between where your equity curve plots and the overall average of the possible routes taken tells you the degree to which your solution is curve fit. So water down your future expectations and do not rely on the singe result of the backtest....it is curve fit I tells ya!!!!!....but still has a kernel of real value in it. Don't throw it away....just take heed of this system 'risk' weakness. At the portfolio level, this dispersal variation is all sorted out.

Randomise Backtest Starting Bar
This is a big issue with divergence as during periods of stability there is not a 'drop to drink'. Where the anomalies occur matters....and this can never be guaranteed. If you enter too late in the game after a big event like the GFC.....you may be underwater for many years before the chances of another fortuitous event of divergence where you can quickly recover from your drawdown. No surprises here. Where you decide to start has important consequences on the outcome. This creates significant style drift amongst divergent managers but does not necessarily mean they are bad managers. Luck still decides a large part of your success in the divergent space.



Ok guys that wraps up Monte Carlo...but hopefully you can see that it is a very useful tool that you can apply to evaluate a strategy in 'visual ways'. There will always need to be a degree of qualitative judgement made about whether a strategy is 'good or not'. At least this approach is a rigorous way to make that judgement as opposed to a knee jerk thoughtful assumption about how things work.

The next post is going to discuss the MSAccess portfolio as we need to get ready to accept data from our testing

Stay tuned.

Shift over....down tools




C

Okay Nick, your testing has dispelled the specter of broker costs.

How do we trade with it?

Hi J

You must assess the importance dependent on the nature of strategy you employ. For instance if you have a scalping strategy the MC array should be able to give you info about how sensitive your equity curve is to spread variability and where your curve plots in the context of the array. If it plots along the top of the array, you can be guaranteed that it is curve fit as *if spread varies from the history of spread that you have conducted your test on*.....you will find you plot in the guts of the array well away from the supposed backtest path. So then you need to assess how important is spread variability to your overall result. Am I in the game to predict spread variation or market condition? You need to then critically examine your philosophy and objectives and decide whether the outcome you see from this test is what you are after....or not.

In my world of divergence in D1 and H4, variation of spread has bugger all impact as displayed in terms of the Law of Large Numbers. Brokers have a limited say in my overall result. My results are more contingent on market condition than execution issues.

Just food for thought mate that might help some ...but not all :-)

Have a closer look at the entire context in terms of what occurs during bull runs versus bear runs in equities. Look at the S&P500TR Index return compared to the total Index of CTA's. Look at what we have experienced since QE and the progressive split into extended convergent and divergent phases. Look at the overall impact on the lines of best fit.

Importantly look at the risk weighted return of the S&P Total Return Index (includes dividends). A CAGR of 5.87% for the 18 year period and a max draw of 50.95%.



Now compare this to the performance characteristics of CTA's that are diversified and divergent players. I have included the S&PTR Index equity curve for comparison. A CAGR of 9.07% and max draw of 16.97%



I have annotated things here to make my interpretation clearer. of what I mean by different market phases of divergence (trending) and convergence (mean reverting).

Now have a look at a particular class of CTA which is where this thread is heading as within that group of CTA's that comprise the index, there are different classes or groups of performers. The ones we are focusing on are the risk-weighted performers. Compare their risk weighted performance to the S&PTR Index. A CAGR of 8.91% and a Max Draw of 7.09%.



If you have a different opinion then that is just a difference in how we interpret what we mean by long term and overall sustainability and volatility of returns.

I think you may find that he is re positioning his funds towards those styles that attract allocation. He is not saying trend following is dead but that it is overcrowded.

When I talk about the style drift in CTA's, the response from Harding is exactly what has been happening in the industry. A style drift away from traditional cross asset diversification towards systematic diversification. If you read up on Clenow and Harding and the likes, they sculpt their equity curves to be attractive to allocators as FUM is their bread and butter and to be attractive you need a different correlated offerring to the 60:40 equity bond mob. They do this by system design as simple diversification across asset classes and timeframes cannot guarantee a desired solution.

Anyway...thanks for the input but it hasn't dampened my enthusiasm about the path I am taking in this project.

Those charts are not of the 70's, 80's and 90's where CTA's did even better due to preferential interest rates on dormant cash.

Those charts are from 2000 to now. The TF Index smashed the S&P in 2015 and 2016 from a divergent phase....also in 2010 and 2011. So occassionally it outperforms between 2010 to now. What has changed however is the frequency of divergent conditions and the swings between divergent phases and convergent phases post 2010 have been more protracted. Notably probably by huge QE whose impact has spread across asset segments and as a result, every asset 'and there dog' are becoming more correlated with this loose money. There is also probably an impact from the swathe of automation that has entered the markets with HFT and convergent algo's on the short timeframes. Not surprising then really that times have been tough for traditional trend following over the entire period post 2010. So the traditional CTA that was 'traditionally diversified' has suffered as a consequence.

If the market **** itself again..then the correlations will break and once again the traditional divergent solution will work again. Until that happens, you need other solutions to stay afloat....hence system diversification. We will still apply cross asset and timeframe diversification like the traditionalists for when or if that event occurs...but in the meantime we will manage our drawdown profile through system diversification.

Past performance is certainly no indication of future performance so risk management needs to be on top of the agenda...not in last place. Searching for performance is not the way to do this as survivorship bias ensures you have a limited investment lifespan. For example I can guarantee that those 64 CTA's that I track have been in existence since 2000 to now. Can you guarantee that the composition of the S&P500 Index has not changed over the same period?

PS My investment amounts are my own business and no one else's. I am not a commercial member.

A Brief Overview of the DBase

Attached Image

Guys we are getting closer to the point where we need to start using our Dbase to house research generated from this process. I thought I would give a quick overview now as unfortunately I am out of action tomorrow but will return to continue things on.

Given that we need our organised wits about us in this project and the need to remember why we selected particular options etc., a database is pretty essential as it ensures that the valuable research we do and the decisions made are always retrievable in an efficient manner. We therefore use the database as our storage tool for this entire project to house research, EA's, strategy performance results etc. We don't simply store the files in folders on our PC as we need to identify relational dependencies that matter.

This database which you can retrieve from the opening post is version 1.2. Original versions have been updated since commencing this exercise. It is a bit deceptive as the simple menu and forms pack a bit more punch than what you might expect so I will give a brief overview here of it's capability.

When you extract the dbase from the rar file you store it somehwhere in a dedicated folder for this project. You need MSAccess2013 or later to open it.

When you double click to execute the file you will be presented with this screen. It is not beautiful but it does the job.

Main Menu


We have a 3 level relational hierarchy in this database of tables which commences at the portfolio level. We can have as many different portfolios as we like. Each portfolio comprises a collection of strategies. Each strategy comprises a collection of trade results (eg. different broker sources, different time ranges of data etc. )

That should all make sense.

The analytics button is where we go to play around with the strategy configurations in our search for different configurations that we would like to export to our analytical spreadsheets.

That is about it....so lets have a look at the Portfolio button when we click it.

Portfolio Views
We are presented with the following list view. As you can see, there is a sample portfolio already created in it.



There is powerful search capability on all fields to restrict the list to our specified criteria when things get huge...as they do. These restricted lists all default back to original list view when you leave this option and then return.

We can enter and amend details here or create new portfolios by adding lines...or if we double click on the relevant portfolio line, we will be presented with a different screen view.



In this view we can see the strategies that are attached to the portfolio and other fields relevant to that portfolio.

Let's return back to the list view as this view by clicking the close button. I like the list view as I like to have control over my lists. I am continuously in this view playing with search criteria to restrict lists to limited options......anyway I digress........

Let's add a new Portfolio and call this 'Copernicus Divergent 1", make it active by clicking on the active button and filling in other details. We will make it effective from today.

Now I make each portfolio specific for the data mining settings I use in EA studio.....so remember when I said that we can save a setfile for the EA Factory settings that are currently being applied, well I save the setfile produced here by double clicking on the attachment and finding the location of that PC and embedding it in the Dbase. When you double click on the DataMiningSets attachment link you will be led through a process to find the file and attach it. You will see that there is now one attachment in the field. If you double click on it you will be able to retrieve the attachment for use elsewhere. This feature is virtually unlimited in the number of attachments you can embed so go for broke.

The collections field is where we save collections generated by EAStudio which contain hundreds of individual strats. This is how we at least retain an entire collection of strats produced by EA studio and re-refer to it when we are looking for more solutions than we presently need.

"Other attach" is what is says.....for any other relevant attachment that is connected to that portfolio such as your favorite music clip while you were creating your portfolio or other important attachment.

Ok so we have created a new portfolio to house the strategies we are generating from EA Studio. Now it is time to load a strategy from EA Studio. Close from ther list view by clicking on on cross x on the right just above the list then you will return to the main menu. Then select the button Strategies.

Strategies
You are presented with the following list view of all strategies in the dbase and which portfolios they are connected to. As you can see we have 3 strategies in the dbase already (samples only) that are all connected to the portfolio "Copernicus 15".



If we double click on the line you are then presented a detailed strategy view (like the portfolio form view we looked at earlier. I have double clicked on the strategy with the Magic number 1338475 and am presented with the following screen.



In this view you can see what trade history is linked to this strategy and from what source and timeframe it relates to.

Let's return to the list view by clicking on the cross x just above the field 'Long and short' at the right of the screen.

Now let's add a new strategy that appears to hold promise from our data mining activities. In this case the strategy is the EA name that is generated by EAStudio. By convention each strategy is named according to the following syntax in EA studio. We like to keep that syntax for times when we need to return the strat to EA studio for further work. " Robot Factory EURUSD D1 33099827.mq4" The syntax contains the EA generator name "Robot factory", the instrument that was data-mined "EURUSD" and it's timeframe "D1"and the Magic number of the EA of "33099827". As this info is supplied by the naming convention of EA Studio we apply this info to the fields of the strategy. We also enter other details in the relevant fields and include the EA code as an attachment.

Note that the prior samples were executable files ex4....disregard this as we prefer to retain source code so we save the files with an mq4 extension.

We also want to link this new strategy to the new portfolio we created so select this portfolio from the list box in the field "PortfolioID". There are other fields we can discuss later but we are not there yet. This is just an overview to get you familiar with the workings.



Intermission




C

When an animal is too colorful is it often poisonous. When the package is too colorful is it often marketing :-)
Would you really eat those pastel colored spaghetti? Can I really trust this Monte Carlo thingy?

Let's have some fun. Here is a bunch or hypothetical returns (understand dice rolled values. Excel did roll the dice)

0.495218562,2.171557788,-2.271068689,1.011278942,0.343697057,0.80772178,0.211672564,1.597438462,0.040537333,-0.677360292,
-0.597352413,0.901665311,1.547482498,-0.083209742,0.134378452,-0.749661126,-0.087777907,-0.294070084,2.3148578,-0.549110626,
0.570290691,-0.221723313,-1.076875338,0.376181544,0.655302243,-0.578411364,0.864651754,-0.622366409,1.35523967,3.212077267,
2.332169182,-1.325774749,-0.015232649,-0.395786003,0.416298897,-1.150960228,-0.02222116,-0.180763303,-1.413097717,
1.318575743,-0.647017956

population............. 41 samples
sample mean......... 0.2370354
sample var............. 1.283465
sample skewness.. 0.4382471

The mean estimates the edge. The variance is an indication of risk. C is gonna be happy skew is positive. Let's plot that. Beautiful, isn't it?


Let's resample these returns (random draw with replacement) and let's plot a few of the equity curves that produces. The red line is the original above. The black one is the average of the resampled curves. All good they are all going up. Test passed! check! Happiness. Music. Beer.


Unfortunately this is all Fool's Gold my dear! The actual distribution is the good ole Normal N(0,1). The real mean is 0, aka no edge. The real skewness is 0 (symmetrical). It just happens that I chose a curve that was going up. The bootstrap failed at recovering the true distribution.

--------------------------------------------------------
Let's now pretend the data is market returns instead. Let's model the market with a 25th degree polynomial. Needless say it can only be overfitting the dataset. 41 returns make 40 segments. I can cut in 4 pieces of 10, shuffle and stick. There are 24 combinations. I plot 6 of them. See how my model fits the market!


I could have as well underfit the dataset by simply fitting a line: market is just going up. It would have shown the same. Buy and Hold is great mate! Just look at my Monte Carlo validated backtest on Netflix, Inc.! (can somebody please write me a buy-and-hold EA?)



Why does it fail? Simply because there is a random trend in the dataset which is present in each of the returns. Whatever the way you rearrange them this "trend" will appear.


If it is so easy to fool a trivial toy example how can I trust a model (trading strategy) based on a much more complex system (the market) with so many more degrees of freedom?

Pip's in da house....listen up y'all. You're about to be thwacked with a numberphiles hammer. :-)



It is possible to make spaghetti appealing. Here try this nice ordered spaghetti using this special purpose spaghetti fork. Normally $100....today only $24...but call now. Offer doesn't last!!!!

A Brief Overview of the DBase.....continued



Stick with it guys. Unfortunately there is always some pain in your R&D efforts. We are almost done with the database for now..but we have to leave this post on a happy note. Hopefully you are starting to get a feel for the underlying structure of it so you can apply some whizbang search criteria as we mine deeper into this bedrock.

Trades
Select this option from the Main Menu and you go to the trade data history of your factory. I have supplied an excel tool we use to format data in a standard way from your MT4 Strategy Tester or the FxBlue environment to quickly dump data into this repository. You will see a live example in action in future posts.

The structure of the trade data history is mapped to the Magic# and we define the type of test undertaken, namely whether it was a backtest, walk forward test or Live test. In this project we define a backtest as any insample test undertaken by a data mining operation, a walk forward test as any out of sample test conducted in a demo or simulated environment and a live test....well that's a live test with real money on the line. :-)

We also define the test method such as open prices only, control points, tick by tick or live environment. The rest is pretty straight forward. The Deposit$ amount is the deposit amount you establish in the MT4 testing environment. We have a different way to calculate allocation towards strategies at the portfolio level and when live, this Deposit$ amount isn't used.

For the purposes of this post I have added trade history for the same strategy we just created from both the Pepperstone Demo account using open prices and the Dukascopy Demo tick by tick data source that uses Birts Tick Data Suite. (sorry guys....you might need this tool as well if you don't have it already). At this mine we need lot's of tools.

Analytics, Selection Criteria and Excel Dump Options
This is the guts of the database where you spend most of your time compiling different options to throw into the deeper analytics that are compiled in a spreadsheet (the analytic tool I have already supplied to you). You can access the portfolio or strategy details by double clicking on the relevant portfolio or strategy line and get further details that are missing on this analytic summary.



Lets have a look at each section.

In the portfolio details on the left hand side, if you want to exclude a complete portfolio from your analysis then untick the 'active box'. For example I want to have a look at the details of the strategy I just created so I untick 'Copernicus 15' as this relates to a different set of strategies. I then hit the 'refresh' button in the title of the form.



The result summary then reduces to only those results emanating from 'Copernicus Divergent 1'. If you use the dropdown list of 'select broker' and then hit the refresh button again you will find two data sets in the analytics. That data set from the tick by tick run from Dukascopy and that resxult set from the Pepperstone Demo run. This was the trade history I just added for this demonstration. Note the difference in the results....this is where things get interesting. We now have a quick method to sieve lots of data and find out what it is telling us.

Now place a tick again on the portfolio 'Copernicus 15' to resurrect all the trade results in the database. Let's restrict the list to a few strategies from two portfolios as we want to assimilate strategies into another new portfolio after we test them in detail.

Go to the strategy section on the left hand side and let's say we select Magic 1338475 and 33099827. We don't want to look at the others so we inactivate them. Hit the refresh button....and hey presto



You can start seeing the power of this beast when you play around and it is super fast in crunching alternatives.

The whole purpose of this database is to assimilate compilations to thrust into your analytic spreadsheet tool to do heavy crunching and lifting.

Once you have a list that you are happy with it's compilation, you then hit those buttons at the top 'Analytics Dump', 'ReturnsDump' and 'TradeDetailsDump' which will quickly compile this compilation for export into your Analytics tool......but that's for another day.

Shift over again guys



Be back in a couple of days. Unfortunately will be offline tomorrow. Need a break anyway as I am starting to foam at the mouth.

Cheers

C

PS Please update your tools we now are at version 1.3.

It loads from their server, but all calculations it does are done locally on your CPU / RAM (check your Taskmanager and you´ll see that once you start any processing, the CPU load and RAM usage of your browser increases) - it´s Javascript, those are always run locally. It´s not multi-threaded however, so you can benefit from running multiple instances of it in separate windows (not tabs as they will otherwise be background-suspended by modern browsers). I have a 16 core CPU, so I ran 16 instances of it to max the CPU out.

Almost:-)

Agree with some of it....not all if it. Look up the definitions of convergence and divergence.

Science is about empirical methods. Models to describe observables...not whether those observables are necessarily repeatable. Have a look at Chaos Theory and its applications in science for example. Try and repeat the behaviour of the mechanics of the double pendulum which is frequently cited as displaying chaotic behaviour in non linear dynamics.

Thanks for this info GT. Very helpful :-)