My progress as a full-time investor & trader

Dear mim2005,

also being a Hanover's "fan" I can see the kind of rigorous and wonderful work you've (you both have!) done.

This is a good demonstration of efforts and serious procedure, but, of course, I am all ears here for the "final teaching".

Waiting for that!

As Mim2005 has stated, I helped contribute to the design of these EAs. While waiting for Mim2005 to comment further himself, here is my own take.

There are certain concepts that I take care to avoid, when it comes to designing trading systems. Many of these are listed here. Specifically, in the case of these EAs:

#1. Curve fitting. It’s my belief that one should start out with seemingly plausible trading concept(s), and then use all available data to test their viability; rather than attempt to manufacture concepts from mining the data itself — a subtle, but vitally important distinction. This is the best guarantee that I know of, for developing robust systems that avoid curve fitting. Avoiding the use of indicators (MAs, MACDs, stochastics, etc), which require calibration, also helps to reduce dependency on undue (or unwitting) optimization. Both the entries and exits in Mim2005’s EAs rely more on the use of broadbrush principles rather than accurate math.

#2. Money management (position sizing) can NOT provide an edge in itself.
I strenuously avoid the use of the all of the following:

1. grids (evenly spaced levels have no technical or statistical basis, i.e. are not recognized by the market),
2. averaging down (the famous Paul Tudor Jones pic says it all!),
3. any focus on recovery from recent losses (recent losses have no greater effect on eventual bottom line than any other losses),
4. hedging (same-pair hedges can be replicated by a single net position, and with lower costs),
5. baskets (can likewise be broken down into simpler net positions),
6. variable or progressive staking (e.g. martingale variants — these merely trade guaranteed short term recovery for eventual ruin).

Conversely, all three EAs make rigorous use of hard stoplosses, and employ consistent, fixed fractional position sizing (X% equity risk per trade).

#3. Consistency of result: while low drawdown (and hence smoothness in the equity curve) is desirable, and can justify more aggressive sizing (a higher value of X%), I’m happy enough provided that each period of drawdown is always followed by a new equity high. At every point in the curves, the account equity is higher than it was 6 months ago — for me, that is a satisfactory measure of consistency, given the way that long term trend following strategies operate.

#4. Understanding system strength and weakness: Following on from point #1, IMO it’s important to understand under what market conditions a system prospers, and under what conditions it is likely to struggle. Then, with regard to having confidence in a system, it becomes a case of satisfying yourself that ‘favorable’ market conditions occur often enough to allow the entries/exits to provide an overall edge. While there are never any guarantees with any system — markets can do whatever they please — this at least provides some assurance that past results offer some indicative range as to what the future will probably hold.

In the case of these EAs, all of them use trend-following type exits, that are broadly based on the ‘infinite yield’ concepts employed by systems like ‘DIBS’ and ‘millipede’. In lieu of choosing pairs based on currency strength/weakness analysis, the EAs simply trade EURUSD, because EUR and USD tend to remain negatively correlated (creating daily/weekly trends) for a reasonable amount of time. Hence, for as long as macroeconomics maintain this negative correlation — if EURUSD ‘trends’ far enough, often enough, as it did over the 6 years of the backtest — then the EAs should remain profitable. Whenever EURUSD goes into prolonged sideways patterns, losses are the result. That’s what I mean by a simple system SWOT analysis allowing a pragmatic outlook about prospective results.

#5. Allowance for costs and inaccuracies in backtesting: I acknowledge the problems and inaccuracies involved in backtesting, but I believe that these can be alleviated somewhat by (i) using quality tick data (Mim2005 has used data from Dukascopy; note the 99% modeling quality and no mismatch errors in the reports), (ii) building in a generous margin for error for spread costs (these tests have been run on EURUSD, for which the EAs allow 1.8 pips), and (iii) trading over longer timeframes, where inaccuracies, costs and execution vagaries (slippage etc) become relatively less significant. I agree that backtesting, like any other form of testing, provides no iron-clad guarantee of future profit, but given the above points, I believe it can offer a fair indication as to whether a system stands a reasonable chance. I also believe that any “backtesting is worthless” comments are likely to be partly indicative of the nature (scalping? curve fitting? indicator-based signals? dodgy MM? etc) or impotency ("saddle blanket" concepts?) of the systems being tested.


EXITS: All 3 EAs use the same exit algorithm. The ‘infinite yield’ objective is to maximize profit while conditions are favorable, and this means letting profits run as long as possible, and adding more entries (on valid setups), even if it means forfeiting part of that profit on all component positions, during retracements along the way. Those who are familiar with strategies like DIBS and millipede will know that they seek to find low risk intraday entries, and allow those that survive the opportunity to catch longer term trends that occasionally develop, for very high RR trades. Joel Rensink, the inventor of DIBS, explains this further here.

ENTRIES: The goal of the entries is to kick-start the trades in the right direction, and thus avoid early stopout, improving the probability of survival into the medium term, after which the (trend following) exits take over the management of the trades. Each EA uses a different intraday entry mechanism. The first EA uses a quirk in price movement at a certain time of the day; the second uses an ingenious breakout strategy devised by Mim2005; the third uses a confluence of key levels that appears to be recognized by the big players. The third EA has a lower profit factor, but compensates because of its ability to find a higher number of suitable setups, and thus generate a greater turnover of trades. I’ll leave it to Mim2005 to elaborate further on all this, if he wishes.


For me, the biggest challenge with EAs is to somehow translate human trading knowledge into a mechanical algo, without too much of that knowledge being eroded in the translation process. There are other great concepts that I’d like to experiment with: pair selection based on currency strength/weakness; supply and demand levels (as popularized by Sam Seiden); orderflow-type ideas; tell-tale intraday patterns created by heavyweight player agendas; news trading; multiple time frame analysis — these are a few, just off the top of my head.

Despite their inability to apply intelligent ‘contextual’ analysis or adapt to different market conditions, I like EAs, and for all of the obvious reasons. They provide objectivity and detachment; accurate, consistent, emotionless, around-the-clock execution; they’re never afflicted by illness, stress, tiredness, nerves or negative attitudes; and they allow the ‘trader’ to spend time away from the computer, both increasing leisure time and greatly reducing personal involvement.


I’m not trying to prove anything with this post; rather, I hope it gives other system developers some ideas and guidelines to consider and experiment with. Mim2005 and I have been collaborating for only a short while, and our R&D is merely a work in progress. Let's face it: all these test results demonstrate is that certain entry/exit concepts would have provided a fairly solid edge over the last 6 years, and that there’s no obvious reason why their effectiveness shouldn’t be totally eroded in the immediate future. I don’t claim to be a profitable trader — certainly not yet. I simply accept the results for what they represent, nothing more, nothing less.

Finally, I just like to say that it’s been brilliant working with Mim2005. As the saying goes, two heads are always better than one, and the task is certainly made a lot easier when your trading buddy holds similar trading philosophies to your own. Mim2005’s IT background has been a huge help; his ability to code MT4 EAs has allowed me time to further the development of other ideas and projects.

David

Footnote --- Other References:

(1) IMO this is an essential read for anybody who's contemplating the development of automated trading systems. What went wrong?
"The lesson here and my message to you is that no matter how much you test, no matter how far you go back there will always be a market condition that your test does not have."
Again, this highlights the need for risk management. However, it seems that 4 months' worth of drawdown obliterated 40 months' of consistent profit. Now this can occur only if either position sizes were 10x larger during the drawdown, or trade setups occurred approx 10x more frequently. I note this comment:
"the bigger the drawdown the better the system works - thats because of a very clever money management strategy which i cannot reveal."
Sound like a recovery MM? I suspect that it was the "clever MM strategy" that contributed to the downfall — another good reason to stay away from recovery based MM. Consistent position sizing is imperative.

(2) Prolific author Daniel Fernandez' blog posts offer some great insight into the dos and don'ts of mechanical trading. I will comment on some of these when I get some spare time.

(3) There's also some outstanding info on system development here and here.

What about mine? Pretty new, own developed i like to call it the holy grail

Whats the MAR?
i 6doubled the account in 86trades.

MAR: see the link in post #15. It's annualized return divided by worst case drawdown. Here is the formula that I use (plugging in the values from your screenshot):

years=???;
deposit=10000;
netpl=43034.40;
maxdd=27.72;
annualroe=(((deposit+netpl)/deposit)^(1/years)-1)*100;
print "annual return =", annualroe, "%";
print "MAR ratio = ", annualroe/maxdd

Congratulations on what appears to be a very good start (but MT4's report doesn't provide a lot of info on MM and risk management, both of which are crucial to long term success). Come back and let us know how well it's performing after 860 trades.

You know the market constantly evolves and tomorrows data is not the same as yesterdays right? Your best backtested parameter set could easily go straight to worst case scenario drawdown going forward. You cannot escape risk this way. You cannot trade the market without the risk of losing

Hi MAC Doug,

Thanks for your comments, the main intention of the post you've quoted is more to give some insight into the amount of effort I went into when I started my study of technical analysis. As I mentioned I only found I could develop somewhat profitable backtest using those more traditional TA approaches and I didn't feel comfortable that they were necessarily going to be profitable in a live environment - hence why ultimately I stopped with those type of approaches.

Completely agree in regards to markets constantly evolving and the risks around optimising parameters. I also agree that accepting losses is the only way of becoming successful in this game.

The EA's I posted results on in earlier posts don't use this traditional type of technical analysis or indicators. As mentioned before I really don't know if they will be consistently profitable in a live environment. So far I have only been running them on a live account for a little over a week and there up around 10% (using 0.5% of account risk) but they got lucky on the tapering announcement.

I will make the trade explorer public ... although I fully realise these EA's could crash and burn but I will be happy that I at least gave it a go - never be scared to fail.

The 3 live EA's discussed in earlier posts: http://www.forexfactory.com/mim2005

NOTE: The objective of this thread is not to focus on these EA's as they are only a small part of my overall investment and trading - I pretty much consider this an initial test of some of the work I've done in forex automated trading.



Take Care

mim2005

mim2005,
hanover,



Well done. I know you know I am a long time estimator , but this sort of "assay" is really very well done. I am sorry not to be able to vouch you both, yet (for what it might mean). Thanks, Hanover, for the plenty of links to reference material, sort of "your path into this whole" that can be taken as a possible roadmap or blueprint.

I perfectly agree with all the QC criteria listed for the backtesting. I have the feeling that most detractors of historical testing here around are not even aware of the possible pitfalls, that you here somehow rule out given the adopted criteria that you mention. Professional work, congratulations.

I have a question concerning the backtests: have you tried some Monte Carlo simulation of them? You cited Daniel Fernandez, and you know that, within his members section, there are quite nice tools for performing such kind of test. It could give more reliable results of expected drawdowns and profit potential.

Always thinking about Daniel (yes, I am a member of his "asirikuy" community), the question he's currently focusing on is finding a method for correlating in sample results with out of sample (live) performance. The way he's making it is quite radical, surprising, and sometimes shocking. His last post on the public blog speaks for itself. I will be very interested in following the evolution of your systems with time.

There are also claims around telling that exactly these so nice-looking in backtesting systems give then, in real time, less returns than more "conservative". But actually I cannot find any concrete demonstration of this notion. Again congratulations and pips (actually: dollars!) be with you.

Bye

BL

Quality posts and a beautiful live trading start

I'll be following with interest. Thanks to both mim2005 and Hanover.

Hi Pip Commando,

At this stage in my investment/trading experience I lean towards agreeing with you. I have found programming EA's to be backtesting profitable exceptionally challenging.

Although I don't think it's impossible to develop EA's that are consistently profitable in live situations.

I have worked as a professional programmer for much of my lifetime and being a bit of a technology geek I have always believed that software can be designed to do some fairly amazing things ... it's just it requires a huge amount of time and effort to make it happen. Having said that, at the highest levels nothing can compete with the human brain.

In my own experience I currently live trade a completely discretionary system that I suspect is more consistently profitable than the forex EA's I have developed but I can't say this conclusively as I haven't got enough time under my belt to compare yet.

I think your approach of using a combination represents a good balance.

Take Care

I agree that it is not the quantity of backtests that are run that gives them meaning. Proper backtesting procedures aren't followed by the vast majority of traders. The problem is compounded by using deceptive tools such as the MT4 backtester. And thus those traders get poor results going forward. Overfitting is the most common culprit. It is very easy to do if you are optimizing a parameter that doesn't lead to consistency of results. Optimizing consistency generally solves this problem. There are ways to prevent or mitigate overfitting in system design and testing and more importantly to improve on the generalizability of the backtest. Diversification is one of the best methods to prevent single system failure but it is often underused or not used at all by retail traders. But generally speaking, this is an art more than anything and lots of disinformation abounds in trading books, articles and forums.

I think a lot of people optimize the past to create the most smooth / most profitable equity curve possible. Using an out of sample period to get some form of validation is a must but as others have pointed out this is not a panacea if the overall testing procedure or system concept is not sound. Until a developer can follow a particular method of system creation / optimization that produces profitable systems on unseen data of varying stripe each and every time, there is potential that the future has been overfit, that the process does not generalize well and thus for any given system to fail regardless of how the backtest looks. The design of a robust framework for system testing is no trivial matter!

Just some thoughts as to future topics I plan on spending some time on ...(not necessarily in this order)



EA related ...

1. Some specifics about how these EA's work (covered by Hanover to some extent)

2. Some of the indicators which I've developed (although none of them by themselves are a holy grail )

3. Some discussion around how I visualise the EA's tend to work

4. Risks in regards to the three EA's I'm currently using.

5. The two or three concepts that made the biggest difference to the profitability of the EAs I had developed (I think there will be a little gold here for some that are stuck where I was a few months ago)


Other ...

6. Comparing my discretionary trading vs. my mechanical EAs

7. Why I approaching investment trading using various approaches - not just forex

8. The psychological challenges I've had

9. Where to from here...

mim2005

SOME NOTES AROUND HOW THE EAs WORK
----------------------------------------------------------

Please note that I am broadly documenting how the current EAs work (without identifying their edges) ... but this doesn't mean that it's the "correct" way of working or that it's the best approach, simply what I have found works to date in my own experience. I am sure there are many ways to skin this cat and other EAs may be using totally different concepts more successfully.

Please bear with me as some of this will have already been presented by myself earlier or Hanover in his recent post.

A) CURRENCY PAIR SELECTION
The three EAs are designed to run on the EURUSD primarily because of the lower spreads available and in one particular case the importance of the session in which it trades.
Two of the EAs exploit edges that are fairly generic and I am fairly confident could be applied to work on other currency pairs. This means that there may well be opportunity to build further EAs for different currency pairs and hence leverage the edge concept for more profit. This is something I plan to do if the live EAs can be shown to be consistently profitable.


B) ENTRIES
All three EAs exploit fairly small intraday entry edges. i.e. at the point the entry is made statistically there is slightly more chance of price moving temporarily in their direction rather than not. The 'amount' of edge is small and last for a relatively short period of time. The entries are not in any way based on any sort of candle pattern, or what I would call traditional technical analysis.

The EAs use practically NO information in regards to broader context or higher time frame price movement etc. (i.e. trend detection, overbought oversold etc.).


C) MONEY MANAGEMENT
The three EAs all use simple fixed fraction of account balance risk. They all use stop losses placed at time of entry.

D) EXIT
The three EAs use the same concept for exits (although with slightly varying parameters in one case). The EAs are based on the concepts of letting your winners run as long as possible and the exit strategy is essentially a unique form of trailing which seems to provide much better results than pretty much any conventional trailing mechanism I have used and tested before (I have tried so many different trailing strategies in my time writing EAs and found the current one I use via a defect in my programming although I do understand broadly why it is effective.

The trailing stop mechanism provides a very wide opportunity for the position to go in the EAs favor i.e. They don't move to protect profits quickly and in fact a trade that is reasonably successful initially will often turn around and get stopped out for a full loss as the trailing SL hasn't even kicked in yet (even several hours or more later). At the same time it's not uncommon for the most profitable trades to last many days or week or more without being stopped out. This leads to a much lower % Win Rate but ultimately higher profits from the Risk/Reward (ala DIBS and millipede - which I thank for putting me onto these concepts).

Essentially the trailing exit is designed to catch as much of a higher time frame trend as possible.

The exits are not related in any way to the entries and in fact regardless of the entry location in most cases open trades will all use the same trailing SL.


E) MULTIPLE ENTRIES
Each EA entry is based on it's own logic and has no knowledge of prior entries the system may have made. Hence all trades are unique and distinct and based on it's merits alone.
Once the trades are placed the SL is simply trailed and the trade forgotten, the trade will always be exited via it's SL. The systems can make multiple trades at once and hence there will often be overlap between trades. It happens often that in the event of a trend the systems will enter multiple times throughout the trend and hence effectively pyramid earnings. It may also happen that trades in opposite direction are taken essentially hedging a part of the position.


Take Care

mim2005

Hi mim,


Entries solely on one TF. I reckon its lower TFs. Hm, when trend occurs or the price already started do rise or descent the system keeps on adding. You have three EAs. One is solely a trailing one? When contra signals occur you have a jumping trailing SL in place?


Im having difficultizz with trailing strategy with my system. No trailing works atm. Im thinking of a jumping SL as a trailing strategy but again it will keep me 2 months of forward testing to prove at least something. Hm, cant backtest my system that is why I have high cost of time and cost of RAM . Can you show some direction where you decided on a strategy.


Identify the range. Whats the system take on this. I see you have a range identification atm with hedged position on eu.


in your own time,
thanks
21

Hi Betterlife,

Thanks for your thoughts.

Yes I am a member of Asirikuy and I have a huge amount of respect for Daniel Fernandez and the work he has done with the Asirikuy community. I think Daniel is incredibly hard working and scientific in his approach and it amazes me how much effort he puts into sharing his work ... the amount of work he does makes me feel as though I am taking it easy by comparison. As your probably aware the quantity of information on his site is huge and I found it overwhelming. Whilst I didn't get into it into detail I did scan most of the topics I was interested in and in most cases I gained insight into things that I hadn't considered.

In short, I haven't used Mote Carlo simulations or any of the good tools Daniel has available, not because I don't think the're useful, but simply because I don't have enough time. I haven't visited the community for a few months so I'm not up to date on the latest. I did get the impression Daniel is more focused on higher time frames and don't think that the types of EAs I have developed (i.e. intra day edges) would appeal to him for this reason.

I personally believe that the returns of my EAs in live environments will be much, much lower than those in backtesting (let's hope the're not negative) - simply because as soon as you develop a profitable EA based on back testing almost by definition you are doing some curve fitting. How much is the performance going to degrade in future? I have no idea, although the only way to find out is to test it real time I guess. All I can say is, let's hope the EURUSD trends soon as the type of balanced range that's been occurring over the last few days is not good for these EAs.

Lots of pips and real money to you too

Take Care

mim2005

Hi 21vs7,

Tough situation trying to find a generic and effective trailing stop approach without backtesting.

If you've decided to use a trailing stop as an exit strategy the first thing you will need to consider is what type of system you are trying to build, if your looking to catch longer daily trends you will need to choose something that trails more loosely, and to capture short moves you will need to trail tightly.

The other two factor you will need to design are a) when to move the stop ? b) where to move the stop?

a) The best I can tell you from my experience is that the EAs I developed use an approach whereby the SL is usually only moved when the price goes substantially in their favour. So if price stays in the same spot more or less ... my EA SLs don't get moved.

b) Where to move your SL to ... well my EAs move them to a spot that on the longer term time frame they think if price moves back to then it has more or less invalidated the idea of a continued trend ... not sure if that makes sense but it's a common approach used to place any SL i.e. the point at which your original trend idea is invalidated.

I don't use %, or ATR, or Moving Averages etc. as there's no intelligence in them and I don't think your getting any additional edge with these.

If the above is too difficult then consider a higher timeframe pSAR ... it's where I started but there are definitely much more effective approaches as I touched on above.



In regards to your question regarding range detection and hedging. In most cases the EAs take a trade if they think price will move in a certain direction short term. In the last few days it just so happens that they figured to sell near the high side of the range and buy near the low side but they didn't necessarily "detect" the range to make this decision, if you follow my meaning. They EAs also do not use prior trade knowledge hence they EA by nature doesn't know that it's hedging, but by effectively combining the EAs the result is you get hedged situations and often the net position your actually holding is the sum of the directional convictions that the different EAs have at the time i.e. if an EA is thinking of trading 1 lot short and a bit later it or another EA decides it should trade 2 lots long you will end up with some hedging but a net long position. As you can see right now the EAs are net long.


Hope this helps a little.

Take care,

mim2005

Yes, tell me about it. I have 3 computers working all days. Im gonna add extra, but my lady will kill me

Plus, after some results one cant be sure that 2 weeks/months are representative. Of course they are not! So, its even harder. Had several phases of building this and atm relatively good.


Ok, im also looking at the same direction here.. I need trend and I need trend on at least 2 currencies..Im also having hedged positions, which EAs dont know about Cus of several EAs.



Ok, I understand few points you highlighted.


Im not really targeting those extra extra trends, Im just betting to cap some of that..as im trading the whole market.. And will easily reenter. Sometimes of course is on the tops, so need to work on that also. But still im gonna work to move TP up and away. But first thing first.


Concerning trailing strategy: When and where. Japs. I tryied tight trailing ones, I tried BE ones. This is the last time Im gonna consider trailing as an exit, otherwizze im thinking of moving to exit signals as a form of exit. Ill try to get agressive sell signals from the system and im gonna cap them there aroundish.


About ATR. I am thinking of ATR more as a volatility filter. I would like to filter few trades out, as Im trading heavily. pSAR im gonna look into but im gonna guess that eventually everything im gonna find is gonna be in the price alone. Talking about the index currency prices and that movement should give the clearer picture of moving the TP higher OR capping it at agressive exit signals.

Plus, im thinking to make TM trade manager who would control all this. lots of entries, lots of currencies so need an extra filter to play that. Also a lot of work in this direction.



Thanks for the info and insights,
Keep collecting.

Nice work compared to the rest of the "herd" guys...

I've been working on automatic system generation based on genetic programming for several large institutions for the past decades (being independent and successful using automatic generation approaches for the past 10-15 years). Just some warning:

1. Manual system generation approaches are old-school. You will be predated and front-run by other traders if you do this.
2. Using genetic programming you can mine systems with extremely good historical statistics and very high chances of success in live market conditions (took me a long time to figure out how). Highly profitable algo trading using data-mining without dredging is what you need to be in the game.
3. The higher the TF, the easier this is. You can make a killing trading the weekly.
4. Multi-instrument is key. If something works on 5 symbols it is incredibly more robust than a single one.
5. Synthetics are key.

From your work I would expect system failure within the next 1-3 years, with actual live trading statistics giving you a profit of about only 30% what you expect and drawdowns will be at least 2.5 times as they've been in your back-tests. Your approach is too weak to lead to significantly higher profitability but you're clearly headed down a better path than the rest. Best of luck guys.

Erik1, do you really don't know Daniel Fernandez? Sure?

You are pointing out, here and elsewhere, exactly the problems the community is getting crazy about recently.

As you are no person (as far as one can see) spitting out judgments without being backed by numbers and stats, how are the figures above calculated?

1-3 years (do you have a metric for OS failure prediction?)
profit about 30% of what backtested
dd 2.5 times than indicated?

Is it a result of a Monte Carlo?

High chances of succes in live market: please, again, is it a metric or an experience?

What do you mean for "highly profitable": can you give an AAR?

And then, gush! Synthetics...

It's hard to believe you're just guessing...

All ears, here.

Bye

BL

P.S.: sorry mim for hijacking this thread... you would concede these are quite interesting questions, after all.

Recently Hanover suggested the blog mate, but I don't know the fellow.


Yes, but not sharing trade secrets mate. I won't do your homework. This comes from my experience with exploitation of edges in backtesting data and then trading them live, trade frequency and the relationship they have with profitability under new trading conditions. When you develop apparently highly profitable, higher frequency edges in simulations you fall into a trap related to the pitfalls of consecutive losing trades. Think about it. This does not depend on how you develop your edge. I can tell what you might expect from a system under unknown conditions within certain confidence intervals depending on its characteristics. This has been fairly accurate for me.


Not precisely. Think more about bootstrapping, synthetic data using Frequency domain replacements, etc.


I'm not guessing mate, it's just my experience. Many people use synthetics.

Not interested in prolonging this discussion, was just dropping some small tips to help these fellows. Cheers.

Well, thanks a lot for the limited but significant contribution.

Just, if you want, consider that "providing help" might happen in some form that is a bit more constructive than saying "it will not work for these good reasons".

I credit you that, by following all your posts, one gets a good mine of hints about the direction of a productive research path.

Thank you for that.

Bye

BL