MA Cross Optimization EA (very cool) | Page 5 | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

New very cool CandleTime and session time Clock indicators 189 replies

Rsi cross indicator? (NOT cross 50, etc) 44 replies

very cool Trading Panel EA needs some adds 19 replies

Royal Cross/Fresh Cross 7 replies

World Market Hours - Very Cool 8 replies

Attachments: MA Cross Optimization EA (very cool)
Exit Attachments

MA Cross Optimization EA (very cool)

  • #81
  • Jul 20, 2008 4:49am Jul 20, 2008 4:49am
do you think a central banker gives a hootin-nanny about a turtle system? he's concerned with his country's economy, not some small fry speculator.... forex is not a real market.

Quoting Rabid
There are a number of systems that have become so well traded that it makes more sense to fade against them than to trade with them.
all this "random entry" talk has me a bit worried about your philosophy. direction is key.... not managing an exit. we could go round and round debating this, and i really don't want to go there. good luck w/ it, rabid!

As for this:
Quoting rabid
If you have a system that works well in the past there is no guarantee that it'll continue to work in the future.
nothing i wrote disputes this. note my words were "gamble" and "speculate."
  • #82
  • Jul 20, 2008 5:22am Jul 20, 2008 5:22am
Quote
do you think a central banker gives a hootin-nanny about a turtle system? he's concerned with his country's economy, not some small fry speculator.... forex is not a real market.
And do you think I give a hootin-nanny about what a "central banker" does? Regardless of whatever they do, good speculators will trade it accordingly. I don't care to predict the markets, there are 2 kinds of people in this business it seems... analysts and traders. You can't be both. I trade other markets besides forex, have traded equities for going on... well, dang, nearly 7 years now in one form or another. Am also getting interested in commodities a bit, but regardless of "the market" or "central bankers" or some other mystical outside force there's still money to be made somewhere. Analysts write great entertaining works of fiction, often with believable protagonists and a deep sense of drama... but in no way would I trust them to trade for me.

Quote
all this "random entry" talk has me a bit worried about your philosophy. direction is key.... not managing an exit. we could go round and round debating this, and i really don't want to go there. good luck w/ it, rabid!
ROFL. Yes but the market tells you direction without any need to guess. Direction is primitive, you could be 100% right on direction and still lose all of your money if you don't get it at the right time and with the right momentum. But there is no holy grail of entries, there are no systems with 0% risk. Even if you got the direction correct 95% of the time, if the other 5% had 100x the loss... you're done for. So yes quite clearly you need to be worried about your exits too, anything counter is sheer foolishness.

Anyway, I'm just trying to help you out here. When I stopped caring about direction and "being right" and just started trading price and became, well, obsessed with controlling risk, I started being profitable. That's all. I don't think direction means a hill of beans, up or down... it's all the same to me. I take a trade when the price tells me to, and I get out when price tells me to. Forex moves a lot more than any other market I've traded... so it seems you have to be a lot more sensitive to these things here than elsewhere, but the basic concepts are the same.

Quote
nothing i wrote disputes this. note my words were "gamble" and "speculate."
Well, I'm not going to debate semantics or what you intended, I can only point to what you've said. It's seems the intent of this thread is to demonstrate how moving average cross-overs with the right settings can be profitable. I don't debate that at all, I just caution about over-optimizing. You seem to recognize that when you mentioned that all of this could just be coincidence... to which I could only respond w/ a big fat "well yea, duh, that's the rule of back testing."
  • #83
  • Jul 20, 2008 6:59am Jul 20, 2008 6:59am
Tdion, I think it's time you read a book on system development =D.

Rabid makes some valid points--not that you haven't--and a little something I noticed about my fantasy EA:

I watched the entire 6 months play back at me, and I noticed that EURUSD has this tendency to move in short bursts for a period after an equally long period ranging. Ok, that's a unique characteristic about this pair for this period, how then do I take advantage of it?

Develop a system to consistently capture those breakouts, and collect the pips. That's exactly what the EA does, and it worked beautifully--then. Now, it captures about 70% of pips compared with the backtest.

I guess I'm telling you to just continue to run these optimizations. I suspect that there will be this 'eureka' moment where you discover a robust combination of settings and indicators which can properly time the market. I've had the pleasure of temporarily going through that moment--until it hit a division by 0 error.
  • #84
  • Jul 20, 2008 7:05am Jul 20, 2008 7:05am
OK, guys, let me throw in a few more logs into the fire, before trying to put things back on track

Past performance, back-testing and curve-fitting: we all seem to agree that surely we cannot predict the future, nor can we rely 100% on what we discovered in the past (in terms of "predictable conditions") for our future trades. Testing a system/concept with a big historic data could, in fact, serve little purpose - most of us trade with shorter timeframes in mind, in terms of profit, than a decade. Who would be willing to sit on a negative position for 7-10 years even if it will eventually give profit?

So, we expect and need the results today/this week/this month, not in 10, not even in 2 years' time. Hence, if system performed well over a given 10 year period, it could totally fail during the next few weeks/months, wiping out our balance. Quick example: MA Optimizer gave good results for UJ on H1 with 60/86 MAs when tested on 2007 data. The same set-up led to a loss during the first 6 months of 2008. To make profit, one had to change (in 2008) MAs to 55/85 - "small" difference which makes ALL the difference in terms of profit, with no way of knowing that on 1 January 2008 one had to change one MA from 60 to 55.

Fibos: in my opinion, they work at least for one simple reason (the same goes for any S/R level) - "the madness of crowds". Majority of personal and institutional traders have the same S/R and Fibo levels on the screen and they EXPECT smth to happen around those "magic" levels. So, when the price does reach e.g. certain Fibo level there could be sufficient number of traders puttings things "on hold", and hence the price sticks to that level too. So, we should keep those indicators in mind purely because dozens/hundreds of large institutional traders keep an eye on them too ("herd mentality", unfortunately - we are too small to ignore what big boys are possibly thinking).

Exit/entry: determining a trend (in principal) is a piece of cake (well, more or less ). Getting the right entry point is more difficult, but there are proven system to facilitate that. Exiting at the right time IS the key to success/profit. How many times did we turn a winning trade (even +10-15 pips) into a 20-30 pips looser? How many time did we jumped to lock in a 30 pips profit to see the price moving another 100 pips? Search for DIBS thread, they have good ideas on Risk/Reward and Money Management. The key concept is a "free trade" - you exit part of your position to lock in some profit, you let the remaining part "run", if it's SLed - you at least break even; if it makes the home run, you have a BIG winner. Other people trail market too, moving SL to lock in more and more profit as it becomes available. My personal opinion is: 10 pips of sure profit is MUCH better than loosing 10-20 pips chasing a 100-pip gain.

Central banks: whatever they do, the FX market does not react with an instant change of trend. True, price does jump a lot during major news (e.g. interest rate decisions), but (a) we know well in advance when that's gonna happen, so we can close our positions and sit on the fence for a while; and (b) with a bit of fundamental knowledge it's not hard to see what decision CBs are likely to make. Then, of course, we don't have to follow the initial move right away - we can see "what happens" first.

Mechanical trading vs discretionary: both work for different people, depends on one's circumstances, pocket, expectations and timeframe (how soon one needs real profit which can be spent on food/car/plane/school bills whatever).

OK, enough side-tracking. The subject of the thread is MA crosses (with additional indicators now thrown in). I see Tdion's efforts as a good base for analysing various parameters using user-selected TFs and pairs (and back-testing period). It's not a holy grail, it's a not an algo trading platform. By now, we seem to all agree that no matter how attractive the results of any backtesting are, it could be dangerous to jump in blindly on every signal created by such a system, or to fully entrust all the money to an algo. Surely, human factor brings in the issues of emotions and discipline, but opening a position with a perfect signal 10 minutes before CB's rate decision would be suicidal. Algo could miss that (unless it's very complex, but then we are talking about AI ) - good trader will spot it.

Sorry for such a long post. It's slightly off the topic (thread's theme), but I felt the above points do show that the same thing can be viewed differently from various angles (and there is no "ultimate" truth - we are all different, with different needs, believes, and expectations).

So, let's see what each one of us can apply newly discovered knowledge to what we know already to increase our chances/profit.

One Chart is Worth Ten Thousand Words.
  • #85
  • Edited 7:33am Jul 20, 2008 7:22am | Edited 7:33am
Quoting Rabid
........
That's the point of a Monte Carlo simulation, you run a bunch of random data inputs and see how a system performs given a purely random sample over large periods of time. Some systems are profitable even given purely random data.
But how? If the data is a random walk, then every system is reduced to a net expectancy of exactly zero. Trying to devise a system to predict a coin toss, based on past "trends" of heads and tails, is a complete waste of time.
.........
I've been saying it for a long time now... a good exit and good risk management can turn a losing system into a winning system.
But if all that's required to be long-term profitable is a good exit and MM, then writing a profitable EA is child's play. In fact, just about every EA would be profitable.
........
I'm not trying to attack you, Rabid, I enjoy reading your posts. But I'll eat my gardening shoes if anybody can show me a system that's capable of attaining positive expectancy, indefinitely, when confronted by a randomly generated data set.

In terms of altering expectancy, MM is impotent. To anybody who's not convinced, try the XLSs attached to posts #31, #32, #43 in this thread:
http://www.forexfactory.com/showthread.php?p=2093339
Again, I'll consume my gumboots if anybody can come up with a valid mathematical proof that demonstrates otherwise.

On second thoughts, forget the gumboots, I'll just go out and bankrupt every casino in Vegas by playing only Roulette.

David

[EDIT] Footnote: my thoughts on randomness here http://www.forexfactory.com/showthread.php?p=2080271, for whatever they're worth.
  • #86
  • Jul 20, 2008 8:12am Jul 20, 2008 8:12am
Quote
But how? If the data is a random walk, then every system is reduced to a net expectancy of exactly zero. Trying to devise a system to predict a coin toss, based on past "trends" of heads and tails, is a complete waste of time.
You don't base it on past trends. Past data is useless in a random walk. Remember random data follows a normal distribution. That alone is sufficient to make money on purely random data.

Quote
But if all that's required to be long-term profitable is a good exit and MM, then writing a profitable EA is child's play. In fact, just about every EA would be profitable.
Unless of course they don't manage their risk or don't react quickly enough to a loss.

Quote
I'm not trying to attack you, Rabid, I enjoy reading your posts. But I'll eat my gardening shoes if anybody can show me a system that's capable of attaining positive expectancy, indefinitely, when confronted by a randomly generated data set.
If all you're concerned about is a purely random data set, that's easy. Think regression to the mean. A purely random data set will always regress back to a fixed mean. The markets are not so simple of course, so that won't help you trade in real life. In the fantasy world of random data you can simply wait for price to exceed 2 standard deviations and fade it with proper odds-based MM.

Quote
In terms of altering expectancy, MM is impotent. To anybody who's not convinced, try the XLSs attached to posts #31, #32, #43 in this thread:
There is a very big flaw in your analysis. Can you see it? In a casino environment the length of bets are fixed. You put your money in, the dice is rolled, the result is output and you win or lose. Trading is a very different environment. Trading is an exercise in conditional probability.

Consider the following game:

You have 2 dice rolls. The first roll, then the 2nd. You can bet before the first roll and after the first roll, you can cannot remove money placed on the first bet but you can add to it if you'd like. In order to be considered for a pay out you must "buy" a slot by putting at least some money before the first roll, but there is no substantial minimum requirement. You can put 1 dollar or 100 at the first roll, you can put 1 dollar or a 1000 after the 1st, doesn't matter. Ok?

Now the goal of the game is simple. Get a 9 or higher. If a 9 or higher is rolled between 2 dice, you win a 1:1 payout. If not, you lose any money you've put on the table.

Knowing this you can create an optimal strategy for playing the game, right? Put some small amount to buy into the game. If a 6 rolls on the first dice, increase your stake. Why? Because the odds of getting a 9 or higher with a 6 already on the table... a 3, 4, 5 or a 6, that's 4/6 or 66.6%. In that case, increase your stake to, say, $10, and capitalize on the increased odds.

This is the difference between a flat statistical outcome and a conditional probability outcome. The market gives you an opportunity for conditional probability, where casinos try very hard not to. The most obvious exception being blackjack, which has some notorious weaknesses for skilled players.

In short, if you fail to take advantage of the conditional opportunities given then your odds of a 9...

1st / 2nd
3 / 6
4 / 5 6
5 / 4 5 6
6 / 3 4 5 6

1+2+3+4 = 10. 10/36 or 28% versus 66.6%. Scaling in at the right time makes all the difference in the world, and scaling in is nothing but money management.
  • #87
  • Jul 20, 2008 8:12am Jul 20, 2008 8:12am
Quoting Forex007
...........
Past performance, back-testing and curve-fitting: we all seem to agree that surely we cannot predict the future, nor can we rely 100% on what we discovered in the past (in terms of "predictable conditions") for our future trades. Testing a system/concept with a big historic data could, in fact, serve little purpose..........
Surely one either has a provable "edge" - by somehow exploiting whatever aspect(s) of non-randomness might exit in price movement - or one doesn't. If the latter, trading is nothing more than a gamble. Likewise, if past performance really offers no guarantees, then back-testing is a complete waste of time.

TheRealThing explains the concept of an edge very eloquently here: http://www.forexfactory.com/showthread.php?p=205799

Whether a system is mechanical or discretionary, the assumption is that past behavior will, on balance, continue to repeat itself indefinitely, i.e. the aspect of non-randomness that's being harnessed (whether it's an MA crossover, a pin-bar entry, or a trailing-stop exit) is providing some sort of 'perpertual' edge.
.................
Exit/entry: determining a trend (in principal) is a piece of cake (well, more or less ). Getting the right entry point is more difficult, but there are proven system to facilitate that. Exiting at the right time IS the key to success/profit. How many times did we turn a winning trade (even +10-15 pips) into a 20-30 pips looser? How many time did we jumped to lock in a 30 pips profit to see the price moving another 100 pips? Search for DIBS thread, they have good ideas on Risk/Reward and Money Management. The key concept is a "free trade" - you exit part of your position to lock in some profit, you let the remaining part "run", if it's SLed - you at least break even; if it makes the home run, you have a BIG winner. Other people trail market too, moving SL to lock in more and more profit as it becomes available. My personal opinion is: 10 pips of sure profit is MUCH better than loosing 10-20 pips chasing a 100-pip gain.
DIBS has apparently been traded by Peter Crowns and TheRealThing for more than 15 years, and (assuming they're telling the truth) they have proven it to be profitable. It involves trading a low risk (pip-wise) breakout entry from an IB, and (like you say) taking half the position off at 1:1 R, and allowing the rest to ride. PC and TRT prefer to trade "perverse" (Peter's own word) systems that have prolonged losing sequences, but count on the fact that when the market trends exceptionally, they have positioned themselves to catch the trend with a "long tail" home run trade. (See: http://www.forexfactory.com/showthread.php?p=2108541).

Every system ultimately comes down to expectancy, i.e. the product of win rate and R-value. Creating a "free trade" is fine and dandy, but ultimately it comes down to how frequently one is able to attain this, relative to the number of times one gets stopped out for a loss. Every system is ultimately a trade off between win rate and R-value. These two factors are inversely proportional to each other, and an "edge" comes only by having an entry and/or exit methodology that somehow overcomes this inherent equilibrium.

If the market trends often enough to allow the home run trades to compensate for the frequent losses, then a "cut losses and let profits run" approach will ultimately deliver an edge. In this case, the occasional "extraordinary" trend that facilitates the home run is what offers an opportunity that is qualitatively different from anything being offered by a randomly generated data set.
Nice post, 007, I couldn't resist replying to a couple of your points.

David
  • #88
  • Jul 20, 2008 8:48am Jul 20, 2008 8:48am
Rabid, in response to post #86:

Ok, I (hopefully!) understand where you're coming from, but apparently we see this differently.

You say that random data regreeses to a mean. Coin-toss analogy: let a head represent an uptick (+1), a tail a downtick (–1). We toss repeatedly, and by keeping a running total we have a randomly generated "price" chart. Now supposing I start by tossing 20 heads in a row, so the total is +20. Is the probability that "price" will regress back to the mean (zero point) now greater than it trending away? I would say that it's 50/50; the past is irrelevant.

Conditional probability. In your dice game, it's effectively a case of increasing bet size when you can be sure that the probabilities are in your favor. (Like a Blackjack pro increasing his bet when there are more high cards remaining in the shoe, because the odds favor the player at that point). To have that kind of advantage in trading, we need to be able to know that a certain characteristic within the data is non-random, i.e. (for your analogy to be valid) something akin to the fact that we have "already rolled a 6 and we need to total >=9".

In other words, for any exit strategy to yield an expectancy other than zero, there has to be something about the data that's qualitatively different from a data sequence generated by a random process (like I tried to explain in my last couple of paragraphs in my reply to Forex007 above).

If the data sequence was random, i.e. any "trend" would be equally as likely to end at any point as any other, then a frequency distribution (over a very large sample size) would show that the longer the trend, the lower its frequency, in a smooth curve, in such a way that a "cut losses and let profits run" strategy would (like any other exit strategy) deliver a net expectancy of zero. To whatever extent traders are able to benefit by cutting losses, and letting profits run, the curve is skewed, a symptom that the market does "trend" occasionally, i.e. the data set that is price movement is not entirely random.

David
  • #89
  • Jul 20, 2008 9:11am Jul 20, 2008 9:11am
Another angle (hopefully; it's 2am here in New Zealand and I'm feeling a little sleepy).

Rabid, to apply the dice game analogy to trading, price is at 100, I buy 1 lot, and it moves to 106 (corresponds to a dice roll of 6). Now, everything else being equal, the probability of its reaching 109 or better is greater than it was when it was 100. But if I buy another X lots now, the probability of a rise or a fall is 50/50, so the move to 106 doesn't provide me with the same winning edge that's created artificially by the rules of the dice game. In that sense, I think the analogy is flawed.

In the dice game, the rules link the two rolls, so that the second roll (although being random) generates a total that is dependent on the result a past event (the first roll). For this to be analogous to trading, we would have to say that one price movement is in some way affected by a prior price movement. Therefore the outcome of each tick is not completely independent. But I would argue that it is.

What am I missing?

David
  • #90
  • Edited 9:34am Jul 20, 2008 9:17am | Edited 9:34am
Quote
You say that random data regreeses to a mean. Coin-toss analogy: let a head represent an uptick (+1), a tail a downtick (1). We toss repeatedly, and by keeping a running total we have a randomly generated "price" chart. Now supposing I start by tossing 20 heads in a row, so the total is +20. Is the probability that "price" will regress back to the mean (zero point) now greater than it trending away? I would say that it's 50/50; the past is irrelevant.
For any given coin flip the odds will always be 50/50, absolutely. Law of independent trials, gambler's fallacy and all that. But we're not talking about a single coin flip, or even 20, we're talking about the sum of hundreds of flips as "ticks." As such the length of a trend will follow standard distribution variance.

At 2 standard deviations w/ a normal distribution we will account for 95.45% of the sample. At 3, that's 99.73. That means if we trade the fade at the 2nd and put a stop loss at the 3rd, we will account for 4.28% of all trades, but a whopping 94% of all trades that reach the 2nd stdev. Since the risk would be 1 stdev while the return would be 2, we'd have a nice 2:1 RR and an amazing win/loss ratio.

This is all bound by something called the "central limit theorem" which states that the sum of a large number of independent and identically-distributed random variables will be approximately normally distributed following a Gaussian distribution (bell curve).

More info can be found here:
http://en.wikipedia.org/wiki/Central_limit_theorem

The problem here comes in applying this theorem to the markets. The markets do not follow a normal distribution, which is proof enough that they are not random. It gets a lot tricker from there, but models have been developed to help compensate for this. This problem is very well understood and studied in the options markets as it directly effects volatility pricing models.

Quote
Conditional probability. In your dice game, it's effectively a case of increasing bet size when you can be sure that the probabilities are in your favor. (Like a Blackjack pro increasing his bet when there are more high cards remaining in the shoe, because the odds favor the player at that point). To have that kind of advantage in trading, we need to be able to know that a certain characteristic within the data is non-random, i.e. (for your analogy to be valid) something akin to the fact that we have "already rolled a 6 and we need to total 9". In that case I would argue that, if it's possible to make such a forecast, then the data can not be a completely random walk.
Well here we're getting into the differences between trading and random walks. But more to the point, you can certainly know the odds in a random game. I can tell you the odds of any fair dice roll is 1/6, that's simple enough. You do not need a non-random set in order to predict odds, infact the whole purpose of odds is to give you an estimate in exactly that case.

Granted, in a coin toss game the situation is different. Because each case is 50/50, the best case conditional probability for say, 2 heads where the first flip is heads is 50%. That doesn't give you an advantage. But if you expand from a coin flip to a dice game, each roll is completely random yet the odds are easily predicted. What you're saying is that the dice game can not yield such odds because it cannot be completely random when in fact we know a fair dice game would be completely random, you cannot predict in advance what an individual dice roll would be, and yet we can clearly distinguish such odds.

It is because our sample is assumed to be completely random that we can in fact assume such odds with clarity. That's why I said trading a random data set is much easier than trading the markets, in the markets things don't follow a normal distribution. "Extreme" levels of volatility or trend persistency occur with much greater frequency than a normal distribution would allow.

Quote
In other words, for any exit strategy to yield an expectancy other than zero, there has to be something about the data that's qualitatively different from a data sequence generated by a random process (like I tried to explain in my last couple of paragraphs in my reply to Forex007 above).
But that's not true, mathematically speaking. While it's true that any given coin flip or "tick" has an equal chance of being heads or tails... and if we were betting only on a single tick it would not be possible to make money, the sum of a large number of flips follows the central limit theorem. This is a relatively proven fact. Why do you think I requested a "price density index" indicator on my MT4 thread?

Quote
If the data sequence was random, i.e. any "trend" would be equally as likely to end at any point as any other, then a frequency distribution (over a very large sample size) would show that the longer the trend, the lower its frequency, in a smooth curve, in such a way that a "cut losses and let profits run" strategy would (like any other exit strategy) deliver a net expectancy of zero. To whatever extent traders are able to benefit by cutting losses, and letting profits run, the curve is skewed, a symptom that the market does "trend" occasionally, i.e. the data set generated by the orders that are placed is not entirely random.
You're still thinking in terms of fixed time outcomes. A random data sequence is not equally likely to end at any point as any other, that's the point of normal distribution. Granted you cannot predict exactly when it will end, but you can give odds that it will end within a range and over millions of trials it certainly will bare out over time.

Quote
Rabid, to apply the dice game analogy to trading, price is at 100, I buy 1 lot, and it moves to 106 (corresponds to a dice roll of 6). Now, everything else being equal, the probability of its reaching 109 or better is greater than it was when it was 100. But if I buy another X lots now, the probability of a rise or a fall is 50/50, so the move to 106 doesn't provide me with the same winning edge that's created artificially by the rules of the dice game. In that sense, I think the analogy is flawed.
Actually the fact that it has moved to 106 reduces your odds of a further increase (unless it goes to trend persistency). Keep thinking about it tho, I think you're getting closer to figuring it out... The rules of the real market are certainly far more complex than any brain sim we've come up with so far.

The markets certainly follow a varied approach in real life. They aren't a coin flip, they aren't a dice game, they're anything and everything they chose to be. Which is partially the problem, since you cannot know the number of possible outcomes in advance there is no fixed average.
  • #91
  • Jul 20, 2008 2:54pm Jul 20, 2008 2:54pm
i think everyone here is giving intelligent input, and i am not going to challenge anyone's opinion in the previous posts.

i feel it necessary to repeat this sentiment again:

if you truly have a winning strategy, it will have performed well in the past, and will in the future. therefore, it is a blanket statement to say that curvefit systems are garbage. within the entire set of all curve-fit systems are some legitimate profitable strategies

at some point you must decide your speculation strategy, and i think one w/ optimized parameters with a large trade population size is the way i will go. that's just me.

ultimately, i can see myself combining 5 or 6 strategies running on different pairs to create a single system. i can also see myself trading several of these hybrid systems (different strategies in each one) across multiple accounts.
  • #92
  • Jul 20, 2008 3:19pm Jul 20, 2008 3:19pm
Interesting thought. A curve fitted strategy is usually fitted to a particular market, which is why it blows up if applied to another market. If you have a curve fit for each market and trade that market only w/ that fit - one of them is bound to work? If you can keep your losses small and let that occasional winner run - your system just might work, but you will need to trade a wide variety of markets - so that you do not miss the one that is adhering to the curve fit.
Basically your gonna be a trend follower - welcome to the club

Quoting tdion
i
ultimately, i can see myself combining 5 or 6 strategies running on different pairs to create a single system. i can also see myself trading several of these hybrid systems (different strategies in each one) across multiple accounts.
In trading, there is no bullshit. You either make money or you don't.
  • #93
  • Jul 20, 2008 3:45pm Jul 20, 2008 3:45pm
mr. green,

i think it's interesting you call it trend following, because "trending" to most traders means the market moves up or down a signficant amount. but in this case, the only trend we are concerned with is whether the strategy continues to work, which would be trend in a different context.

we can see the same strategy applied to multiple pairs in Pipboxer, where aparsai is trading each pair optimized to its own setting. i will do the same, but put different strategies into each system instead of the identical one.

when you don't have a clearly defined edge, this optimization technique is the best i can come up with. besides, the strategies i am optimizing make sense (IE the RSI buys overbought and sells oversold, looking for breakouts...) so it is inevitable that the strategy will experience winning periods.

-td
  • #94
  • Jul 20, 2008 4:01pm Jul 20, 2008 4:01pm
With out a trend an MA cross will whip you to death. How do u define the strategy "working"?

Quoting tdion
mr. green,

i think it's interesting you call it trend following, because "trending" to most traders means the market moves up or down a signficant amount. but in this case, the only trend we are concerned with is whether the strategy continues to work, which would be trend in a different context.

we can see the same strategy applied to multiple pairs in Pipboxer, where aparsai is trading each pair optimized to its own setting. i will do the same, but put different strategies into each system instead of the identical one.

when you don't have a clearly defined edge, this optimization technique is the best i can come up with. besides, the strategies i am optimizing make sense (IE the RSI buys overbought and sells oversold, looking for breakouts...) so it is inevitable that the strategy will experience winning periods.

-td
In trading, there is no bullshit. You either make money or you don't.
  • #95
  • Jul 20, 2008 4:10pm Jul 20, 2008 4:10pm
green,

the MA cross was the first optimization EA i wrote..... i have moved onto RSI optimization strategies, whose performance graphs can be found on previous pages.

some strategies can be countertrending (IE sell when overbought, and buy when oversold)

i thought you meant that you hope for any of the strategies to continue working, which is a trend unto itself.
  • #96
  • Jul 20, 2008 4:27pm Jul 20, 2008 4:27pm
My Apologies Tdion - I am acting like a noob and not reading the whole thread. I just made the assumption you were still on subject to the thread title. Should have known, like most threads, that by page 7 you cant recognize what the thread was originally supposed to be about

Quoting tdion
green,

the MA cross was the first optimization EA i wrote..... i have moved onto RSI optimization strategies, whose performance graphs can be found on previous pages.

some strategies can be countertrending (IE sell when overbought, and buy when oversold)

i thought you meant that you hope for any of the strategies to continue working, which is a trend unto itself.
In trading, there is no bullshit. You either make money or you don't.
  • #97
  • Jul 20, 2008 4:37pm Jul 20, 2008 4:37pm
Rabid, I’ve hijacked Tdion’s thread too far already (sorry, Tom!), so I’ll make this my absolute final post on this topic. Just want to make sure that I understand CLT.

The coin flip is a binomial distribution, and the probabilities are shown in each row in Pascal’s Triangle. If we extend the number of rows indefinitely, it approximates a normal distribution.

The probability of the next flip is 50/50, making forecasting impossible. But if we consider the collective outcome of the next 16 flips, PT shows us that the probability of flipping between 8 and 10 heads is (8008+11440+12870+11440+8008)/2^16 = 51776/65536 = ~ 0.79. So if a bookie was dumb enough to offer me odds of 1:1 against this happening, I should gleefully accept the bet.

Now if we extrapolate to a hundred, a thousand, a million, etc flips, we approach a normal distribution, with the appropriate confidence levels, i.e. 68.2% for ±1 SD, 95.4% for ±2 SD, etc.

So if we were to apply this to the market (analogy: head = uptick, or +1; tail = downtick or –1), we could write some kind of options bet on the assumption of mean reversion, structuring it with the expectancy in our favor, and hoping that there was somebody dumb enough to buy the option. But, like you say, greed/fear psychology introduces the possibility of trends, so any assumption of "pure" randomness is flawed.

So the conclusion is that a "random" data set offers predictability by way of mean reversion, but an "unpredictable" data set does not. Terminology!

Thanks, Rabid, for your patience, and your clear explanation. Now I have to find those gardening shoes……

David
___________________________________________________

Quoting tdion
......
if you truly have a winning strategy, it will have performed well in the past, and will in the future. therefore, it is a blanket statement to say that curvefit systems are garbage. within the entire set of all curve-fit systems are some legitimate profitable strategies
.......
I think that's a reasonable assumption. If you'd started trading your USDJPY crossover system last year (assuming a similar optimization [16/18] was available then), the equity curve shows that you'd now be in profit. Same if you started 2 years ago. Same 3 years ago. And so on. So why not the upcoming year also? At each point on your equity curve, the future was every bit as unknown as is the case currently.

If long-time successful traders like PeterCrowns and TheRealThing are simply applying the same tried-and-trusted formula (even if parts are discretionary, there's still a consistency of rationale) over many years, and their justification for continuing to do so is nothing more than their past results, then what is different here? Does an indicator-based system lack the necessary "intelligence"? What if additional rules (whether indicators, line studies, multiple time frame analysis, chart/candle patterns, etc) were progressively added, by analyzing a large number of setups taken by a "master" (technical) trader, and gradually honing, via an inclusion/exclusion process, the specific patterns that determine his entries and exits? What makes one setup valid, and another not? What pattern or rule determines a certain type of exit? Sure, there may be a myriad of factors, but the number is far from infinite. And there are profitable traders who use relatively simple technical systems.

There's a mathematical approximation to any rationale; and, while a trader's learning process may be endless, the knowledge needed to reach a satisfactory level of profitability is (apparently) attainable within a reasonable time.

David
  • #98
  • Jul 20, 2008 4:39pm Jul 20, 2008 4:39pm
david,

you could post that the desert is wet and the ocean is dry, and i wouldn't have a problem with it. you are always welcome on any of my threads.

cheers....
  • #99
  • Jul 20, 2008 4:53pm Jul 20, 2008 4:53pm
Tom, what happens if you run your EA over the full data window, and then re-optimize the parameters (MA1, MA2, TP, SL) while reducing the window progressively by one year each time, i.e. removing the most recent year on the first run, the most recent two years on the second run, and so on?

How consistent do the parameters remain?

If, conversely, you retain your current parameters (16, 18, etc), but run the test, starting with the first (say) 10 years only, then the first 11 years, then the first 12…..

(Or use 2-year, or 3-year, increments, if you wish, to shorten the process).

This should at least give you as starting point for gauging to the extent and nature of any future re-optimization that might be required.

David
  • #100
  • Jul 20, 2008 5:21pm Jul 20, 2008 5:21pm
the optimal parameters change as you reduce the size of the data set. the fact that the 18/16 for USDJPY did well over all years was remarkable. in fact, it is the only one size fits all cross i've found (although i haven't spent much time exploring other pairs.)

i once read when you curve fit a system, do a mitten fit and not a glove. don't go for the highest profit parameters, but something in between the optimal parameters and breaking even. it makes some sense, but no matter how you approach it, future chaos can sink your ship.
0 traders viewing now