Page 1 of 5

Statistics, Statistics and more Statistics

Posted: Sun Dec 05, 2010 12:43 pm
by Relativity
Here are some stats I generated using MT4. Lets see what anyone can see the obvious. No interpretations please. Questions regarding the calculation are welcomed.

Image

Posted: Sun Dec 05, 2010 1:09 pm
by Naynay
Other than that scalping on the lower timeframes is actually more efficient than the higher ones ? ;)

Posted: Sun Dec 05, 2010 4:26 pm
by pika
Naynay wrote:Other than that scalping on the lower timeframes is actually more efficient than the higher ones ? ;)

Don't forget each trade has a transaction cost which is the spread. On a lower timeframe, the transaction cost is much higher than trading the hgher timeframe with respect to their average ranges, hence, trading in lower timeframe is actually less cost efficient. However, the lower cost efficiency in a lower timeframe is offset by higher trading opportunities from a volatility perspective, since there are more "convolutions" in the formation of the higher timeframe pattern. It is these "convolutions" that make the average range of a higher timeframe PER UNIT OF TIME lower than that of a lower timeframe, as shown in the results. The results is not a measure of trading efficiency in a lower timeframe from a cost perspective and the impact of spread is also not included.

Posted: Sun Dec 05, 2010 8:27 pm
by TheRumpledOne
How are you calculating pip/time efficiency?

Posted: Mon Dec 06, 2010 12:00 am
by Relativity
pika wrote:
Naynay wrote:Other than that scalping on the lower timeframes is actually more efficient than the higher ones ? ;)

Don't forget each trade has a transaction cost which is the spread. On a lower timeframe, the transaction cost is much higher than trading the hgher timeframe with respect to their average ranges, hence, trading in lower timeframe is actually less cost efficient. However, the lower cost efficiency in a lower timeframe is offset by higher trading opportunities from a volatility perspective, since there are more "convolutions" in the formation of the higher timeframe pattern. It is these "convolutions" that make the average range of a higher timeframe PER UNIT OF TIME lower than that of a lower timeframe, as shown in the results. The results is not a measure of trading efficiency in a lower timeframe from a cost perspective and the impact of spread is also not included.


Yes, because spread + commission can vary, I have to exclude these for now. The main purpose now is to spot price action in relation to time. I have found this study helpful in making me a lot less greedy when taking profits + leaving a lot of pips on the table.

Posted: Mon Dec 06, 2010 12:05 am
by Relativity
TheRumpledOne wrote:How are you calculating pip/time efficiency?


Avg Range / Timeframe in mins

e.g.
M60 or H1 Avg Range : 30.4 pips -> P/T e : 30.4/60 = 0.5066667 round up to 0.51

Posted: Mon Dec 06, 2010 12:15 am
by Relativity
Here are some more interesting results. Very very surprising! Now do you see it? Hint : think about why the new rat rules actually work.

Image

Image

Posted: Mon Dec 06, 2010 1:15 am
by pika
Relativity wrote:Here are some more interesting results. Very very surprising! Now do you see it? Hint : think about why the new rat rules actually work.

Image

Image

Relativity,
Your calculation method has one bias that gives the results as such, which is the sample size of the lower timeframes are much larger than that of the higher timeframes. e.g. Number of M1 bars over 3 months = 5x that of M5 bar sample = 15x that of M15 bar samples, and so on. Try standardizing the bar sample size for all the timeframes to the same number (e.g. 1000) and the results should be quite similar for all timeframes. Think about the "convolutions" (i.e. the bends and curves) in a lower timeframes that eventually forms the higher timeframes and you will better understand why the statistics turns out the way they are. I can only conclude that the higher transaction cost of trading is offset by the opportunity for higher trading frequency in a lower timeframe chart.

Please don't be mistaken that I am criticising your study. I am glad you are sharing your thoughts and I am just sharing mine since I have pondered over this area before some time ago when I wanted to understand why the ATR of different timeframes are not linearly related based on their time periods. If you may share your thoughts with more clarity, I would gladly learn from you if the evidence is sound.

Posted: Mon Dec 06, 2010 10:48 am
by Relativity
pika wrote:
Relativity wrote:Here are some more interesting results. Very very surprising! Now do you see it? Hint : think about why the new rat rules actually work.

Image

Image

Relativity,
Your calculation method has one bias that gives the results as such, which is the sample size of the lower timeframes are much larger than that of the higher timeframes. e.g. Number of M1 bars over 3 months = 5x that of M5 bar sample = 15x that of M15 bar samples, and so on. Try standardizing the bar sample size for all the timeframes to the same number (e.g. 1000) and the results should be quite similar for all timeframes. Think about the "convolutions" (i.e. the bends and curves) in a lower timeframes that eventually forms the higher timeframes and you will better understand why the statistics turns out the way they are. I can only conclude that the higher transaction cost of trading is offset by the opportunity for higher trading frequency in a lower timeframe chart.

Please don't be mistaken that I am criticising your study. I am glad you are sharing your thoughts and I am just sharing mine since I have pondered over this area before some time ago when I wanted to understand why the ATR of different timeframes are not linearly related based on their time periods. If you may share your thoughts with more clarity, I would gladly learn from you if the evidence is sound.


Thanks Pika, you input is especially welcomed. I like to improve my accuracy of this study, so critism is useful. Honestly, I am no math wizard so I currently do not know any better method of sampling such data, other than using basic averages.

I have considered your approach previously : to use a fixed number of bars as the base for the sample size. I don't think this sampling approach is sound. The results did not defer too greatly, other than higher TFs. It was initially the approach I used when I first did the MT4 coding. The following problems were found :

1- If the number of bars is set too high, MT4 may not have the data available, especially for higher TFs. e.g. 500 bars for monthly bars. Even if the data is available, it might not be relevant enough as market conditions are different.
2- If the number of bars is set too low to accomandate higher TFs, the overall results can suffer lack of accuracy.
3- The lower TF bars relationship to the higher TF bars is questioned. Say we use 100 as our base. We consider 100 M1 bars and 100 M5 bars. But one must also consider the other 'outsider' 400 M1 bars since they have some relationship to the 100 M5 bars we are sampling. If understanding the relationship is essential, then it is just not consistent to exclude those 'outsider' 400 M1 bars.

Hence the choice to pick a duration as a sample e.g. 1 to 3 months for relevancy, accuracy and consistency.

I did found 2 problems with my current approach.

1- My current broker might not be providing enough bars for sampling, especially the lower TFs. This will skew the results, so I wrote something up to check the number of bars from all TFs available for sampling. There must be sufficent bars available to sample. Then from here, work out a proper duration to sample. Current now it stands at 1 month, unless I can find some other way to draw more data.

2- MT4 takes M10080 as number of mins in a week. But this isnt the case as 60 mins x 24 hours x 5 trading days = 7200. To get around this problem, weekly bars can be excluded from the study from this point. So say we extract 1 month worth of data, which means :
23 D1/M1440 bars (since we have around 23 trading days in a week)
23*6 = 138 H4/M240 bars
23*24 = 552 H1/M60 bars
23*48 = 1104 M30 bars
23*96 = 2208 M15 bars
23*288 = 6624 M5 bars
23*1440 = 33120 M1 bars

Posted: Mon Dec 06, 2010 10:54 am
by Relativity
Anyway, here are the results with the modified mq4. Attached is the mq4. Please do check it to see what you can find.

Image

What I saw has not changed significantly :
- P/T e decreases when we go higher TFs
- P/T e increases when we go lower TFs
- The 'breakoff' point TF is between around M5/M15, sometimes M15/M30. The optimal pip/time efficiency.

So which means :
- The breakoff point TF is very useful to help finetune entry/exit plus to remind when one is being too greedy.
- To catch trends, study from the lower TFs to the breakoff point TF.
- To catch corrections/retracements, study from the higher TF to the breakoff point TF.

When we look at the old and new rat rules :
- Old Rat trades off the wick of D1 bar. There is so many pips available for the taking.
- New Rat works very well when one catches a good M15 bar and does not get any greedy. Don't try to push your luck further. Avg Body P/T e tells a lot about this truth. Take your profit and go. Even it is 5 pips (after minus spread).

I hope I concluded this correctly. Please sound me out if I am wrong anywhere!