Discussion of article "Exploring Seasonal Patterns of Financial Time Series with Boxplot" - page 27
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
I have bolted on my OLAP to the bar analysis via the adapter for MqlRates and some other updates. For EURUSD M15 for the period from 2010 to 2019, I decided to calculate the ProfitFactor aggregator by Close-Open bar ranges, broken down by hours and days of the week. Since this aggregator gives the ratio of positive amounts to negative amounts, its maximum (greater than 1) and minimum (less than 1) values can be interpreted as suitable for buying and selling respectively (for selling, from the PF less than 1 shown, the inverse of 1/PF should be taken to get the profitability of selling). Here is the log (I didn't make a graph):
Each line has PF, hour, and day of the week. Marked the most attractive options. You can see that it is recommended to sell at 23 and buy from 0 to 4 on almost all days.
T9
OK, accepted.
on the subject, well, as if everything is ready with the theory, then what to wait for? ;)
OK, roger that.
on the subject, well, if everything is ready with the theory, then what to wait for? ;)
It's quite a normal idea. You can draw and calculate) Draw - scatter diagram, calculate - correlation coefficient and its significance.
I have some econometric cretinism regarding price increments. I have drawn diagrams and hitmaps, made conclusions. But did not understand the subtleties of being, for example, have:
Correlation of hourly single increments - there are no regularities, which is normal.
with a lag of 10 there is, well, and the more the lag, the more correlation between individual clocks or groups.
But we realise that lag > 1 is a two-way stick - on the one hand it grows, and on the other hand, for example, it falls. That is, we need to look at the history and conduct a separate statistic, or somehow include it in the current one.
For example, if we have a positive increment with a lag of 10, we predict, with high probability, the same increment value at the next bar. Of course, this does not oblige the price to be at the same level. Do we look back 10 bars and subtract the -9th bar from the -10th bar, then add the difference to the last zero bar and get a correct prediction? How do I incorporate this into the stats\ analysis.... I'm floating. I.e. to calculate the real output in some statistical equivalent, probabilistic or something else. To be kind of scientific.)
Or is this a wrong way of thinking?
For example, it is easy to get such diagrams on the interval of 5 years with a lag of 25, for some hours.
H.I. I think I understand. It is necessary to restore the predicted series of increments, by hours (it will be averaged) and compare it with the initial one. The signals there will be by delta. We can use the same boxplots for clarity.
Theoretically, it's a kind of grail. Practically, let's see. The prediction error is obtained only because of the scattering of points on the scattering diagram, it is mystical.
with a lag of 10 there is, and the greater the lag, the greater the correlation between individual clocks or groups.
What is a lag in this case? The increments ten days ago of the 7-hour bar have a correlation with the increments of the 8-hour bar for the current day?
What is lag in this case? The increments ten days ago of the 7-hour bar have a relationship with the increments of the 8-hour bar for the current day?
Lag = lagged closing price, e.g. 0 bar divided by the 10th bar. 1-close[0]/close[10]
The increment of the current hour is highly correlated with the increment of the previous hour, it turns out. The greater the lag the greater the correlation, for certain hours 1-close[0]/close[10] correlates with 1-close[1]/close[11]
lag = lagged closing price, e.g. 0 bar divided by the 10th bar. 1-close[0]/close[10]
The increment of the current hour is highly correlated with the increment of the previous hour, it turns out. The greater the lag the greater the correlation, for certain hours 1-close[0]/close[10] correlates with 1-close[1]/close[11]
So high correlation purely mathematically will be observed on any series with this type of lag. Take SB for verification.
It is not good when an interval from one series overlaps with another. I.e. 0-10 overlaps with 1-11.
Ran it not by increments, but by signs from early 2014. Best result
There is no connection by signs on H1.
So high correlation purely mathematically will be observed on any series with this type of lag. Take SB to check.
If you remove the volatility clustering, the BP at closing prices is already SB, so I believe it.
but it's still interesting, and it turns out that not all hours have the same correlation increases/decreases as the lag changes. I guess we need to find the minimum robust lag.
If you remove the volatility clustering, BP at closing prices is already SB, so I believe it is
but it's still interesting, and it turns out that correlation doesn't grow and fall equally for all hours as the lag changes. I guess we need to find the minimum robust lag.
You can measure the correlation on a sliding window. And then look at its statistical characteristics. I think it will float a lot.