The time sequence of high and low in EOD data

Here is an interesting question. Without intraday minutes data, what is the best way to guess which of high and low happens first assuming only regular EOD data, that is, open,high,low,close,vol data is available? The EOD data does not have any information to tell which of high or low comes first.

From my experiments tested using thousands of EOD data points using multiple markets data, and validated with intraday data, the following algorithm shows around 86% accuracy.

d1=DIFF(DIFF(high, open), DIFF(open,low));
c1=LT(d1, 0.0);
d2=DIFF(DIFF(high, close), DIFF(close, low));
c2=GT(d2, 0.0);
hl=SEL(GE(ABS(d1), ABS(d2)), c1, c2);​

Here DIFF, LT, GT, and SEL are mathematical operators for subtraction, less-than, greater-than, and selection (condition, value1, value2).

Just wonder someone might have a better approach. Thanks

 

Maybe it's just me but I prefer 100% accuracy that intraday data gives. Any reason not to use it?

 

you can't - you have to make daily bars from minute bars.

 

Agreed. Me too think so. But when intraday data is not available or missing, at least there is some way to help. Also very interesting if EOD data were reported in a format of open,extreme1,extreme2,close,vol, here extreme1 or extreme2 is just one of high or low but tell which comes first. Easy to convert to anything from there to the standard format of any sort of a flavor.

 

From your equations, how do you determine which came first? I understand how the equations work but once you have "hl" what do you do?

 

This could be a good candidate for machine learning, i.e. pattern recognition. You have a large data set you can use for training and testing and you have a well defined output you're looking for (high/low first). As an added bonus you could add in other data points like volume, index sentiment, etc. to see if that improves results. Implementing this is not necessarily easy but it could be something worth looking into.

 

I am half-tempted to say that if it really does provide you with 86% accuracy, you can back out a trading strategy out of it

 

Not really it helps to predict, but it just rebuild some of the timing data from the lost EOD so that it helps in some back-test only which I am trying to figure out when no intraday data is available to check.

 

As I said it is an interesting problem to solve. I might not be any use to most of people, but I just wanted to use standard EOD data in a slightly improved way during some back-test using EOD data.

 

Totally agreed. I still think a simple easy and fast approach that could be more efficient for my purpose.