At what point in the day can you actually trade using this? Obviously there is a minimum amount of data required to calculate a variance ratio which can be used for a decision matrix. If you are using 5 minute bars how many hours do you need to have enough data to trade from? Or do you do 1 minute bars?
You are using an out of date browser. It may not display this or other websites correctly.
You should upgrade or use an alternative browser.
You should upgrade or use an alternative browser.
Quantifying randomness: variance ratio
- Thread starter stephencrowley
- Start date
I don't use this for forecasting.. I use it to quantify many days in the past to see if the amount of randomness is stable and then use it to decide what types of rules to use in the near future.
Also, I don't use bars.. tick/quote data
Also, I don't use bars.. tick/quote data
Quote from mysticman:
At what point in the day can you actually trade using this? Obviously there is a minimum amount of data required to calculate a variance ratio which can be used for a decision matrix. If you are using 5 minute bars how many hours do you need to have enough data to trade from? Or do you do 1 minute bars?
Quote from stephencrowley:
I don't use this for forecasting.. I use it to quantify many days in the past to see if the amount of randomness is stable and then use it to decide what types of rules to use in the near future.
Also, I don't use bars.. tick/quote data
Why look for stable randomness?
I would have thought looking for stable non-randomness would be more useful.
Using tick quotes makes sense as it eliminates the lag and bid-ask bounce that is intrinsic to last prices.
If some 'thing', whether it be a stock, etf, spread, etc is consistently trending, or consistently mean reverting every single day then you can probably assume it will continue trending or reverting..
But if something trends half the time and mean reverts half the time, you cant look into the future and say 'today the market will mean revert so i'll use X strategy.'
If something is a complete random walk every day then you can probably do no better than buy and hold.. I mean, you can probably "beat" the market a few days, but it'll be due only to chance.
But if something trends half the time and mean reverts half the time, you cant look into the future and say 'today the market will mean revert so i'll use X strategy.'
If something is a complete random walk every day then you can probably do no better than buy and hold.. I mean, you can probably "beat" the market a few days, but it'll be due only to chance.
Quote from squeeze:
Why look for stable randomness?
I would have thought looking for stable non-randomness would be more useful.
Using tick quotes makes sense as it eliminates the lag and bid-ask bounce that is intrinsic to last prices.
No.. you cant calculate VR with 10 ticks.. you need shitloads of data. Like I said, it is purely an after-the-fact analysis.. if your sample is large enough you might be able to split it in half or something.. but as the number of points decreases the VR accuracy decreases as well.
Quote from GTG:
By this, do you mean you calculate the variance ratio in "trading time" instead of real time sort of like this:
vr = last 10 tick variance / last 20 tick variance
?
Depends on what im doing. When analyzing variance ratios you just need a lot of evenly spaced single-variable data, that can be anything. When trading, I use sub-second quote data which feeds into the algorithms, indicators, whatever and that in turn figures out position sizing, etc.
Quote from GTG:
Well I don't literally mean 10 ticks or so, but that is what you mean when you say that you don't use bars....you use x number of last trade price ticks divided by y number of last trade price ticks without regard to time in any way?
GTG, I don't understand why "trading time" would not be "real-time"? To your point about using tick data, I agree with you that you could take a certain number of ticks as your base period, if not 10 then perhaps 500 or somewhere in between, sort of like tick charts instead of minute bars, but I'm not sure that is what Stephen is doing.
Maybe if we had a few stocks or futures he is trading we could look at the number of ticks during an average day, but why use tick data anyway?
So what is "evenly spaced single variable data"? Assuming the single variable is price and you are working with tick data, how would you evenly space that? It would seem that the higher frequency the data, the more noise would show up in your VRs, and thus the less significant. Again, why not use minute bars? I am not aware of any academic papers using "high frequency" data that go lower than 5 minute bars.
And how does "trendiness" and "mean-reverting" relate to the random and non-random terminology?
Maybe if we had a few stocks or futures he is trading we could look at the number of ticks during an average day, but why use tick data anyway?
So what is "evenly spaced single variable data"? Assuming the single variable is price and you are working with tick data, how would you evenly space that? It would seem that the higher frequency the data, the more noise would show up in your VRs, and thus the less significant. Again, why not use minute bars? I am not aware of any academic papers using "high frequency" data that go lower than 5 minute bars.
And how does "trendiness" and "mean-reverting" relate to the random and non-random terminology?
Im trading liquid stocks/etfs.. all have at least 2-3 mil shares a day traded.
I realize most academics never look beyond 5 or at most 1 minutes because they make absurd assumptions.
I don't use trade ticks at all, they are pretty uninformative I've found. I use full order book depth quotes from amex, inet, arca, totalview and brut.
Quotes can easily be just moved to the next period because they are valid until moved.
If I don't have a quote in one period, I just use the last value, if i have multiple quotes in a period I use some sort of averaging to arrive at a "realized" quote for that period.. which can be based on duration of quoted prices in that interval, mean, weighted, all sorts of options. I also look a bit deeper into the book to get the 'realized' prices because if the best is only showing 50 shares then its not really the best, I simulate a market order that walks the book based on the average size i'll be trading to get a realistic quote.
Any VR profile that is flat around 1 is "random", anything moving above one is trendy and anything below 1 is mean-reverting.
Sometimes they are trendy for 5 to 10 minutes and at longer horizons are completely random.
I've found that nearly nothing is completely mean reverting.. I've found below 0.8 is very tradable.
I realize most academics never look beyond 5 or at most 1 minutes because they make absurd assumptions.
I don't use trade ticks at all, they are pretty uninformative I've found. I use full order book depth quotes from amex, inet, arca, totalview and brut.
Quotes can easily be just moved to the next period because they are valid until moved.
If I don't have a quote in one period, I just use the last value, if i have multiple quotes in a period I use some sort of averaging to arrive at a "realized" quote for that period.. which can be based on duration of quoted prices in that interval, mean, weighted, all sorts of options. I also look a bit deeper into the book to get the 'realized' prices because if the best is only showing 50 shares then its not really the best, I simulate a market order that walks the book based on the average size i'll be trading to get a realistic quote.
Any VR profile that is flat around 1 is "random", anything moving above one is trendy and anything below 1 is mean-reverting.
Sometimes they are trendy for 5 to 10 minutes and at longer horizons are completely random.
I've found that nearly nothing is completely mean reverting.. I've found below 0.8 is very tradable.
Quote from mysticman:
GTG, I don't understand why "trading time" would not be "real-time"? To your point about using tick data, I agree with you that you could take a certain number of ticks as your base period, if not 10 then perhaps 500 or somewhere in between, sort of like tick charts instead of minute bars, but I'm not sure that is what Stephen is doing.
Maybe if we had a few stocks or futures he is trading we could look at the number of ticks during an average day, but why use tick data anyway?
So what is "evenly spaced single variable data"? Assuming the single variable is price and you are working with tick data, how would you evenly space that? It would seem that the higher frequency the data, the more noise would show up in your VRs, and thus the less significant. Again, why not use minute bars? I am not aware of any academic papers using "high frequency" data that go lower than 5 minute bars.
And how does "trendiness" and "mean-reverting" relate to the random and non-random terminology?