The Brent-WTI spread is a highly traded entity of its own. Always good to check the ratio and spread data when looking at “highly correlated” stuff. Theoretically you’re more concerned with cointegrated positions. The high vol of correlation measures is well documented. Check out rolling correlations for SPY and TLT 😀
Given: outliers can occur in any market, you want the largest universe of market choices available regardless of any individual correlation. Correlation is irrelevant when adding markets to the tradable instrument universe you're working with.
Assuming you're a "return maximalist" (a symptom of this is not vol adjusting your positions because it lowers returns): you don't really care if your open positions are correlated or not. The more correlated they are the more likely they are to move in tandem, but should you care? I posit you don't because each tradable instrument is a potential individual outlier in the longer term.
When does correlation matter?
Assuming the two items above are true; you want the largest universe and you don't care about correlations when "in a position," you only care about "entry/exit timing correlation." Consider the extremes: 1) at any given time only one of your tradable instruments is in a trade (in a trend) or 2) all your tradable instruments are in trends and they all drop out of trends together. In the first case your universe doesn't have enough correlation to trigger simultaneous trades. In the latter, a very high correlation drives your system to be in all tradable instruments at once. So what you really want is a low correlation of "dates spent in the trend." In other words, returns correlation is irrelevant with once caveat. Too high a correlation across your universe will drive entries to be "lumpy" and make capital allocation very difficult! Solution for this last issue, make sure your universe is either highly diversified or very large, or both. Sidebar: with a big universe it becomes important to reduce the number of active positions from the "ideal" of "take all trades." Carlos Mata on the "Trading Strategies" substack recently went over the effects of limiting positions to N (N being 10, 15, etc.) Turned out the "reduced set" performed better than the "take all trades" version.
The Darvas Box is a trading strategy and technical analysis tool developed by Nicolas Darvas in the 1950s. It is used to identify potential breakout stocks by focusing on price movements and volume.
The Darvas Box is formed when a stock’s price moves within a specific range. A box is drawn around this range, with the top line representing the highest price during a defined period and the bottom line representing the lowest price.
So it’s just “highest highs” and “lowest lows” for a period.
Wouldn't that be the definition for Donchian breakout? I read through many articles about Darvas box method but I could not find any satisfactory explanation on how it would translate into a mechanical rule. In this article the example parameters are BoxLength=10 and BoxHeightMulti=4. What does it mean? Highest high/lowest low of last 10 trading days? (seems very short term). Where does the height parameter come into play here?
So the difference is that with the Donchian a signal can occur today, again tomorrow, again the next day with subsequent break outs. With a
Box, you are requiring a consolidation phase and only want the first break out. So the rule might be that the H and L of the period are within 4 ATRs of each other. After the first break out after that consolidation - there will no longer be a box, and subsequent breakouts are not signals. You’d need to wait for another 10 (say) bars to all be constrained within a range before the next signal could occur. At least that my understanding!! : ) We can get Rich to chime in but he’s away atm.
Yes, we need Rich to shine some light on this :) As per my understanding, 10 day range will likely be within 4 ATRs more often than not (square root of 10 ~= 3.16 - you'd expect average 10 day range to be x3.16 of average daily range on average)
The time periods are not the point though - we are just trying to understand the diff between the two systems. Parameters are to be discovered by testing alone right.
I understand that it's just a random example parameters. What I'm curious about is how Darvas box is implemented as a system. All I've seen is that it is a semi-discretionary method in various online sources.
The Brent-WTI spread is a highly traded entity of its own. Always good to check the ratio and spread data when looking at “highly correlated” stuff. Theoretically you’re more concerned with cointegrated positions. The high vol of correlation measures is well documented. Check out rolling correlations for SPY and TLT 😀
Cheers Vic. Will do :-)
Outlier hunters shouldn't measure correlation.
Given: outliers can occur in any market, you want the largest universe of market choices available regardless of any individual correlation. Correlation is irrelevant when adding markets to the tradable instrument universe you're working with.
Assuming you're a "return maximalist" (a symptom of this is not vol adjusting your positions because it lowers returns): you don't really care if your open positions are correlated or not. The more correlated they are the more likely they are to move in tandem, but should you care? I posit you don't because each tradable instrument is a potential individual outlier in the longer term.
When does correlation matter?
Assuming the two items above are true; you want the largest universe and you don't care about correlations when "in a position," you only care about "entry/exit timing correlation." Consider the extremes: 1) at any given time only one of your tradable instruments is in a trade (in a trend) or 2) all your tradable instruments are in trends and they all drop out of trends together. In the first case your universe doesn't have enough correlation to trigger simultaneous trades. In the latter, a very high correlation drives your system to be in all tradable instruments at once. So what you really want is a low correlation of "dates spent in the trend." In other words, returns correlation is irrelevant with once caveat. Too high a correlation across your universe will drive entries to be "lumpy" and make capital allocation very difficult! Solution for this last issue, make sure your universe is either highly diversified or very large, or both. Sidebar: with a big universe it becomes important to reduce the number of active positions from the "ideal" of "take all trades." Carlos Mata on the "Trading Strategies" substack recently went over the effects of limiting positions to N (N being 10, 15, etc.) Turned out the "reduced set" performed better than the "take all trades" version.
This article was in part inspired by our discussion with Perry Kaufman, who presents an alternative approach! Suggest listening to that here: https://thealgorithmicadvantage.substack.com/p/a-wealth-of-experience-trading-diversified.
What is this Darvas Box breakout system you mention in the article? Why does it have box length and height parameters?
The Darvas Box is a trading strategy and technical analysis tool developed by Nicolas Darvas in the 1950s. It is used to identify potential breakout stocks by focusing on price movements and volume.
The Darvas Box is formed when a stock’s price moves within a specific range. A box is drawn around this range, with the top line representing the highest price during a defined period and the bottom line representing the lowest price.
So it’s just “highest highs” and “lowest lows” for a period.
Wouldn't that be the definition for Donchian breakout? I read through many articles about Darvas box method but I could not find any satisfactory explanation on how it would translate into a mechanical rule. In this article the example parameters are BoxLength=10 and BoxHeightMulti=4. What does it mean? Highest high/lowest low of last 10 trading days? (seems very short term). Where does the height parameter come into play here?
So the difference is that with the Donchian a signal can occur today, again tomorrow, again the next day with subsequent break outs. With a
Box, you are requiring a consolidation phase and only want the first break out. So the rule might be that the H and L of the period are within 4 ATRs of each other. After the first break out after that consolidation - there will no longer be a box, and subsequent breakouts are not signals. You’d need to wait for another 10 (say) bars to all be constrained within a range before the next signal could occur. At least that my understanding!! : ) We can get Rich to chime in but he’s away atm.
Yes, we need Rich to shine some light on this :) As per my understanding, 10 day range will likely be within 4 ATRs more often than not (square root of 10 ~= 3.16 - you'd expect average 10 day range to be x3.16 of average daily range on average)
The time periods are not the point though - we are just trying to understand the diff between the two systems. Parameters are to be discovered by testing alone right.
I understand that it's just a random example parameters. What I'm curious about is how Darvas box is implemented as a system. All I've seen is that it is a semi-discretionary method in various online sources.
What system was this backtested in?