Refinements to Hedgefundie's excellent approach

Discuss all general (i.e. non-personal) investing questions and issues, investing news, and theory.
MotoTrojan
Posts: 5495
Joined: Wed Feb 01, 2017 8:39 pm

Re: Refinements to Hedgefundie's excellent approach

Post by MotoTrojan » Wed Jul 17, 2019 1:41 pm

Hydromod wrote:
Wed Jul 17, 2019 1:34 pm
MotoTrojan wrote:
Wed Jul 17, 2019 12:38 pm
Those sorts of changes are a lot more hand-wavey IMHO. TMV and SPXU both have expected negative returns, not something I want to play with. Dynamic allocation/volatility though shows some promise and still maintains a positive expected return.
It's pretty clear that one should want to use the funds that have expected positive returns.

It's also true that TMV and SPXU have expected negative returns under current conditions (1982 to present). But the question is what happens if conditions change?
Long-term treasury yields are set by the market expectation and thus always have a positive expected return. Inverse funds thus will always have a negative expected return. Do not mix up expected return and realized return. TMV and SPXU will always have a negative expected return.

If you held long-treasuries forever you'd expect to make money. If you held an inverse of that forever, you'd expect to lose money. Of course many of us aren't investing forever, so it is possible that actual/realized returns will not be positive.

Topic Author
Hydromod
Posts: 152
Joined: Tue Mar 26, 2019 10:21 pm

Re: Refinements to Hedgefundie's excellent approach

Post by Hydromod » Wed Jul 17, 2019 1:57 pm

MotoTrojan wrote:
Wed Jul 17, 2019 1:41 pm
Long-term treasury yields are set by the market expectation and thus always have a positive expected return. Inverse funds thus will always have a negative expected return. Do not mix up expected return and realized return. TMV and SPXU will always have a negative expected return.

If you held long-treasuries forever you'd expect to make money. If you held an inverse of that forever, you'd expect to lose money. Of course many of us aren't investing forever, so it is possible that actual/realized returns will not be positive.
Hydromod wrote:
Wed Jul 10, 2019 12:10 pm
The monthly values used are: (i) portfolio 2: LTT x3, (ii) portfolio 3: LTT x-3, and (iii) portfolio 1: LTT itself, for reference.

Image
From a pragmatic approach, I'm interested in strategies that will consistently work over five to ten year periods, with or without inflation. For the 30 years prior to 1982, TMV would have been a better bet than TMF. There's a pretty good chance I won't be here thirty years hence...

If I was investing in 1970, say, and had these options available, I'd be saying something like "why would you be investing in LTT or TMF? They've been dropping steadily for 15 years!"

Topic Author
Hydromod
Posts: 152
Joined: Tue Mar 26, 2019 10:21 pm

Re: Refinements to Hedgefundie's excellent approach

Post by Hydromod » Fri Aug 02, 2019 4:22 pm

This is an entry on schemes for addressing volatility with UPRO and TMF, using the UPROSIM and TMFSIM daily simulated dataset previously discussed. This dataset runs from late 1986 through the start of 2019.

I was interested in quantifying the effect of different rebalance periods, volatility windows, and volatility limiting schemes with the UPRO/TMF risk parity scheme. There’s been some discussion of this in the Hedgefundie thread.

To compare apples to apples, I used a consistent time step for rebalancing, considering 10, 20, 40, and 60 days (biweekly to approximately quarterly). To assess the effect of start date, I assumed that the first N days were available to start a sequence and I calculated N sequences, where N is the number of days between rebalancing. If N is 10, I had 10 sequences starting on days 1 through 10 of the dataset. If N is 60, I started sequences on days 1 through 60 of the dataset. This gives me a number of replicates that should have almost identical CAGR.

I considered four different moving windows for calculating volatility (again 10, 20, 40, and 60 days).

I compared the two schemes, symmetric volatility weighting and downward volatility weighting. The symmetric weighting scheme uses the standard deviation of all daily UPRO and TMF returns in the moving volatility window. The downward weighting uses the standard deviation of all daily returns in the moving volatility window, except that returns larger than zero are reset to zero.

I compared the two basic schemes with a volatility-limited scheme that resets the weights when the UPRO standard deviation SDu is larger than a limiting value SDlim. For this, I considered a range of volatility limiting values.

The first idea I checked was resetting the UPRO weight wtu to SDu / SDlim when SDu > SDlim. The TMF weight is simply 1 – wtu. I think that this is the way it is implemented for a riskless asset (CASH instead of TMF). After playing with it a bit, I realized that the downward volatility scheme works better with a criterion of SDlim/2 to account for the smaller standard deviation.

It turned out that the different schemes give almost identical returns on most days, but rolling 5-year CAGRs can be very different. A single scheme can give significantly different rolling CAGR for different sequences, depending on which day that the sequence started. The discrepancy is more marked when comparing different schemes. This is because the volatility limit doesn’t kick in for most days, so occasional large rare differences don’t have a chance to even out. A large event (e.g., Black Monday) that is treated differently with different schemes can strongly affect the rolling CAGR for the next 5 years.

I found that the volatility-limiting scheme tended to deliver worse CAGR than the corresponding scheme without limits, which is unexpected, although for a few combinations it delivered much better CAGR. CAGR values also tend to bounce all over with different volatility limit values. I hesitate to use the scheme because of the inconsistent behavior.

I then came up with the idea of limiting total portfolio variance, including UPRO and TMF, with the scheme

SDlim^2 = w SDu^2 + (1 – w) SDt^2

where SDt is the standard deviation of TMF. This gives a weight of

w = (SDlim^2 – SDt^2) / (SDu^2 – SDt^2)

I only used this when SDu > SDlim. This scheme tended to deliver better CAGR, with more consistent behavior.

The following figure shows the four schemes for various combinations of rebalancing period (plot columns) and volatility window (plot rows) for a volatility limit of 8 percent. The orange and red lines represent the symmetric volatility model with (orange) and without (red) volatility limiting. The cyan and blue lines represent the downward volatility model with (cyan) and without (blue) volatility limiting. The gray line represents the 40/60 scheme with the same rebalance frequency.

In these plots, the gray line is slanted because the different start dates have different returns. Even the difference of start date within a single quarter gives a 32-year CAGR that ranges from 14 to 16.5 percent. This is an example that shows how important it is to be consistent in representing starting points when comparing different schemes.

Image

Takeaways from the figure:

• The symmetric volatility scheme consistently appears to have yielded higher CAGR than the downward volatility scheme, which was not what I expected.
• Both the symmetric and downward volatility schemes consistently had one or two percent higher CAGR than the 40/60 scheme.
• The volatility limit schemes do not necessarily have a consistent benefit, but bounce around based on volatility window and rebalance frequency.
• Increasing the volatility window gives more consistent CAGR values.

The following figures consider different volatility limits. In these figures, the lines represent the difference in CAGR between the volatility scheme and the fixed 40/60 scheme. All colors represent the same schemes as the first figure.

Image

Image

Image

Image

Image

Takeaways from the series of figures

• The volatility limit of 8 percent appears to be a relatively good value for the symmetric scheme. The scheme seems to struggle in some cases with smaller limits, and has less benefit for larger values.
• The downward volatility scheme is much more sensitive to the volatility limit than the symmetric scheme.
• Monthly rebalances arguably do best with 60-day windows and 8 percent volatility window.
• Quarterly rebalances arguably do best with 20-day windows and 10 percent volatility window.
• The benefit of the volatility limit decreases as the rebalance frequency becomes longer, and there may not be much benefit to a volatility limit scheme with quarterly rebalances.

Hope this proves interesting for folks.

Topic Author
Hydromod
Posts: 152
Joined: Tue Mar 26, 2019 10:21 pm

Re: Refinements to Hedgefundie's excellent approach

Post by Hydromod » Sun Aug 04, 2019 11:22 pm

I want to flesh out the idea of a confidence index that can be used to adjust the weights in a stock/fund portfolio, in particular UPRO/TMF. The index should enhance weighting to equities under favorable macroeconomic conditions and enhance weighting to low-risk funds under unfavorable macroeconomic conditions.

The idea is to fuss with the UPRO weighting that is determined through other means, such as risk parity. Let U0 be the nominal weight calculated with other means. With the confidence index CI, where -1 <= CI <= 1, the weight U actually applied becomes

U = U0 + F (ULIM0 – U0)

where 0 <= F <= 1 is a weighting factor and ULIM0 is a limit based on the current value of CI. ULIM0 = 1 when CI > 0 and ULIM0 = 0 when CI < 0.

The value F = 1 in the market timing approach favored by willthrill81, which means that the weights for UPRO are 1 with CI > 0 and 0 with CI < 0.

I propose a more gradual approach, adjusting F incrementally according to changes in CI. I propose using an index f, where (i) f_lo <= f <= f_hi, (ii) -1 <= f_lo <= 0, (iii) 0 <= f_hi <= 1, and (iv) F = abs(f). It might be reasonable to use f_lo = -0.5 and f_hi = 0.5, which equally combines the risk parity and market timing approaches.

Updates to f are signaled based on the rate of change in the CI. The value of f is decremented with a signal that economic downturn is occurring or imminent, and incremented with another signal that economic robustness is occurring or imminent. Increments and decrements are made up to a limiting value.

The unemployment index (UEI) has long been recognized as a leading indicator of recessions. Typically recessions occur after the UEI begins to rise after a period of falling UEI, and recovery begins occurs prior to the peak of the UEI. The rise leads the recession by zero to several months. This is the index used by willthrill81.

The confidence index should indicate weighting towards equities during falling or flat UEI, and should transition to weighting towards bonds when the UEI starts to rise.

Typically the UEI maximum has lagged the start of the equity rebound by a substantial fraction of a year. However, it appears that the start of economic recovery is also associated with a reduced rate of increase in the UEI. This is consistent with recovery starting to increase employment, slowing the increase in UEI. So the inflection point in the rate of change in an increasing UEI is an indicator of recovery.

Usually the UEI is a bit noisy from month to month. Also, the initial estimates may be changed once or twice as more data comes in, so the data available for backtesting may not be the same as is available in real time. Accordingly, back testing results may be a little optimistic.

Moving averages are typically used to detect the upswing from the UEI minimum. Subtracting the moving average from the latest monthly UEI value gives a proxy value that has the same sign as the slope of the UEI. The signal for a potential recession is onset of a positive slope. Generally the moving average is over a year, plus or minus a few months.

The inflection point in the rate of UEI increase is a bit more subtle to detect. Again moving averages can be used. In this case, the difference between successive values of the slope proxy is used. The inflection point is the first time when the current slope is less than the previous slope, even when both slopes are positive. The moving average may be a bit shorter for this calculation, at the cost of more noise. Three successive values of the moving average could be used to estimate a curvature and the criterion would be the change in sign of the curvature.

So the updating rules are pretty simple each month.

If both the UEI slope and curvature are positive, decrement f until f_lo is reached.

If either the UEI slope or curvature are negative, increment f until f_hi is reached.

It remains for the user to select the step size. Presumably the step should be large enough that only a few months are needed to go from f = f_hi to f = f_lo. It remains to be explored if there are strategies for using different steps for the increment and decrement, or different steps when going towards F = 0 (return to nominal from an extreme) and away from F = 0 (go to extreme from the nominal).

I have two figures, one for 1x S&P500 and LTT and the other for 3x (UPRO/TMF). Both of these are based on monthly data from siamond’s dataset. For clarity, I consider from January 1982 through January 2019, omitting the much poorer performance prior to 1982.

Leveraged 1x

Image

Leveraged 3x

Image

In the top plot for both figures, I show the total return for several concepts, all with monthly rebalance. The yellow and purple lines are the equity and bond returns. The gray line is a standard equity/bond weighted portfolio (60/40 for 1x, 40/60 for 3x). The red line is a parity line using variance instead of standard deviation. The heavy blue line is the parity case, adjusted by taking into account the UEI information.

I assumed that both the previous month’s returns and unemploymenet rate were available when calculating the weights used for the current month. This may be a little optimistic, because there is typically a few days lag before reporting the index.

The middle and bottom plots are the same in both figures. The middle plot indicates the UEI, with red dots flagged as increasing recession and the blue dots flagged as decreasing recession. The bottom plot is the unadjusted equity weight (red) and the adjusted equity weight (blue).

For these plots, I allowed the equity weights to rise halfway to one and to drop to zero. After playing with possibilities a bit, I selected a drop rate to go from fully optimistic to fully pessimistic in two months. The rise occurs over 20 months for these plots, although good results were obtained from 4 months to 25 months (another good value was a rise over four months). The averaging period for calculating recession start was 16 months, and the averaging period for calculating recession termination was 6 months.

For the 1x case, the overall CAGR values for 1982 to 2019 were

13.2 Variance parity adjusted with UEI index
11.9 Variance parity
11.4 Equity
9.44 LTT
11.1 60/40 weighting

For the 3x case, the overall CAGR values for 1982 to 2019 were

27.3 Variance parity adjusted with UEI index
23.6 Variance parity
14.8 UPRO
14.7 TMF
19.5 40/60 weighting

So there appears to be some merit to using the UEI adjustment. Presumably the improvement over straight parity will be less in the future, without the benefit of hindsight.

I am reluctant to set the maximum UPRO weight to 100%, because of the possibility of events such as Black Monday that are very poorly handled.

perplexed
Posts: 53
Joined: Thu Feb 02, 2012 4:32 pm

Re: Refinements to Hedgefundie's excellent approach

Post by perplexed » Mon Aug 05, 2019 11:25 am

Excellent post. Thank you.
In principle, I like incorporating macro indicators into the allocation formula, thus going away from the sole dependence on pricing stats.


A few questions from a novice.

Any chance you may be able to post stddev, drawdowns, and sortino?

Topic Author
Hydromod
Posts: 152
Joined: Tue Mar 26, 2019 10:21 pm

Re: Refinements to Hedgefundie's excellent approach

Post by Hydromod » Mon Aug 05, 2019 11:47 am

Shouldn't be a problem. I'll see what I can do tonight.

perplexed
Posts: 53
Joined: Thu Feb 02, 2012 4:32 pm

Re: Refinements to Hedgefundie's excellent approach

Post by perplexed » Tue Aug 06, 2019 4:34 pm

Curious, if incorporating unemployment index is a possibility in the PV?
That way I can simply click and get the allocation monthly.

Of course, if you are willing to share your matlab files .... many blessings!

Regards!

Topic Author
Hydromod
Posts: 152
Joined: Tue Mar 26, 2019 10:21 pm

Re: Refinements to Hedgefundie's excellent approach

Post by Hydromod » Tue Aug 06, 2019 4:52 pm

perplexed wrote:
Tue Aug 06, 2019 4:34 pm
Curious, if incorporating unemployment index is a possibility in the PV?
That way I can simply click and get the allocation monthly.

Of course, if you are willing to share your matlab files .... many blessings!

Regards!
I fell asleep last night before getting anything done, and got sidetracked by looking at bands because of the eventful market yesterday. Sorry about that.

I don't think there is any way to use the unemployment index in PV, except maybe as a plotting comparison.

I can give you my matlab files if you want. They're a little untidy and poorly documented, since they are only for quick and dirty private consumption. I'd probably want to clean them up a bit first.

It would be easy enough to implement the indicator in Excel. You'd have to fuss a bit to combine the monthly indicator with the daily UPRO and TMF series though.

perplexed
Posts: 53
Joined: Thu Feb 02, 2012 4:32 pm

Re: Refinements to Hedgefundie's excellent approach

Post by perplexed » Tue Aug 06, 2019 5:17 pm

Thank you for your kindness. MATLAB is better!

While looking at target volatility, I saw that some periods (months) when UPRO (or TQQQ) are at a high allocation, the portfolio suddenly goes through a bad month. However, it is not always the case (and hence target vol wins over 4/60)! I wonder, if there are other indicators, such as unemployment index, to withhold high allocation in those months.

Yeah last two days have been interesting. 40/60 hands down better than target vol (end of month transactions) for these two days.

MotoTrojan
Posts: 5495
Joined: Wed Feb 01, 2017 8:39 pm

Re: Refinements to Hedgefundie's excellent approach

Post by MotoTrojan » Tue Aug 06, 2019 6:44 pm

perplexed wrote:
Tue Aug 06, 2019 5:17 pm
Thank you for your kindness. MATLAB is better!

While looking at target volatility, I saw that some periods (months) when UPRO (or TQQQ) are at a high allocation, the portfolio suddenly goes through a bad month. However, it is not always the case (and hence target vol wins over 4/60)! I wonder, if there are other indicators, such as unemployment index, to withhold high allocation in those months.

Yeah last two days have been interesting. 40/60 hands down better than target vol (end of month transactions) for these two days.
I think it was March or May of this year which had an almost -19% month with 100% UPRO via vol targeting... it'll happen, but then your average equity exposure over the last decade was closer to 55% which boosted returns quite a bit. It also avoided the December 2018 incident thanks to the volatility (but modest returns) that precluded it.

User avatar
LadyGeek
Site Admin
Posts: 55784
Joined: Sat Dec 20, 2008 5:34 pm
Location: Philadelphia
Contact:

Re: Refinements to Hedgefundie's excellent approach

Post by LadyGeek » Tue Aug 06, 2019 6:50 pm

For the MATLAB fans, may I recommend the open-source competitor? See: GNU Octave

I haven't run Octave in a very long time, but I know you can get it to work.
Hydromod wrote:
Tue Aug 06, 2019 4:52 pm
...I can give you my matlab files if you want. They're a little untidy and poorly documented, since they are only for quick and dirty private consumption. I'd probably want to clean them up a bit first.
See this wiki article: Using open source software for portfolio analysis

If you are willing to share your files (after clean up), upload it to a file sharing site (like Google Drive) and share the link. Include a description in your post and I'll add it to the wiki.

(This offer is for anyone who would like to share a script which uses open-source software. R, python, etc.)
Wiki To some, the glass is half full. To others, the glass is half empty. To an engineer, it's twice the size it needs to be.

Topic Author
Hydromod
Posts: 152
Joined: Tue Mar 26, 2019 10:21 pm

Re: Refinements to Hedgefundie's excellent approach

Post by Hydromod » Wed Aug 07, 2019 7:53 pm

perplexed wrote:
Mon Aug 05, 2019 11:25 am
Any chance you may be able to post stddev, drawdowns, and sortino?
I have the same plots with added information. It took a bit of coding to get here, and I've been fussing with volatility things. Sorry for the wait.

The heavy blue lines represent the risk-parity model with risk adjustments based on the UEI, the red lines represent the same model without the UEI adjustment, and the gray line is the 40/60 model. Orange is S&P and purple is LTT.

The right column is (i) 5-year moving CAGR, (ii) drawdown, and (iii) a Sortino ratio.

It seems that there are various ways to calculate the Sortino ratio. In this case, I subtracted the 1x or 3x LTT CAGR from the other 1x or 3x asset CAGRs, and I used the 5-year-moving annualized downward standard deviation based on just the subset of monthly returns that were negative.

Note that the 2008 drawdown for the UEI model looks severe, but it is due to a spike that dissipates rapidly. Not at all the same as the other models.

1x case

Image

3x case

Image

Hope you find this useful.

klaus14
Posts: 226
Joined: Sun Nov 25, 2018 7:43 pm

Re: Refinements to Hedgefundie's excellent approach

Post by klaus14 » Thu Aug 08, 2019 12:35 am

Hydromod wrote:
Wed Aug 07, 2019 7:53 pm
perplexed wrote:
Mon Aug 05, 2019 11:25 am
Any chance you may be able to post stddev, drawdowns, and sortino?
I have the same plots with added information. It took a bit of coding to get here, and I've been fussing with volatility things. Sorry for the wait.

The heavy blue lines represent the risk-parity model with risk adjustments based on the UEI, the red lines represent the same model without the UEI adjustment, and the gray line is the 40/60 model. Orange is S&P and purple is LTT.

The right column is (i) 5-year moving CAGR, (ii) drawdown, and (iii) a Sortino ratio.

It seems that there are various ways to calculate the Sortino ratio. In this case, I subtracted the 1x or 3x LTT CAGR from the other 1x or 3x asset CAGRs, and I used the 5-year-moving annualized downward standard deviation based on just the subset of monthly returns that were negative.

Note that the 2008 drawdown for the UEI model looks severe, but it is due to a spike that dissipates rapidly. Not at all the same as the other models.

1x case

Image

3x case

Image

Hope you find this useful.
Hi Hydromod,
If you summarize your main findings in simple language, would be very helpful for readers.

Topic Author
Hydromod
Posts: 152
Joined: Tue Mar 26, 2019 10:21 pm

Re: Refinements to Hedgefundie's excellent approach

Post by Hydromod » Thu Aug 08, 2019 9:17 am

klaus14 wrote:
Thu Aug 08, 2019 12:35 am
Hi Hydromod,
If you summarize your main findings in simple language, would be very helpful for readers.
Sorry, these plots were basically a requested follow-on to previous plots. I'll try to summarize in one post.

All of this thread is devoted to exploring refinements to the Hedgefundie 40/60 UPRO/TMF quarterly rebalance strategy. I've accepted the idea of using adaptive weights to adjust the weights up and down based on trailing volatility (e.g., risk parity strategies), and perhaps more frequent rebalancing.

I'm exploring with these last plots the feasibility and reliability of incorporating macroeconomic information into what I'll call an Equity Confidence Index (ECI). Ideally the ECI would point to climates when equities are likely to perform well and when they are likely to perform poorly relative to long-term treasuries. If this could be done with some confidence, I'd have an actionable basis for tilting towards equities under favorable climates and tilting away from equities under unfavorable climates. These tilts would represent adjustments to whatever weights are otherwise calculated based on volatility (e.g., adjustments to the 40/60 or risk parity weights).

I'm looking at both 1x and 3x leveraged ETFs here.

willthrill81's market timing thread pointed to the unemployment index (UEI) as a fairly robust and well-known leading indicator of recessions, which points to a strategy for reducing market exposure. This market timing approach basically switches from 100% equity to 100% LTT, potentially from month to month, based on UEI indicators. I'm trying to understand this process and see if I can use the UEI to develop an ECI that gives an actionable strategy for increasing market exposure, reducing the tendency for whipsawing back and forth that market timing strategies are prone to.

I outlined a strategy some posts back that sets the equity weight preferentially towards 1 or towards 0 based on the behavior of the UEI. It appears that a reasonably actionable strategy is to (i) incrementally change the weight towards 0 if the UEI is increasing AND accelerating and (ii) incrementally change the weight towards 1 if the UEI is decreasing OR decelerating.

I set limits on how much the increment can be, and set limits on how much the adjustment to the equity weight can be. I've compromised on allowing the equity weights to increase a maximum of halfway to 1 on the up side, and allowing the equities to completely drop out on the down side. The limit on upside equity weighting reduces the consequences of black swans like Black Monday in 1987. Going from one extreme to the other over four increments seems to have been reasonable, although under at least one period there appeared to have been significant benefit to being much slower on the rise.

The plots compare the basic behavior of the 40/60, straight risk parity, and the ECI-adjusted risk parity models.

Basically the 40/60 approach mitigates the strong volatility in equities, but still can suffer some large drawdowns. The risk-parity approach further flattens out the ups and downs, with the better behavior in recessions outweighing the somewhat poorer performance near the ends of bull markets.

The ECI-adjusted approach further reduces the drawdowns during recessions, while taking better advantage of bull markets. The three plots in the right column of the figures shows that the ECI-adjusted performance tends to be better than both the 40/60 and risk-parity approaches most of the time from 1982 through the start of 2019.

Although I haven't shown it, I've tried the approach with several different risk parity approaches. The ECI adjustment seems to work similarly with all of the approaches I've tried, as might be expected. Early tests suggest roughly a 10% increase in CAGR would have been seen, maybe as much as 15%.

Hope this helps clarify the idea.

perplexed
Posts: 53
Joined: Thu Feb 02, 2012 4:32 pm

Re: Refinements to Hedgefundie's excellent approach

Post by perplexed » Thu Aug 08, 2019 9:37 am

Thank you so much! You really have no reason to be sorry! :-)

Couple of quick pointers back. I may be mistaken in these, so please do correct me if I am wrong.

* It looks like 2009 recession, UEI enabled allocation had a greater drawdown (blue line looks to go lower than red). Same I see in 1987ish (?). Is that correct? Is this a technicality from a slight lag of UEI inflection or second derivative impacting the allocation? Or are there more deeper insights? From the formula you gave, it seems more of a linear adjustment to allocation from a UEI. A stronger or step wise adjustment may fix this?

* I think you are plotting Sharpe rather than sortino (?), or is just a Y-label thingy. Sortino only looks at downward variance relative to performance ...

This is beautiful work, Hydromod.

Topic Author
Hydromod
Posts: 152
Joined: Tue Mar 26, 2019 10:21 pm

Re: Refinements to Hedgefundie's excellent approach

Post by Hydromod » Thu Aug 08, 2019 9:58 am

perplexed wrote:
Thu Aug 08, 2019 9:37 am
Thank you so much! You really have no reason to be sorry! :-)

Couple of quick pointers back. I may be mistaken in these, so please do correct me if I am wrong.

* It looks like 2009 recession, UEI enabled allocation had a greater drawdown (blue line looks to go lower than red). Same I see in 1987ish (?). Is that correct? Is this a technicality from a slight lag of UEI inflection or second derivative impacting the allocation? Or are there more deeper insights? From the formula you gave, it seems more of a linear adjustment to allocation from a UEI. A stronger or step wise adjustment may fix this?

You are correct in both observations. The 2009 one is misleading; there was a fast overall spike in LTT that bounced the risk allocations for a short time. The UEI-enabled spiked higher before returning to its starting point, so technically its drawdown was lower. However, UEI-enabled increased overall from 2008 to 2010 while the other decreased, so I'd say it overall did better.

The Black Monday 1987 is the example where the UEI approach is helpless. It was out of the blue. That's precisely my example of a black swan that points to not going 100% equities.


* I think you are plotting Sharpe rather than sortino (?), or is just a Y-label thingy. Sortino only looks at downward variance relative to performance ...

It's a Y-label thingy. To be clear, I did the standard deviation version with just downward returns rather than the version with upward returns included but set to zero.

This is beautiful work, Hydromod.

perplexed
Posts: 53
Joined: Thu Feb 02, 2012 4:32 pm

Re: Refinements to Hedgefundie's excellent approach

Post by perplexed » Thu Aug 08, 2019 11:24 am

Great! Thank you.
Today is a good day for monthly setting from target vol --- going to whole 80% upro :-)

MotoTrojan
Posts: 5495
Joined: Wed Feb 01, 2017 8:39 pm

Re: Refinements to Hedgefundie's excellent approach

Post by MotoTrojan » Thu Aug 08, 2019 11:33 am

perplexed wrote:
Thu Aug 08, 2019 11:24 am
Great! Thank you.
Today is a good day for monthly setting from target vol --- going to whole 80% upro :-)
I went 80% on 7/31 and while today is nice, this week wasn't :sharebeer .

perplexed
Posts: 53
Joined: Thu Feb 02, 2012 4:32 pm

Re: Refinements to Hedgefundie's excellent approach

Post by perplexed » Thu Aug 08, 2019 11:35 am

One more day before this week is over ...! :moneybag

klaus14
Posts: 226
Joined: Sun Nov 25, 2018 7:43 pm

Re: Refinements to Hedgefundie's excellent approach

Post by klaus14 » Thu Aug 08, 2019 12:07 pm

Hydromod wrote:
Thu Aug 08, 2019 9:17 am
klaus14 wrote:
Thu Aug 08, 2019 12:35 am
Hi Hydromod,
If you summarize your main findings in simple language, would be very helpful for readers.
Sorry, these plots were basically a requested follow-on to previous plots. I'll try to summarize in one post.

All of this thread is devoted to exploring refinements to the Hedgefundie 40/60 UPRO/TMF quarterly rebalance strategy. I've accepted the idea of using adaptive weights to adjust the weights up and down based on trailing volatility (e.g., risk parity strategies), and perhaps more frequent rebalancing.

I'm exploring with these last plots the feasibility and reliability of incorporating macroeconomic information into what I'll call an Equity Confidence Index (ECI). Ideally the ECI would point to climates when equities are likely to perform well and when they are likely to perform poorly relative to long-term treasuries. If this could be done with some confidence, I'd have an actionable basis for tilting towards equities under favorable climates and tilting away from equities under unfavorable climates. These tilts would represent adjustments to whatever weights are otherwise calculated based on volatility (e.g., adjustments to the 40/60 or risk parity weights).

I'm looking at both 1x and 3x leveraged ETFs here.

willthrill81's market timing thread pointed to the unemployment index (UEI) as a fairly robust and well-known leading indicator of recessions, which points to a strategy for reducing market exposure. This market timing approach basically switches from 100% equity to 100% LTT, potentially from month to month, based on UEI indicators. I'm trying to understand this process and see if I can use the UEI to develop an ECI that gives an actionable strategy for increasing market exposure, reducing the tendency for whipsawing back and forth that market timing strategies are prone to.

I outlined a strategy some posts back that sets the equity weight preferentially towards 1 or towards 0 based on the behavior of the UEI. It appears that a reasonably actionable strategy is to (i) incrementally change the weight towards 0 if the UEI is increasing AND accelerating and (ii) incrementally change the weight towards 1 if the UEI is decreasing OR decelerating.

I set limits on how much the increment can be, and set limits on how much the adjustment to the equity weight can be. I've compromised on allowing the equity weights to increase a maximum of halfway to 1 on the up side, and allowing the equities to completely drop out on the down side. The limit on upside equity weighting reduces the consequences of black swans like Black Monday in 1987. Going from one extreme to the other over four increments seems to have been reasonable, although under at least one period there appeared to have been significant benefit to being much slower on the rise.

The plots compare the basic behavior of the 40/60, straight risk parity, and the ECI-adjusted risk parity models.

Basically the 40/60 approach mitigates the strong volatility in equities, but still can suffer some large drawdowns. The risk-parity approach further flattens out the ups and downs, with the better behavior in recessions outweighing the somewhat poorer performance near the ends of bull markets.

The ECI-adjusted approach further reduces the drawdowns during recessions, while taking better advantage of bull markets. The three plots in the right column of the figures shows that the ECI-adjusted performance tends to be better than both the 40/60 and risk-parity approaches most of the time from 1982 through the start of 2019.

Although I haven't shown it, I've tried the approach with several different risk parity approaches. The ECI adjustment seems to work similarly with all of the approaches I've tried, as might be expected. Early tests suggest roughly a 10% increase in CAGR would have been seen, maybe as much as 15%.

Hope this helps clarify the idea.
Thanks a lot for writing this up!

So instead of 40/60 you can start with -let's say- 50/50 because volatilities are equal for some reason.
If you get the up signal from UEI, you may adjust this up to 75/25.
If you get the down signal, you may adjust this down to 0/100.
You do adjustments every month incrementally based on the strength of the signal.

Topic Author
Hydromod
Posts: 152
Joined: Tue Mar 26, 2019 10:21 pm

Re: Refinements to Hedgefundie's excellent approach

Post by Hydromod » Thu Aug 08, 2019 12:53 pm

klaus14 wrote:
Thu Aug 08, 2019 12:07 pm
Thanks a lot for writing this up!

So instead of 40/60 you can start with -let's say- 50/50 because volatilities are equal for some reason.
If you get the up signal from UEI, you may adjust this up to 75/25.
If you get the down signal, you may adjust this down to 0/100.
You do adjustments every month incrementally based on the strength of the signal.
That's precisely the idea. The volatility adjustment and the UEI signal move independently.

Forester
Posts: 260
Joined: Sat Jan 19, 2019 2:50 pm
Location: UK

Re: Refinements to Hedgefundie's excellent approach

Post by Forester » Sat Aug 10, 2019 5:26 am

There was a guy on the Alpha Architect channel with a timing indicator. https://papers.ssrn.com/sol3/papers.cfm ... id=3164081

Using CBOT & CME data, hedgers (smart money) vs nonreportable speculators (dumb money) in 1) S&P futures 2) 30yr treasuries 3) 10yrVs30yr, the three relative positions combined to create one binary signal.

This could be a third modifier. The relevant data is released weekly.

From the paper,
Notably, from 2011–2017, a time period over which the U.S. equity market rose strongly— thus making it potentially detrimental not to have been fully invested in equities—the SMI strategy outpaced its nearest competitor, TREND, by approximately 75 basis points per year while having a time-averaged equity exposure of only 61% compared to 101% for TREND. Moreover, the maximum drawdown of the SMI strategy over this time period, -10.7%, was the lowest drawdown of all the strategies tested, and half the magnitude of the TREND strategy’s -21.2% drawdown.

unranked
Posts: 2
Joined: Sun Apr 14, 2019 1:25 am

Re: Refinements to Hedgefundie's excellent approach

Post by unranked » Sun Aug 11, 2019 4:06 am

Amazing work. I have always had similar ideas but never had the motivation to flesh it out empirically, and with such rigour too!

Allow me to paraphrase my interpretation of the ECI-adjustment scheme, to make sure I did not misunderstand you.

1) Construct a base risk parity portfolio.
Required data: Daily return series of equities and bonds
Parameters: Lookback window


Suppose we have selected a window of 20 days. Say, we observe the standard deviation of the daily returns over the past 20 days for equities (SD_E) is triple that of bonds (SD_B). Then, we will designate a weight of 25/75 to equities/bonds. More generally, our allocation to equities is (1/SD_E) / (1/SD_E + 1/SD_B) and likewise for bonds.

2) Tilt base weights using UE information.
Required data: Unemployment rate
Parameters: Bounds, tilting speed


Classify the current economic regime as either favourable or unfavourable for equities, using unemployment as an indicator. More specifically, the regime is unfavourable if and only if both the first and second derivatives are increasing. From here on, decisions are more or less arbitrary. You have set the bounds to between 0/100 (unfavourable) and 75/25 (favourable), with a step size that would allow you to transition between these two extreme states in a few months.

Is this an accurate characterisation of your strategy? I have several ideas for how this could be extended, but would like to make sure that I am on the same page before writing any further. The endgoal in my mind here is the creation of a fully systematic, tax-efficient implementation of a robust investment strategy that is grounded in sound logic, requiring minimal human discretion when making trading decisions.

Many thanks for your effort, Hydromod.

Topic Author
Hydromod
Posts: 152
Joined: Tue Mar 26, 2019 10:21 pm

Re: Refinements to Hedgefundie's excellent approach

Post by Hydromod » Sun Aug 11, 2019 1:41 pm

unranked wrote:
Sun Aug 11, 2019 4:06 am
Amazing work. I have always had similar ideas but never had the motivation to flesh it out empirically, and with such rigour too!

Allow me to paraphrase my interpretation of the ECI-adjustment scheme, to make sure I did not misunderstand you.

1) Construct a base risk parity portfolio.
Required data: Daily return series of equities and bonds
Parameters: Lookback window


Suppose we have selected a window of 20 days. Say, we observe the standard deviation of the daily returns over the past 20 days for equities (SD_E) is triple that of bonds (SD_B). Then, we will designate a weight of 25/75 to equities/bonds. More generally, our allocation to equities is (1/SD_E) / (1/SD_E + 1/SD_B) and likewise for bonds.

2) Tilt base weights using UE information.
Required data: Unemployment rate
Parameters: Bounds, tilting speed


Classify the current economic regime as either favourable or unfavourable for equities, using unemployment as an indicator. More specifically, the regime is unfavourable if and only if both the first and second derivatives are increasing. From here on, decisions are more or less arbitrary. You have set the bounds to between 0/100 (unfavourable) and 75/25 (favourable), with a step size that would allow you to transition between these two extreme states in a few months.

Is this an accurate characterisation of your strategy? I have several ideas for how this could be extended, but would like to make sure that I am on the same page before writing any further. The endgoal in my mind here is the creation of a fully systematic, tax-efficient implementation of a robust investment strategy that is grounded in sound logic, requiring minimal human discretion when making trading decisions.

Many thanks for your effort, Hydromod.
You have the idea.

I've since found a nice formulation that allows the economic regime to be combined seamlessly with the volatility data. This is based on the idea of a risk budget allocation.

The budget formulation for risk parity is (b_i/V_i)^m / (sum_j (b_j/V_j)^m), where V is variance, b is risk budget (0 <= b <= 1 and sum_i b_i = 1). The m exponent is 0.5 for risk parity with volatility and 1 for risk parity with variance. The standard risk parity we've been using has b = 0.5 for both equity and LTT, and m = 0.5 (that's exactly what you described).

The parameter b is the key link that allows tilting based on risk tolerance or macroeconomic indicators like the UEI. If one is a bit aggressive, setting b > 0.5 for equities emphasizes the equities but keeps the risk balancing active too. So a moderately aggressive investor might simply set b to 0.6 or 0.7.

I've seen three ways that folks have calculated variance for risk parity.
  • Symmetric (the usual variance)
  • Downward-only with upward returns included as zeros
  • Downward-only with upward returns removed entirely
These tend to increasingly reduce the average equity weight as the upward returns are increasingly removed. I've been seeing better results with the first and third approaches than the second approach for some reason.

I'm personally starting to lean towards a scheme with
  • m = 1
  • Downward-only variance with upward returns removed
  • b adjusted according to the unemployment rate index as described for the UEI
  • Rapid drop in UEI-based index on negative signal (max to min in 4 months)
  • Slow rise in UEI-based index on negative signal (min to max in 20 months)
  • Upper bound on b of around 0.75 to 0.8
  • Lower bound on b of 0 (out of market entirely)
  • Volatility look-back period of 3 months
  • Band-based rebalancing frequency of 10 to 15% change from target
  • Minimum rebalancing interval of 10 days
I'm not convinced on the slow rise yet. It may be testing well because it works better than rapid rises during extended doldrums, but it may not be best for a typical recovery.

The longer volatility period tends to help reduce whipsaw, which I suspect is due to volatility noise in short look-back periods.

The band-based rebalancing scheme with minimum rebalance interval seems to give a nice balance between responding to strong signals without rebalancing unnecessarily. A longer volatility period also helps with reducing the need to rebalance due to moving weight targets.

It seems like returns aren't strongly influenced by further tightening the equity band limits below 15 to 20% as long as the minimum rebalance interval is in place.

I've been backtesting with every 5-year period in the UPROSIM/TMFSIM dataset (i.e., starting a separate sequence each trading day).

The nominal 40/60 scheme, with quarterly rebalancing, would have had an average 5-yr CAGR of 17.4%.

The outlined scheme would have pushed average 5-yr CAGR up to 23 or 24% while reducing the hammering from Black Monday. Range of UPRO was 0 to 95%, typical peaks varied from around 60 to 80%, overall average UPRO was ~45%. Maximum drawdown was ~48%.

The market timing approach, allowing b to go to 1 (UPRO to 100%), does a bit better in most periods but would have been hammered badly on Black Monday.

In general, I'm finding that with my risk tolerance that setting an upper limit on b for UPRO around 0.75 or 0.8 (equity risk is 3/4 to 4/5 of the total risk budget) is as far as I would go with any of the schemes. This limit allows UPRO to go higher than 80% when the volatility is favorable, but not reach 100%. During backtesting with the downward-only variance and b in this range, UPRO weighting typically spiked from 60 to 80% but did spike to 95% once.

I've been developing a bunch of tests of the different weighting approaches, but that's a lot of figures to post. I'm not quite sure how to present it best for folks. I'm hoping to post a few more charts soon.

Maybe a long summary table might help.

Hope this helps.

Topic Author
Hydromod
Posts: 152
Joined: Tue Mar 26, 2019 10:21 pm

Re: Refinements to Hedgefundie's excellent approach

Post by Hydromod » Sun Aug 11, 2019 1:43 pm

Forester wrote:
Sat Aug 10, 2019 5:26 am
There was a guy on the Alpha Architect channel with a timing indicator. https://papers.ssrn.com/sol3/papers.cfm ... id=3164081

Using CBOT & CME data, hedgers (smart money) vs nonreportable speculators (dumb money) in 1) S&P futures 2) 30yr treasuries 3) 10yrVs30yr, the three relative positions combined to create one binary signal.

This could be a third modifier. The relevant data is released weekly.

From the paper,
Notably, from 2011–2017, a time period over which the U.S. equity market rose strongly— thus making it potentially detrimental not to have been fully invested in equities—the SMI strategy outpaced its nearest competitor, TREND, by approximately 75 basis points per year while having a time-averaged equity exposure of only 61% compared to 101% for TREND. Moreover, the maximum drawdown of the SMI strategy over this time period, -10.7%, was the lowest drawdown of all the strategies tested, and half the magnitude of the TREND strategy’s -21.2% drawdown.
This is very interesting! I will need to investigate this approach.

Thanks much!

unranked
Posts: 2
Joined: Sun Apr 14, 2019 1:25 am

Re: Refinements to Hedgefundie's excellent approach

Post by unranked » Mon Aug 12, 2019 6:32 am

One step ahead of me, it seems.

Indeed, one of the modifications I wanted to propose was the risk allocation/budgeting idea, where one could modify the risk budget allocated to equities based on whatever market timing indicator or personal opinion that one has. This allows one to nicely integrate both aspects of the strategy.

Is there any reason why you have opted for "variance parity" (m=1) as opposed to "volatility parity" (m=0.5), beyond superior backtesting results? From my understanding, both portfolio construction methods are mean-variance optimal under different assumptions. The former assumes equal expected returns whereas the latter assumes equal Sharpe ratios. I try to be as agnostic about expected returns as possible, so I personally favour the m=0.5 case. Would love to hear your thoughts on this.

(As an aside, we are still operating in the two-asset equity/LTT world. If we wish to extend this to multiple asset classes/strategies, we would require correlation inputs, or at the very least correlation assumptions. Do you have this goal in mind?)

Regarding the transition speed of b, I suspect the outperformance of a fast decline and slow incline that you have observed may be due to the negative skewness and volatility clustering of equity returns combined. Of course, this is entirely my guess, but I think there may be some merit in this asymmetric risking/derisking scheme. I will have to put some more thought into how I want to manipulate b.

--

Another modification that I want to propose is that of volatility targetting. The above portfolio construction method tells us in what proportion should our risks be assumed, but fails to tell us how much of that risk we should assume. As a default, the HEDGEFUNDIE thread favours 3x with a fixed dollar allocation, while your refinement has changed this to (presumably?) 3x with a variable risk allocation.

It is this value "3" that I wish to make a parameter. Consider two universes, identical in terms of the expected returns and volatilities for equities and bonds, but one where their correlation is positive and the other negative. Different joint return distributions will lead to very different portfolio risk/return profiles. An investor should be much more willing to lever up their exposure in a world where equities and bonds are negatively correlated.

Another reason why we may wish to vary our leverage is that volatility varies with time. Thus, sticking to some pre-defined leverage factor causes the amount of risk that we assume to vary with time. If volatility happens to be twice as high in 2021 than 2020, then we are assuming twice as much risk in 2021 when compared with 2020. There is no reason for this behaviour unless we have strong reason to believe that expected returns are higher in 2021, which I don't think anyone can say with any degree of confidence.

To conclude, volatility targetting suggests that instead of sticking to some pre-defined leverage factor, we should vary our leverage factor such that the levered portfolio attains some desired volatility target. This target should be determined by the individual investor. For what it's worth, I personally am fond of setting a volatility target informed by the lifecycle method (ie. higher targets when young to achieve roughly constant exposure over time and higher targets if one's human capital is negatively correlated with the financial markets, vice versa).

If you think this avenue is worth pursuing, let me know. I would be happy to help make this concrete, be it in terms of strategy specification or code.

--

Lastly, I have a somewhat frivolous thought. Any reason why you have 0 as a lower bound for b? I am of the same opinion as well, but I am afraid I have developed this restriction out of an irrational fear of shorting. Maybe we need to validate this constraint.

Once again, thank you. Looking forward to your input.

Topic Author
Hydromod
Posts: 152
Joined: Tue Mar 26, 2019 10:21 pm

Re: Refinements to Hedgefundie's excellent approach

Post by Hydromod » Mon Aug 12, 2019 8:54 am

unranked wrote:
Mon Aug 12, 2019 6:32 am
One step ahead of me, it seems.

Indeed, one of the modifications I wanted to propose was the risk allocation/budgeting idea, where one could modify the risk budget allocated to equities based on whatever market timing indicator or personal opinion that one has. This allows one to nicely integrate both aspects of the strategy.

Is there any reason why you have opted for "variance parity" (m=1) as opposed to "volatility parity" (m=0.5), beyond superior backtesting results? From my understanding, both portfolio construction methods are mean-variance optimal under different assumptions. The former assumes equal expected returns whereas the latter assumes equal Sharpe ratios. I try to be as agnostic about expected returns as possible, so I personally favour the m=0.5 case. Would love to hear your thoughts on this.

This decision is purely due to backtesting. Variance methods tend to lower average equity exposure. I didn't see that much benefit to m=1 over m=1/2 until I tested the specific downward-weighted approach, though. The downward weighting seems to do better with m=1 than m=1/2. With that said, some of the performance may depend on LTT, which is expected to have continually decreasing performance over time as discussed thoroughly in the Hedgefundie thread.

(As an aside, we are still operating in the two-asset equity/LTT world. If we wish to extend this to multiple asset classes/strategies, we would require correlation inputs, or at the very least correlation assumptions. Do you have this goal in mind?)

Not immediately. I likely will in the future, but working out the appropriate tactics with just two assets is my immediate focus. I haven't been encouraged at all by preliminary testing using portfolio visualizer.

Regarding the transition speed of b, I suspect the outperformance of a fast decline and slow incline that you have observed may be due to the negative skewness and volatility clustering of equity returns combined. Of course, this is entirely my guess, but I think there may be some merit in this asymmetric risking/derisking scheme. I will have to put some more thought into how I want to manipulate b.

My suspicion is that it worked better over for one extended period with a sideways market, I think it was the 2000s, and this dominated the better performance on faster recoveries. If so, there are two competing aspects at work here to disentangle. I haven't pinned this down for sure, though.

--

Another modification that I want to propose is that of volatility targetting. The above portfolio construction method tells us in what proportion should our risks be assumed, but fails to tell us how much of that risk we should assume. As a default, the HEDGEFUNDIE thread favours 3x with a fixed dollar allocation, while your refinement has changed this to (presumably?) 3x with a variable risk allocation.

It is this value "3" that I wish to make a parameter. Consider two universes, identical in terms of the expected returns and volatilities for equities and bonds, but one where their correlation is positive and the other negative. Different joint return distributions will lead to very different portfolio risk/return profiles. An investor should be much more willing to lever up their exposure in a world where equities and bonds are negatively correlated.

Another reason why we may wish to vary our leverage is that volatility varies with time. Thus, sticking to some pre-defined leverage factor causes the amount of risk that we assume to vary with time. If volatility happens to be twice as high in 2021 than 2020, then we are assuming twice as much risk in 2021 when compared with 2020. There is no reason for this behaviour unless we have strong reason to believe that expected returns are higher in 2021, which I don't think anyone can say with any degree of confidence.

To conclude, volatility targetting suggests that instead of sticking to some pre-defined leverage factor, we should vary our leverage factor such that the levered portfolio attains some desired volatility target. This target should be determined by the individual investor. For what it's worth, I personally am fond of setting a volatility target informed by the lifecycle method (ie. higher targets when young to achieve roughly constant exposure over time and higher targets if one's human capital is negatively correlated with the financial markets, vice versa).

If you think this avenue is worth pursuing, let me know. I would be happy to help make this concrete, be it in terms of strategy specification or code.

I haven't wanted to get into it before getting the short-term tactics worked out, but I'm beginning to think along the lines of adaptively changing the strategy components based on the position in the business cycle. Changing leverage level is one option, switching asset composition from 3x to -3x is another.

So not exactly the same idea and rationale as volatility targeting, but closely linked.

My impression is that this type of timing has historically been found to be very difficult. I've started trying to understand strategies for switching, but so far the early results have tended to agree with that conclusion. I get essentially zero correlation with predicted returns for the next month, although I may be doing a bit better for longer intervals. One might have done much better switching from positive to negative with gold; equities and LTTs, not so much.

A market index such as the UEI may be just the thing to help with switching.

There's another one that I just ran across, the business cycle index, that appears to have very similar information on timing as the UEI. It's produced weekly.

https://seekingalpha.com/article/426610 ... ay-23-2019.

I compared the two last night. It looks like the BCI may be a little more precise and perhaps actionable than the UEI. I haven't dug into it at all yet, though.


--

Lastly, I have a somewhat frivolous thought. Any reason why you have 0 as a lower bound for b? I am of the same opinion as well, but I am afraid I have developed this restriction out of an irrational fear of shorting. Maybe we need to validate this constraint.

The short answer is it backtested best over 1987 to 2019. The more nuanced answer is that I'm prepping for the idea of using a different asset composition for "in market" and "out of market", as mentioned above. In market being some flavor of the UPRO/TMF mix and out of market being some other mix TBD.


Once again, thank you. Looking forward to your input.

rascott
Posts: 485
Joined: Wed Apr 15, 2015 10:53 am

Re: Refinements to Hedgefundie's excellent approach

Post by rascott » Mon Aug 12, 2019 11:37 am

Have you done more work just looking at dual momentum....or even just a simple moving average signal?

I've been looking at both....and both get somewhat similar results. We know that momentum/trend following strategies don't beat the buy and hold market investor long-term....but they do significantly reduce big drawdown in bear markets. I'm just scratching the surface of this....but my thought is that combining leverage with momentum is the best cocktail, and should more than makeup for the underperformance of a trend-following strategy.

One could up the UPRO as they saw fit for their risk profile. Back to Jan 87....one could go all the way up to 50% UPRO/50% SP500 (2x equities)....and just go to CASH when out of market and still not see as big a drawdown as a 100% buy and holder of SP500. And it's beat the SP500 by 4.65% CAGR (I was using a 9 month SMA signal)

I wouldn't take my leverage all the way to 2x when signaled (probably just 1.4x or so)...and if you go LTTs when "out" of market then the returns shoot the moon on the backtest. But I'm lerry of counting on huge gains from LTTs looking forward.


This article has similar thoughts....but he really pushes the envelope...going all in 3x or all out. Interesting white paper is linked.

https://seekingalpha.com/article/422616 ... since-1928

Topic Author
Hydromod
Posts: 152
Joined: Tue Mar 26, 2019 10:21 pm

Re: Refinements to Hedgefundie's excellent approach

Post by Hydromod » Mon Aug 12, 2019 12:28 pm

rascott wrote:
Mon Aug 12, 2019 11:37 am
Have you done more work just looking at dual momentum....or even just a simple moving average signal?

I've tried playing with dual momentum in portfolio visualizer. There's some tantalizing things there, but I don't yet understand how it works. I've tried a little bit offline but haven't gotten far.

I've been looking at both....and both get somewhat similar results. We know that momentum/trend following strategies don't beat the buy and hold market investor long-term....but they do significantly reduce big drawdown in bear markets. I'm just scratching the surface of this....but my thought is that combining leverage with momentum is the best cocktail, and should more than makeup for the underperformance of a trend-following strategy.

One could up the UPRO as they saw fit for their risk profile. Back to Jan 87....one could go all the way up to 50% UPRO/50% SP500 (2x equities)....and just go to CASH when out of market and still not see as big a drawdown as a 100% buy and holder of SP500. And it's beat the SP500 by 4.65% CAGR (I was using a 9 month SMA signal)

I wouldn't take my leverage all the way to 2x when signaled (probably just 1.4x or so)...and if you go LTTs when "out" of market then the returns shoot the moon on the backtest. But I'm lerry of counting on huge gains from LTTs looking forward.


This article has similar thoughts....but he really pushes the envelope...going all in 3x or all out. Interesting white paper is linked.

https://seekingalpha.com/article/422616 ... since-1928

Thanks for the link. It's quite provoking. I'm not quite that brave yet.

I'm going to be looking into some of this stuff in some more detail soon. I think it may be important to have this part of the strategy resolved in the relatively near future...

I'm also a bit intrigued by what the "out of market" portfolio should consist of. The standard is cash or LTT; is there something that is consistently better than that?

As you mention LTT is looking a bit dicier as a strategy going forward. One thing to consider is that there are institutional market timing strategies that predictably switch from equity to LTT and back, so in future cycles there may be a transient bump in LTT to take advantage of without relying on the long-term returns. Cycle-scale negative correlation, so to speak.

Forester
Posts: 260
Joined: Sat Jan 19, 2019 2:50 pm
Location: UK

Re: Refinements to Hedgefundie's excellent approach

Post by Forester » Mon Aug 12, 2019 12:48 pm

rascott wrote:
Mon Aug 12, 2019 11:37 am
Have you done more work just looking at dual momentum....or even just a simple moving average signal?

I've been looking at both....and both get somewhat similar results. We know that momentum/trend following strategies don't beat the buy and hold market investor long-term....but they do significantly reduce big drawdown in bear markets. I'm just scratching the surface of this....but my thought is that combining leverage with momentum is the best cocktail, and should more than makeup for the underperformance of a trend-following strategy.

One could up the UPRO as they saw fit for their risk profile. Back to Jan 87....one could go all the way up to 50% UPRO/50% SP500 (2x equities)....and just go to CASH when out of market and still not see as big a drawdown as a 100% buy and holder of SP500. And it's beat the SP500 by 4.65% CAGR (I was using a 9 month SMA signal)

I wouldn't take my leverage all the way to 2x when signaled (probably just 1.4x or so)...and if you go LTTs when "out" of market then the returns shoot the moon on the backtest. But I'm lerry of counting on huge gains from LTTs looking forward.


This article has similar thoughts....but he really pushes the envelope...going all in 3x or all out. Interesting white paper is linked.

https://seekingalpha.com/article/422616 ... since-1928
Moving averages still inflict carnage if they trigger when unneeded. Volatility targeting tests better with leveraged ETFs. Trading UPRO on the 200day or 10month moving average would have got massacred in the 2010s.

rascott
Posts: 485
Joined: Wed Apr 15, 2015 10:53 am

Re: Refinements to Hedgefundie's excellent approach

Post by rascott » Mon Aug 12, 2019 1:05 pm

Forester wrote:
Mon Aug 12, 2019 12:48 pm
rascott wrote:
Mon Aug 12, 2019 11:37 am
Have you done more work just looking at dual momentum....or even just a simple moving average signal?

I've been looking at both....and both get somewhat similar results. We know that momentum/trend following strategies don't beat the buy and hold market investor long-term....but they do significantly reduce big drawdown in bear markets. I'm just scratching the surface of this....but my thought is that combining leverage with momentum is the best cocktail, and should more than makeup for the underperformance of a trend-following strategy.

One could up the UPRO as they saw fit for their risk profile. Back to Jan 87....one could go all the way up to 50% UPRO/50% SP500 (2x equities)....and just go to CASH when out of market and still not see as big a drawdown as a 100% buy and holder of SP500. And it's beat the SP500 by 4.65% CAGR (I was using a 9 month SMA signal)

I wouldn't take my leverage all the way to 2x when signaled (probably just 1.4x or so)...and if you go LTTs when "out" of market then the returns shoot the moon on the backtest. But I'm lerry of counting on huge gains from LTTs looking forward.


This article has similar thoughts....but he really pushes the envelope...going all in 3x or all out. Interesting white paper is linked.

https://seekingalpha.com/article/422616 ... since-1928
Moving averages still inflict carnage if they trigger when unneeded. Volatility targeting tests better with leveraged ETFs. Trading UPRO on the 200day or 10month moving average would have got massacred in the 2010s.
Where did I say one would use the UPRO MA? You wouldn't.

The signal asset would be the SP500 MA. Just as mentioned in the white paper provided in the link.

Inflicting carnage sounds a bit odd.... do you mean just underperforming holding UPRO....yeah it would have...but still would have done better than SP500 buy and hold, during 2010s.

Bear markets are when trend following sysytems shine....that's kind of my point. Target volatility may be better.
Last edited by rascott on Mon Aug 12, 2019 1:11 pm, edited 2 times in total.

Topic Author
Hydromod
Posts: 152
Joined: Tue Mar 26, 2019 10:21 pm

Re: Refinements to Hedgefundie's excellent approach

Post by Hydromod » Mon Aug 12, 2019 1:09 pm

Forester wrote:
Mon Aug 12, 2019 12:48 pm
Moving averages still inflict carnage if they trigger when unneeded. Volatility targeting tests better with leveraged ETFs. Trading UPRO on the 200day or 10month moving average would have got massacred in the 2010s.
Maybe the key is to use the underlying S&P index rather than UPRO for the moving average? That seems to be what the linked white paper is suggesting.

I don't think I would have thought of that twist without your comment. Now I'm wondering whether these signals should be based on the actual leveraged asset or on the unleveraged index. I've always just assumed that the leveraged asset should give the signal. But leveraged volatility is indeed more likely to trigger whipsaws.

edit: too slow to type...

But also I'd go to the idea of incremental switches regardless, like with the UEI approach I mentioned. That would slow down whipsaws too.

Lee_WSP
Posts: 621
Joined: Fri Apr 19, 2019 5:15 pm
Location: Arizona

Re: Refinements to Hedgefundie's excellent approach

Post by Lee_WSP » Mon Aug 12, 2019 1:23 pm

Forester wrote:
Mon Aug 12, 2019 12:48 pm
rascott wrote:
Mon Aug 12, 2019 11:37 am
Have you done more work just looking at dual momentum....or even just a simple moving average signal?

I've been looking at both....and both get somewhat similar results. We know that momentum/trend following strategies don't beat the buy and hold market investor long-term....but they do significantly reduce big drawdown in bear markets. I'm just scratching the surface of this....but my thought is that combining leverage with momentum is the best cocktail, and should more than makeup for the underperformance of a trend-following strategy.

One could up the UPRO as they saw fit for their risk profile. Back to Jan 87....one could go all the way up to 50% UPRO/50% SP500 (2x equities)....and just go to CASH when out of market and still not see as big a drawdown as a 100% buy and holder of SP500. And it's beat the SP500 by 4.65% CAGR (I was using a 9 month SMA signal)

I wouldn't take my leverage all the way to 2x when signaled (probably just 1.4x or so)...and if you go LTTs when "out" of market then the returns shoot the moon on the backtest. But I'm lerry of counting on huge gains from LTTs looking forward.


This article has similar thoughts....but he really pushes the envelope...going all in 3x or all out. Interesting white paper is linked.

https://seekingalpha.com/article/422616 ... since-1928
Moving averages still inflict carnage if they trigger when unneeded. Volatility targeting tests better with leveraged ETFs. Trading UPRO on the 200day or 10month moving average would have got massacred in the 2010s.
Yes, exactly. Even 36 month MMA would have triggered in December 2018 and also the 1987 flash crash. And the whipsaw is worse with a leveraged asset.

Forester
Posts: 260
Joined: Sat Jan 19, 2019 2:50 pm
Location: UK

Re: Refinements to Hedgefundie's excellent approach

Post by Forester » Mon Aug 12, 2019 1:24 pm

Hydromod wrote:
Mon Aug 12, 2019 1:09 pm
Forester wrote:
Mon Aug 12, 2019 12:48 pm
Moving averages still inflict carnage if they trigger when unneeded. Volatility targeting tests better with leveraged ETFs. Trading UPRO on the 200day or 10month moving average would have got massacred in the 2010s.
Maybe the key is to use the underlying S&P index rather than UPRO for the moving average? That seems to be what the linked white paper is suggesting.

I don't think I would have thought of that twist without your comment. Now I'm wondering whether these signals should be based on the actual leveraged asset or on the unleveraged index. I've always just assumed that the leveraged asset should give the signal. But leveraged volatility is indeed more likely to trigger whipsaws.

edit: too slow to type...

But also I'd go to the idea of incremental switches regardless, like with the UEI approach I mentioned. That would slow down whipsaws too.
As a hack USMV as a signal could be used for the time series portion https://www.portfoliovisualizer.com/tes ... 0&total1=0

The drawdown looks worse in isolation without the bonds. Success for UPRO in the 2010s should be moreorless replicating the buy&hold performance because there's been no cause not to be invested.

perplexed
Posts: 53
Joined: Thu Feb 02, 2012 4:32 pm

Re: Refinements to Hedgefundie's excellent approach

Post by perplexed » Mon Aug 12, 2019 1:27 pm

Hydromod: shamelessly (and greedy) checking if your wonderful scripts are coming!

Thanks!

Topic Author
Hydromod
Posts: 152
Joined: Tue Mar 26, 2019 10:21 pm

Re: Refinements to Hedgefundie's excellent approach

Post by Hydromod » Mon Aug 12, 2019 1:40 pm

perplexed wrote:
Mon Aug 12, 2019 1:27 pm
Hydromod: shamelessly (and greedy) checking if your wonderful scripts are coming!

Thanks!
Getting there. I spent much of the weekend getting them in order. The basic guts are just about there, now I'm setting up example tests with plots. Hopefully in a few days...

This stuff can grow rapidly...

Topic Author
Hydromod
Posts: 152
Joined: Tue Mar 26, 2019 10:21 pm

Re: Refinements to Hedgefundie's excellent approach

Post by Hydromod » Mon Aug 12, 2019 11:13 pm

I took a couple of hours and implemented the moving average in the same risk-budget style as the UEI example I showed. I ramp the budget allocation up a fixed delta each day that the daily value is above the moving average, and ramp down a fixed delta each day that the daily value is below the moving average. I can allow this to occur like an on/off switch or gradually over months. I can use either UPRO or UPRO/3 as the signal.

Basically the different options I tried give fairly consistent average 5-yr CAGR in the range of 20 to 23% and max drawdowns in the range of 45 to 66%, compared to the nominal 40/60 with 17.4% and max drawdown of 66%. This is roughly in the range of the UEI options I checked. I think they are giving fairly similar signals regarding timing. Again limiting the maximum risk budget to 80% helped mitigate Black Monday drawdowns.

There may not be much more that can be easily gained with just these two fixed assets.

ocrtech
Posts: 23
Joined: Sat Jul 21, 2012 2:18 pm

Re: Refinements to Hedgefundie's excellent approach

Post by ocrtech » Sat Aug 17, 2019 5:13 pm

Have you done any work on using alternate approaches to calculating volatility?

From the papers I've read, the traditional approach of annualizing the standard deviations of logarithmic returns only preserves around 30% of the information around volatility clusters. I've looked at a couple GARCH variants as well as EWMA but even they only encapsulate around 60%+ of the information. I'm curious if you have done any experimenting with these or other approaches and what kind of improvements you might have seen.

no simpler
Posts: 71
Joined: Sat Jul 13, 2019 4:54 pm

Re: Refinements to Hedgefundie's excellent approach

Post by no simpler » Sun Aug 18, 2019 9:04 pm

ocrtech wrote:
Sat Aug 17, 2019 5:13 pm
Have you done any work on using alternate approaches to calculating volatility?

From the papers I've read, the traditional approach of annualizing the standard deviations of logarithmic returns only preserves around 30% of the information around volatility clusters. I've looked at a couple GARCH variants as well as EWMA but even they only encapsulate around 60%+ of the information. I'm curious if you have done any experimenting with these or other approaches and what kind of improvements you might have seen.
Fractionally differentiated series to preserve memory.

HawkeyePierce
Posts: 529
Joined: Tue Mar 05, 2019 10:29 pm
Location: Colorado

Re: Refinements to Hedgefundie's excellent approach

Post by HawkeyePierce » Sun Aug 18, 2019 9:27 pm

While I'm not adopting Hydromod's strategies myself, I gotta say building an application to automatically make these trades according to target vol or UEI using Interactive Brokers' API sounds like a fun weekend project.

no simpler
Posts: 71
Joined: Sat Jul 13, 2019 4:54 pm

Re: Refinements to Hedgefundie's excellent approach

Post by no simpler » Sun Aug 18, 2019 9:42 pm

HawkeyePierce wrote:
Sun Aug 18, 2019 9:27 pm
While I'm not adopting Hydromod's strategies myself, I gotta say building an application to automatically make these trades according to target vol or UEI using Interactive Brokers' API sounds like a fun weekend project.
I'm working on this, but it's eating up a lot more than one weekend! You can go down a real rabbit hole predicting volatility. Simply using trailing realized volatility isn't very effective as the previous poster mentioned.

HawkeyePierce
Posts: 529
Joined: Tue Mar 05, 2019 10:29 pm
Location: Colorado

Re: Refinements to Hedgefundie's excellent approach

Post by HawkeyePierce » Sun Aug 18, 2019 9:50 pm

no simpler wrote:
Sun Aug 18, 2019 9:42 pm
HawkeyePierce wrote:
Sun Aug 18, 2019 9:27 pm
While I'm not adopting Hydromod's strategies myself, I gotta say building an application to automatically make these trades according to target vol or UEI using Interactive Brokers' API sounds like a fun weekend project.
I'm working on this, but it's eating up a lot more than one weekend! You can go down a real rabbit hole predicting volatility. Simply using trailing realized volatility isn't very effective as the previous poster mentioned.
If you plan on open-sourcing your code I might be interested in pitching in.

Forester
Posts: 260
Joined: Sat Jan 19, 2019 2:50 pm
Location: UK

Re: Refinements to Hedgefundie's excellent approach

Post by Forester » Mon Aug 19, 2019 3:09 am

Could the graduated unemployment indicator be implemented using moving averages of different lengths ?

no simpler
Posts: 71
Joined: Sat Jul 13, 2019 4:54 pm

Re: Refinements to Hedgefundie's excellent approach

Post by no simpler » Mon Aug 19, 2019 6:04 pm

HawkeyePierce wrote:
Sun Aug 18, 2019 9:50 pm
no simpler wrote:
Sun Aug 18, 2019 9:42 pm
HawkeyePierce wrote:
Sun Aug 18, 2019 9:27 pm
While I'm not adopting Hydromod's strategies myself, I gotta say building an application to automatically make these trades according to target vol or UEI using Interactive Brokers' API sounds like a fun weekend project.
I'm working on this, but it's eating up a lot more than one weekend! You can go down a real rabbit hole predicting volatility. Simply using trailing realized volatility isn't very effective as the previous poster mentioned.
If you plan on open-sourcing your code I might be interested in pitching in.
Right now I have a semi-automated system. I have a cron job that pulls latest market data from an API, crunches the data, computes some stats, then posts to Slack or email telling me what to do. But I haven't automated trades yet. Happy to share once I've cleaned up my code.

no simpler
Posts: 71
Joined: Sat Jul 13, 2019 4:54 pm

Re: Refinements to Hedgefundie's excellent approach

Post by no simpler » Mon Aug 19, 2019 6:09 pm

ocrtech wrote:
Sat Aug 17, 2019 5:13 pm
Have you done any work on using alternate approaches to calculating volatility?

From the papers I've read, the traditional approach of annualizing the standard deviations of logarithmic returns only preserves around 30% of the information around volatility clusters. I've looked at a couple GARCH variants as well as EWMA but even they only encapsulate around 60%+ of the information. I'm curious if you have done any experimenting with these or other approaches and what kind of improvements you might have seen.
Best signal I've found so far is in fact VIX after all. Nice thing about VIX is you get around all the memory-destroying issues of differenced returns by just using the VIX itself since it's reasonably stationary as is.You can get VIX to be completely stationary using fractional differentiation of order d=.1 (with a decimal point).

perplexed
Posts: 53
Joined: Thu Feb 02, 2012 4:32 pm

Re: Refinements to Hedgefundie's excellent approach

Post by perplexed » Mon Aug 19, 2019 6:58 pm

no simpler wrote:
Mon Aug 19, 2019 6:09 pm
ocrtech wrote:
Sat Aug 17, 2019 5:13 pm
Have you done any work on using alternate approaches to calculating volatility?

From the papers I've read, the traditional approach of annualizing the standard deviations of logarithmic returns only preserves around 30% of the information around volatility clusters. I've looked at a couple GARCH variants as well as EWMA but even they only encapsulate around 60%+ of the information. I'm curious if you have done any experimenting with these or other approaches and what kind of improvements you might have seen.
Best signal I've found so far is in fact VIX after all. Nice thing about VIX is you get around all the memory-destroying issues of differenced returns by just using the VIX itself since it's reasonably stationary as is.You can get VIX to be completely stationary using fractional differentiation of order d=.1 (with a decimal point).
Would you please elaborate on specific values on VIX for buy/sell?
Any back testing?

Many thanks.

ocrtech
Posts: 23
Joined: Sat Jul 21, 2012 2:18 pm

Re: Refinements to Hedgefundie's excellent approach

Post by ocrtech » Mon Aug 19, 2019 7:18 pm

no simpler wrote:
Sun Aug 18, 2019 9:04 pm
ocrtech wrote:
Sat Aug 17, 2019 5:13 pm
Have you done any work on using alternate approaches to calculating volatility?

From the papers I've read, the traditional approach of annualizing the standard deviations of logarithmic returns only preserves around 30% of the information around volatility clusters. I've looked at a couple GARCH variants as well as EWMA but even they only encapsulate around 60%+ of the information. I'm curious if you have done any experimenting with these or other approaches and what kind of improvements you might have seen.
Fractionally differentiated series to preserve memory.
Thanks! This recommendation led me to HAR-RV as well as multifractal volatility models. I'm not sure I want to continuously collect a whole bunch of intraday trading data in order to predict out-of-sample volatility numbers 20 days out but this definitely helps me put bounds around the potential of the overall trading approach.

Post Reply