What the Trend

This post is available as a PDF download here.

Summary

• In this research note, we explore the performance of simple trend equity strategies and funds in the recent market rout.
• We find a significant dispersion in realized performance, with some strategies shifting entirely to cash at the end of February and some remaining entirely invested.
• We explain why we would expect different strategies to behave differently based upon their model specification.
• We also highlight the roll that luck can play in highly volatile environments. Even something as simple as “when you execute your trade” can lead to hundreds of basis points of performance dispersion when markets are trading in 10% intrarday ranges.
• We again argue that specification luck, rebalance timing luck, and even execution luck can all be mitigated through the thoughtful application of diversification.

When compared to the –S&P 500–, it would appear many trend equity strategies have done little to curtail the recent market rout.

Source: Tiingo.  Calculations by Newfound Research.  Returns are hypothetical and backtested.  Returns are gross of all management fees, transaction fees, and taxes, but net of underlying fund fees.  Total return series assumes the reinvestment of all distributions.  Data through 3/17/2020.

But under the hood, dispersion abounds!  Some monthly-rebalanced strategies started the month heavily hedged, whereas others, using just slightly different signals, came in fully invested.  Some strategies have been slowly transitioning into defensive assets throughout the month whereas others have made large, sudden changes.

I have said it before and I will say it again: this is what we should expect when we evaluate a number of different strategies that are, in large part, based on one or two simple signals.  What we’re witnessing here is model specification risk and rebalance timing luck all rolled into one.

Even if we hold the rebalancing piece constant, small model differences have led to large dispersion.  For example, if you followed Meb Faber’s popular timing model – selling when price falls below its 10-month moving average – you entered the month in cash.  On the other hand, if you followed Gary Antonacci’s popular Dual Momentum GEM model – which is based upon prior 12-month total returns – you would have entered the month in equities.

That is by no means an indictment of the Dual Momentum GEM model.  Shift the sell-off by a week and the 10-month moving average would have missed the call as well (tell me that’s not rebalance timing luck).  Rather, it highlights the specification and timing risk that can occur if we are under-diversified in our how and when axes.

Back to Basics

The first question we should ask ourselves is, “what should we expect from trend following in an environment like this?”

If we were pursuing guaranteed protection, we’d prefer something like a put option.  But that contractual nature can make put options prohibitively expensive, especially if we have to roll our protection during periods of high market volatility.

Trend following tries to reduce the cost of protection, but in doing so eliminates the guarantee.

One might think of it this way: a put option is like buying fire insurance on your house (with the added risk that you might have to re-up your insurance while the house is on fire) and trend following is more like an alarm and sprinkler system.  With the former we pay premiums year-in and year-out; with the latter we get false alarms and the risk that we cannot respond fast enough when a fire breaks out.

There is actually a theoretical link between options and trend following that we’ve discussed in the past.  Specifically, a long/short trend following strategy coarsely approximates the delta-hedging strategy of an option strangle, which ultimately replicates the payoff of a straddle.

If that sounds like gibberish to you, just know that it basically means that trend following approximates the return profile of buying both a put and call option through its trading strategy.

With that link in mind, we can ask, “how can we make the false alarm cost of trend following less?”  The same way it is done in the options world: add more bells and whistles.  Want to buy protection but the premium is too expensive?  Add some conditions like knock-out or knock-in barriers.  In the context of our fire insurance example, an example would be, “protect me against fires unless they happen on every 3^rd Thursday of the month.”

By adding more conditions, the cost of the protection goes down.  Of course, it does so in a manner commensurate with the protection we now expect.

Similarly, in the world of trend following, we might see bells and whistles added on a strategy; e.g. “if price falls below the 200-day moving average and stays there for 5 days,” or “sell if price falls below 10-month moving average and 12-month total return is negative.”  Increased conditions will reduce the probability of triggers, which will likely lower whipsaw, but also increases the regimes in which the strategy will not provide protection.

Gotta Go Fast?

For all the performance dispersion, most trend strategies really are not very different at all.  Generally speaking, their specification is based upon a model that identifies the trend signal (e.g. prior total returns, price-minus-moving-average, and moving-average-dual-crossover) and speed (i.e. the lookback period).

Most common models, however, are mathematically linked.  When broken down to basic building blocks, prior total returns, price-minus-moving-average, and moving-average-dual-crossover all simply apply a different weighting scheme to prior price changes.  Whereas prior total returns (“TSMOM”) equally-weights past price changes, price-minus-moving-average (“PMAC”) front-weights and dual-moving-average-crossover (“DMAC”) back-weights.

In the graphic above we can see that for the same model, a shorter lookback prior – i.e. “faster speed” – results in increased forward weighting.  For example, we can see that the PMAC(50) model puts substantially more weight on recent price changes than the PMAC(200) model, and therefore will react more quickly when trends change.

Thus, it is the combination of the model and speed that will largely dictate how well a trend-following strategy performs for a given drawdown.  If we see a 20% sell-off occurring over two weeks, for example, we should not be surprised if slow-moving models do not perform particularly well.

We explored this idea in our note The Speed Limit of Trend. Using a Brownian Bridge model, we simulated hypothetical market drawdowns of a given depth over a given duration.  We can see that for a 20% drawdown occurring over a short time period (e.g. 1 week to 3 months), we would expect faster models to perform substantially better than slower models.

(Note that the table below is for long/short trend following systems; as we’ve discussed many times in the past, a long/flat system is equivalent to a 0.5 beta portfolio plus 50% exposure to a long/short strategy.)

It should come as no surprise, then, given the speed and depth of the drawdown over the last several weeks that we would expect fast models to out-perform slow models.  And that is precisely what we see when we run these long/short strategies on the actual path taken by the S&P 500 year-to-date.  Many of the long-term models are still invested.

Source: Tiingo.  Calculations by Newfound Research.  Returns are hypothetical and backtested.  Returns are gross of all management fees, transaction fees, and taxes, but net of underlying fund fees.  Total return series assumes the reinvestment of all distributions.  Data through 3/18/2020.

Why even bother with long-term models then, if they’re not going to provide the here and now protection we want?  We need only walk back a few years to see what happens if we try to go too fast.  And these figures do not even account for trading costs!

Source: Tiingo.  Calculations by Newfound Research.  Returns are hypothetical and backtested.  Returns are gross of all management fees, transaction fees, and taxes, but net of underlying fund fees.  Total return series assumes the reinvestment of all distributions.  Data through 3/18/2020.

We can connect this empirical evidence back to options again by thinking of trend following as the payoff of the underlying option strategy plus the trading impact required to replicate it.  We should recognize that trading impact will be zero if, and only if, the trend strategy can continuously trade, without cost, to always hold the exact delta of the option strategy.  In reality, these trend strategies trade daily and do so with binary (+1 or -1) positioning.  These discrepancies can lead the strategy to under- and over-shoot its delta target, creating substantial replication whipsaws.  For longer-dated options, this may be less of an issue as the delta changes more slowly.

Nevertheless, the speed of the system ultimately has very important implications for how we should measure its success.  If fast systems work better for sudden drawdowns and slower systems work better for more prolonged drawdowns, then those are the horizons over which we would expect to see the desired convexity properties emerge over.

Indeed, this is precisely what we found in G̷̖̱̓́̀litch.  Fast systems exhibited measurable convexity over 1-month time frames, whereas slow systems exhibited convexity over one-year time frames.  In other words: if we expected a slow system to work over the last month, we had the wrong expectations.

As we begin to connect all these dots together, we generally find that:

• Faster systems should be expected to perform better in faster sell-offs than slower systems.
• Faster systems may not necessarily do well during prolonged market declines.
• Over the long run, faster systems have not fared as well as slower systems; one argument is that they may compound whipsaw costs more frequently.
• Increasing the number of bells-and-whistles on a trend strategy likely reduces whipsaw costs but only because it reduces the environments it will protect in (“risk is not destroyed, only transformed.”)

Market Volatility Puts Timing Risk on Steroids

In the introduction, we mentioned the discrepancy in signals seen at the end of February using Meb Faber’s 10-month moving average model and Gary Antonacci’s 12-month return model.  That discrepancy would lead one person to go to cash while the other would stay fully invested.  When we talk about model specification risk, this is precisely what we mean.

Had the entire market drawdown shifted by a week, however, the 10-month signal would not have triggered either and both strategies would have remained invested.  That’s precisely what we mean when we talk about rebalance timing luck.

But in highly volatile markets, rebalance timing luck can strike in more subtle ways: execution price.

Consider the graph below, which plots the absolute overnight gap (i.e. the absolute value of prior close minus open, divided by prior close) and absolute intraday range (high minus low divided by current close) for the S&P 500.  We can see that over the last two weeks, overnight and intraday ranges have started to exceed 5%.

Source: Tiingo.  Calculations by Newfound Research.  Absolute Intraday Range is |High(t) – Low(t)| / Close(t).  Absolute Overnight Gap is |Open(t) – Close(t-1)|/Close(t-1).

Consider an investment strategy that generates a sell signal after market close on Friday, March 6^th and intends to trade that signal during the day on Monday, March 9^th.  By the time this strategy went to trade, the market had moved 7.5% away from the price it had generated its signal on!  In this example, the market opened lower and went against a trend follower, but given the variable lead/lag relationship between trend and mean-reversion there is nothing to say that this couldn’t have been a positive open and a boon to performance.

Things get worse from here when we then consider the intraday breadth!  Consider March 13^th: a day where the intraday range for the S&P 500 spanned between 2485 and 2714 before closing at 2693.  If two managers ran an identical strategy that triggered a sell order, but one executed near the day’s low and one executed near the day’s high, the captured spread in performance by day’s end would be north of 8.5% between them!

Here is a more concrete example.  If you ran a simple trend following model that said to sell out of the market when the prior 252-day (approximately 1-year) return turned negative, you received a sell signal after market close on March 12^th.  Below we plot the performance of such a strategy depending upon the execution price realized on March 13^th.

Source: Tiingo.  Calculations by Newfound Research.  Returns are hypothetical and backtested.  Returns are gross of all management fees, transaction fees, and taxes, but net of underlying fund fees.  Total return series assumes the reinvestment of all distributions.

The spread in realized performance exceeds 600 basis points simply based upon when you sold intraday.

(As a side note, the TWAP model above is not actual TWAP, but a simple estimate thereof based upon open, high, low, and close information.  You can read more about our implementation here.)

In incredibly volatile markets, not only do we have to concern ourselves with what happens between rebalances, but also between the time we choose to rebalance and when we can actually execute.  Recent days have demonstrated that the later point can account for hundreds, if not thousands, of basis points.

It will come as no shock to anyone who reads our commentaries that we propose the same solution whenever discussing timing luck: rebalance a little, but frequently.  Whenever possible, we try to avoid all-in and all-out calls at discrete time-steps and prefer to continuously transition our portfolios.  This same logic applies to our trading and testing as well, where we generally target VWAP execution in our funds and use the estimated TWAP on the day following a rebalance in our calculated indices.

Conclusion

One of the hardest problems that allocators face is disentangling luck from skill.  Investment performance is often used as a proxy, but herein we hope to have demonstrated that thousands of basis points of relative performance difference can be due to numerous lucky events, namely:

• The luck of which model you use.
• The luck of when you rebalance (e.g. monthly; weekly; continuously).
• The luck of the execution achieved.

Of course, at Newfound we would argue that you can largely avoid these sources of luck through diversification.  Don’t use only one model: use several.  Don’t rebalance in discrete time steps (e.g. monthly): rebalance partially but more continuously.  And if you can avoid it, try not to execute all at once during high volatility days: execute throughout the day.

All of these activities can help curb the impact of bad luck.  But they will also curb the impact of good luck.  It is entirely possible that you could pick the right model, rebalance at just the right time, and get the best execution of the day.  We’d argue, though, that achieving consistency of outcomes requires working to reduce the impact of randomness.

Corey Hoffstein

Corey is co-founder and Chief Investment Officer of Newfound Research, a quantitative asset manager offering a suite of separately managed accounts and mutual funds. At Newfound, Corey is responsible for portfolio management, investment research, strategy development, and communication of the firm's views to clients. Prior to offering asset management services, Newfound licensed research from the quantitative investment models developed by Corey. At peak, this research helped steer the tactical allocation decisions for upwards of $10bn. Corey holds a Master of Science in Computational Finance from Carnegie Mellon University and a Bachelor of Science in Computer Science, cum laude, from Cornell University. You can connect with Corey on LinkedIn or Twitter.