2020 Q1 Investor Letter

­Clients & Friends –

The first two thirds of the quarter were uneventful.  I will spend the remainder of this commentary discussing the other 99% (or, what felt like 99% of the quarter).

Rather than re-hash information that is more publicly available (e.g. monetary and fiscal policy, economic conditions, market fragility, et cetera), I want to reflect upon some lessons learned.  While I suspect that we will continue to learn from this event over time, I believe an early post mortem can provide some insight into what worked, what didn’t, and implications going forward.

But first, a quick firm update.

Disclosure: Any performance discussed herein is hypothetical model performance does not reflect the results achieved by any investor.  Performance is gross of all fees, expenses, sales charges, or trading expenses, but net of underlying ETF expense ratios (where applicable).  Returns assume the reinvestment of all distributions.  GIPs compliant returns available upon request.

Firm Update

We are excited to announce the addition of Dillon Pierce to our team as our VP of Advisor Solutions, providing further education   Prior to joining Newfound, Dillon worked at 361 Capital and NorthStar Securities, where he specialized in providing education on managed futures, long/short equity, nontraded REITs, and closed-end funds.  Dillon is based out of our Denver, CO office.  Dillon can be reached at dpierce@thinknewfound.com.

Steven Braun, who joined our investment team last year, has recently revamped our quarterly updates to provide more thorough performance decompositions.  If you are interested in receiving these updates, please reach out.

What Worked

Broadly speaking, trend following helped cut portfolio risk in equity mandates.  Coming off a strong 2019 and all-time highs, however, meant that most trend models did not begin to significantly react until early March.  Even then, your mileage varied significantly given your specific trend model implementation.

As a broad benchmark for these types of strategies, we publish a generic US Trend Equity index.  This index adopts a diversified approach, mixing a variety of specifications (i.e. models and speeds) to arrive at its final allocation blend between –US Equities– and –Short-Term US Treasuries–.  As one might suspect, and as seen in the allocation graph below, the transition from –US Equities– to –Short-Term US Treasuries– occurred gradually through March as trend signals of different speeds turned off at different times.

Portfolio allocations above are model holdings and therefore are hypothetical and do not represent actual client portfolios.

The ensemble nature of the model means that rather than cutting risk outright, the strategy tends to cut risk a little but frequently, seeking to bring the portfolio to a soft landing during turbulent periods.  This can work well when drawdowns are more prolonged (e.g. 2007 and 2008) or the market makes a rapid V-bottom prior to all the signals turning off (e.g. 2018).  It can even potentially benefit from short-term mean reversion that is often seen after trend signals turn off, selling into strength as the market rallies.

March proved to be none of these cases.  The sell-off was rapid and relentless and by the time markets hit bottom – one short month after hitting all-time highs – the US Trend Equity Index had only cut approximately 50% of its exposure.  So, while it did prevent some drawdown, and is quite well prepared for any further drawdown, it did not cut out a meaningful amount of risk.

However, as we mentioned above, mileage varied dramatically among different trend equity implementations and we were pleased that our own proprietary trend models fared better.  Below we plot the year-to-date equity curves of the –S&P 500–, the –Newfound U.S. Trend Equity Index–, our Risk Managed U.S. Sectors (“RMUS”) model (with –Weekly– and –Daily– rebalancing mandates), and our –RMUS + Portable Beta– model (which is overlaid with our U.S. Treasury futures portable beta mandate).

Source: Tiingo. Calculations by Newfound Research. Past performance does not guarantee future results. Performance is hypothetical and gross of all fees, expenses, sales charges, or trading expenses with the exception of underlying ETF expense ratios.  Returns assume the reinvestment of dividends.

We can see that on 3/23, when the –S&P 500– was down more than 30% on the year, the –RMUS (Weekly)– model was down 22.2%, performing better than the –Newfound U.S. Trend Equity Index– which was down -26.2%.  The –RMUS + Portable Beta– model was down just 18.5%, benefiting from the overlay of U.S. Treasury exposure.

A meaningful driver of this performance difference was the dynamic nature of the trend models we employ.  We have written a bit about our approach in the past, but the philosophy can largely be summed up by Lenin’s quote, “There are decades where nothing happens; and there are weeks where decades happen.”

In moving from philosophy to practical implementation, our aim is to recognize that markets move on information flow, and that information flow is not constant in time, either in magnitude or velocity.  Consider that a simple trend model – e.g. a 200-day simple moving average – always uses the same information window.  One might argue that events of the past 30 days more-or-less make the prior 170 irrelevant.

Rather than trying to measure trends over a constant amount of time, our aim is to measure them over a constant amount of information flow.  Quantifying information flow, however, is not trivial.  But some domains – for example price ticks, volume, or volatility – may serve as suitable proxies.  This translation is precisely what we do.

With some care, a trend model built in a different domain can be translated back into the time domain.  In doing so, we end up with a model that is dynamic: sometimes reflecting slow trend models and sometimes reflecting fast trend models.

Prior to March, most of our models reflected slower moving averages, which, in our experience, is typical for slow and steady bull markets.  During March we have seen some of the fastest transitions in model speeds (i.e. slow to fast) as well as some of the fastest speed level readings ever (exceeding even those we saw in 2008).

Source: Tiingo.  Calculations by Newfound Research.

Above we plot the model speed dynamics for our trend models when applied to large-cap US equity sector ETFs.  Entering 2020, the model speeds ranged between 150 and 250 days, reflecting intermediate-to-slow reaction speeds.  By early March, those speeds had accelerated to between 100 and 200 days.  By end of quarter, they had plunged to as low as 20 days, with the median sitting closer to 50.

As a further example, below we plot the price of a –U.S. Financials Sector ETF (XLF)–, its ­–200-Day Moving Average–, and –Newfound’s Trend Model–.  We can see that prior to March, Newfound’s model acted almost identically to the 200-Day model, but sped up dramatically during the sell-off, reflecting the information cascade that occurred.

Source: Tiingo.  Calculations by Newfound Research.

This rapid acceleration in model speed meant that by February 28^th, eight of the ten sector signals we evaluate had turned off.  The remaining two (Technology and Utilities) turned off by March 9^th.  By comparison, only 38% of the signals underlying the US Trend Equity Index had turned off by February 28^th, and just 66% were off by March 9^th.

This rapid acceleration should also serve us well if markets make a rapid rebound.  In one of our recent commentaries, we discussed expected performance of different trend following speeds given different recovery shapes (i.e. V, U, W, or L-style recoveries), drawdown depths, and drawdown durations.  We found that for short-term recoveries, quicker models were well suited.

As a reminder, we implement our tactical changes over time in an effort to avoid rebalance timing luck.  For these models specifically, equity signals are transitioned over a one-month period.  We discuss the implications of this choice in the What Didn’t Work section.  Spoiler: this transitional approach was not effective in March.

Finally, mandates implementing our tactical U.S. Treasury futures overlay (e.g. our mutual funds) benefited from the market’s rapid flight to safety, with 10-year yields falling from 1.88% at the beginning of the year and going as low as 0.38% intraday. Unfortunately, entering the year we carried a significantly reduced notional position.

The futures strategy is driven in equal parts by three primary categories of signals: value, momentum, and carry.  Our value signals, which are based upon real yield estimates, turned bearish in Q1 2019 and we have been systematically reducing our position ever since.  Our carry signals, which capture both term spread and roll yield, looked thin on both an absolute and historical basis.  Only momentum signals remained largely positive.  The cumulative effect (shown below) was a half-sized position.

Portfolio allocations above are model holdings and therefore are hypothetical and do not represent actual client portfolios. There is no guarantee that the strategy will achieve its objectives. There are no guarantees that the strategy will be positioned correctly for any given market environment. The strategy utilizes various rebalance techniques designed to reducing transaction costs and turnover, which may result in the strategy’s actual allocation straying from its target allocation.

Nonetheless, even at reduced size, the position provided meaningful returns over the quarter and helped buffer some of the equity losses.  We estimate that the model returned approximately 3.9%.  For our mutual fund mandates, which implement just a 60% notional overlay (e.g. a 50% position is equivalent to 30% notional exposure), this translates into an approximate 2.3% benefit.

Source: Stevens Futures. Calculations by Newfound Research. Past performance does not guarantee future results. Performance is hypothetical and gross of all fees, expenses, sales charges, or trading expenses.

In real time, however, the drawback to this overlay was felt from 3/9 to 3/18, when it gave up -1.9% while the S&P 500 also fell approximately -12.5%.  Rather than providing a buffer, the overlay exacerbated losses.  Whether this retracement reflected a dysfunctional fixed income market, forced de-leveraging, or just a mad dash for cash is unknown.  But it is by no means unprecedented.  From October 6^th through October 10^th, 2008, the S&P 500 fell over -15%.  At the same time, 7-10 Year U.S. Treasuries fell approximately -3.5%.  While we believe that the flight-to-quality trade captured by U.S. Treasuries is alive and well, it is worth recognizing that when markets completely break down all correlations do go to one.

What Didn’t Work

In our research on rebalance timing luck, we have identified three key variables: (1) the frequency of rebalancing; (2) the intrinsic turnover of the strategy; and (3) the volatility of a long/short portfolio that captures what the strategy is invested in versus what it could be invested in.

For a simple long/flat trend equity strategy (i.e. one that allocates between equities and short-term U.S. Treasuries), that last piece can be proxied by the volatility of equities.  When the volatility of broad equities increases, rebalance timing luck goes up.

When the volatility of equities goes from placid to levels not seen since the Great Financial Crisis in a two-week period, you can imagine that the implications of rebalance timing luck explodes.

We have long argued that the best solution for dealing with rebalance timing luck is to implement a tranched rebalancing schedule.  Simply put, the idea is to divide your capital among a number of sub-portfolios and rebalance the sub-portfolios in a rotation.  For example, rather than rebalancing once a month, we could rebalance ½ of our portfolio every-other week.  Or we could rebalance ¼ of our portfolio each week.  As the number of sub-portfolios goes up, the impact of rebalance timing luck goes down (for N sub-portfolios, timing luck is reduced to 1/Nth of its original magnitude).

Below we plot the returns for our RMUS model implemented with different rebalance strategies.  The –Monthly– strategy rebalances only at the end of each month.  The –Bi-Weekly– strategies rebalance one of their two sub-portfolios at the end of every other week.  The –Weekly– rebalances one of its four sub-portfolios each week.  Finally, we also plot a –Daily– variation, which rebalances one of its twenty sub-portfolios each day (which is likely only feasible when implemented inside a pooled vehicle, as we do in our mutual funds).

 Source: Tiingo. Calculations by Newfound Research. Past performance does not guarantee future results. Performance is hypothetical and gross of all fees, expenses, sales charges, or trading expenses with the exception of underlying ETF expense ratios.  Returns assume the reinvestment of dividends.

The dispersion in performance is fairly significant.  The –Monthly– strategy performs best, with a year-to-date return of just -7.81% and the –Weekly– strategy performs the worst, with a year-to-date return of -20.82%.  But we can see that the –Bi-Weekly–, –Weekly–, and –Daily– results are much more similar to each other, potentially indicating that the –Monthly– strategy was merely a benefactor of rebalance timing luck.

Note, however, that controlling for timing luck does not guarantee the best outcome.  In fact, the best outcome in this case came from being a benefactor of timing luck.  Just as diversification across assets means you give up the best to avoid the worst, diversification over time means you give up the good luck to avoid the bad.

Consider, for example, this alternative scenario: instead of the market selling off in the last week of February, everything was shifted later by 5 days.  Now, if we re-run the same scenario, we find that the Monthly strategy is the worst performer, having remained invested throughout the whole period.

(Note that while the dates below are listed through 3/31/2020, the data has been shifted forward by five trading days, so the actual last five trading days of March are not reflected).

Source: Tiingo. Calculations by Newfound Research. Past performance does not guarantee future results. Performance is hypothetical and gross of all fees, expenses, sales charges, or trading expenses with the exception of underlying ETF expense ratios.  Returns assume the reinvestment of dividends.

In a highly volatile market, significant performance dispersion can even occur based upon when you execute your trades within a day.  In the graph below, we plot the performance of the –Daily– rebalance strategy in March assuming different execution prices: the next day’s open, the next day’s close, or an estimate of time-weighted average price (“TWAP”).  The performance difference over this one-month period is 85bps.

Just as rebalance timing luck can lead to disparate outcomes, the timing of trades can exacerbate the end results.

Source: Tiingo. Calculations by Newfound Research. Past performance does not guarantee future results. Performance is hypothetical and gross of all fees, expenses, sales charges, or trading expenses with the exception of underlying ETF expense ratios.  Returns assume the reinvestment of dividends.

And this is to say nothing at all of the tactical model actually employed herein (which we touched upon a bit in our What Worked section above)!

One of the hardest problems facing allocators going forward will be disentangling luck from skill.  While most might expect skill to shine in a crisis, it is also the time when the influence of luck can be largest.

Going Forward

There are no do-overs or retries in investing.  Dwelling on the past may only prepare us to fight the last battle and leave us blind to the next.  Nevertheless, we should carry forward lessons learned.  Evaluating what worked and what did not is a first step, but more critically we should ask ourselves, “what was intentional?” and, “what was luck?”

For example, systematically rebalancing our mutual funds daily and our models weekly was an intentional decision: it married both our research on rebalance timing luck and operational constraints.  Rebalancing models each Monday, however, meant being influenced by significant weekend gaps (e.g. -7.4% on March 9^th and -10.4% on March 16^th).

Similarly, our weekly models first undershot, then overshot, our target cash allocations, which proved to be a drag on performance.  How does this occur?  Consider a simple example where a strategy can invest either in equities or cash.  When the tactical signal turns off, and stays off, the strategy transitions over time out of its equity position.  If implemented daily, the strategy makes a consistent transition, selling 5% of the portfolio each day.

Now consider a weekly implementation that rebalances on a Monday.  If the signal turns off on a Friday, the strategy will sell 25% of the portfolio on Monday.  In doing so, it overshoots the daily implementation, which will catch up by week’s end.  However, if the signal turns off on a Monday, the strategy will remain invested all week while the daily implementation de-risks, only catching up on the subsequent Monday.  Hence, it constantly undershoots.

For illustrative purposes only.

In practice, signals can change mid-week and signals may flip on and off, leading to dynamic lead/lag behavior.  When we evaluate the target equity allocation in our –Daily– and –Weekly– implementations of Risk Managed U.S. Sectors model, we see precisely this.

Portfolio allocations above are model holdings and therefore are hypothetical and do not represent actual client portfolios. There is no guarantee that the strategy will achieve its objectives. There are no guarantees that the strategy will be positioned correctly for any given market environment. The strategy utilizes various rebalance techniques designed to reducing transaction costs and turnover, which may result in the strategy’s actual allocation straying from its target allocation.

Undershooting from 3/10 to 3/20 meant the weekly variation was subject to more drawdown.  Overshooting after 3/20 meant that it participated less in the subsequent rally.  Combined, this created an approximate drag of 200 basis points versus the daily implementation (not accounting for implementation costs).

Internally, we aim not to criticize too harshly the poor outcomes of intentional decisions just as we do not laud ourselves for the good fortune from unintentional ones.  Nevertheless, we believe that all decisions should be subject to review: re-evaluating the basis for what is intentional and aiming to further minimize the impact of what is not.

With that in mind, we plan to explore (or re-explore) the following research items over the coming weeks and months to help better inform our strategy construction.

• Weekend Gap Risk: Between market close and the next day’s open, there are 17.5 hours.  Over the weekend, this expands to 65.5 hours.  During a fast-moving crisis, this may lead to a higher probability of gap-down events to the detriment of a weekly rebalancing model that is undershooting its target.
• Dynamic Rebalancing Schedules: For our separate account and model clients, rebalancing on a weekly schedule still invited significant timing luck into our realized performance in Q1. We plan to research both the performance and operational implications of changing our rebalance frequency based upon prevailing market volatility.
• Tranching Speed: In our existing models, we prefer to rebalance “a little, but frequently.” Specifically, in our trend equity mandates, this means transitioning our models over a 20-trading-day schedule (e.g. if the portfolio target changed from 100% invested to 100% divested, and stayed that way, the portfolio would build a 5% short-term US Treasury position per day). The purpose of this approach is to avoid rebalance timing luck.  But it also embeds the assumption that our models maximize their forecast accuracy over a one-month holding period.  Without falling prey to recency bias, we want to revisit this assumption and explore the implications of other – be they static or dynamic – choices for this speed.
• Embedded Convexity: Our trend equity mandates can be best explained as a 0.5 beta equity portfolio overlaid with a long/short trend following model. This approach is particularly weak to sudden and rapid changes in market direction.  In this project, we aim to explore the use of options strategies to provide trades that provide some hedge in the direction opposite of our trend signals.

One of the questions we have received frequently as of late is, “when would you override your models?”  While we have never done so to date, it does not preclude the opportunity.  While our preference is systematic over discretionary, we have internally debated the ability for systematic models to adequately adapt to idiosyncratic, exogenous shocks.  And certainly a pandemic coupled with a complete shutdown of the global economy, a complete breakdown of market structure, and the swift and overwhelming response of the Federal Reserve counts as an exogenous shock.

The problem with discretionary action is that it implies we have greater insight than our models, which are informed by the collective behavior of market participants.  I, for one, am not that confident that I do.

While +7% days are more comforting that -7% days, I would argue that when the public company valuation of the world’s largest economy is swinging by high single digits in either direction, it’s not a healthy sign.

Nevertheless, we must ask: is there an opportunity for the discretionary implementation of complementary systematic models?

For example, we’ve argued that long/short trend following models are inherently linked to options and exhibit mechanical convexity because they approximate the delta exposure of a straddle (i.e. long a put and call).

Let’s say we purchased a straddle at the price of the S&P 500 12 months ago, set to expire in 1 month.  The delta of that straddle would be about -0.33, despite being $15 in-the-money on the put.  Why?  Because volatility is so high right now there is a strong probability that we get knocked back towards at-the-money.

For a long/flat strategy – which is approximately 0.5 beta + 50% exposure to a long/short strategy – this is effectively equivalent to saying we want to have 0.335 beta right now.  A straddle struck at the S&P 500’s price 6 months ago has a delta closer to -0.752.  This would imply a long/flat strategy would want to have a beta closer to 0.125.

Despite negative trends, neither of these are zero simply because volatility is so high.  Markets are too uncertain, increasing the range of possibilities, and leading to more neutral exposure.

Of course, if we were to apply such an overlay in a discretionary manner, it raises the question as to when we would elect to take it off.  We currently have no plan to take this, or similar, courses of action, but are actively debating the ideas internally.

Conclusion

As systematic investors, we have no strongly held views on current macroeconomic conditions.  Rather, our goal is to keep watch for market developments – manifest through quantitative signals – that we believe pose a significant threat (or opportunity!) to investor risk and return.

Though our strategies are implemented systematically, our research remains a human endeavor.  It is through our ongoing research that we aim to evolve by identifying those areas in which our models are currently blind.  We continue to believe that tangential explorations can lead to some of the best breakthroughs, allowing us to circumvent our own biases by looking at a topic from a different angle.

At our core, we remain dedicated to helping investors proactively navigate the risks of investing through industry-leading research and investment acumen.  To achieve that end, however, we must be willing to ask ourselves not only difficult questions about future market dynamics, but also our own process.  Revisiting our closely held views is important for remaining pragmatic about evolution, rather than staying blindly dogmatic to our beliefs.

Thank you for reading and for allowing us to earn your business.  If you have any questions at all regarding our strategies or anything else, please do not hesitate to reach out.

Sincerely,

Corey M. Hoffstein
Chief Investment Officer
corey@thinknewfound.com

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.