Thursday, March 05, 2020

Why does our Tail Reaper program work in times of market turmoil?

I generally don't like to write about our investment programs here, since the good folks at the National Futures Association would then have to review my blog posts during their regular audits/examinations of our CPO/CTA. But given the extraordinary market condition we are experiencing, our kind cap intro broker urged me to do so. Hopefully there is enough financial insights here to benefit those who do not wish to invest with us.


As the name of our Tail Reaper program implies, it is designed to benefit from tail events. It did so (+20.07%) during August-December, 2015’s Chinese stock market crash (even though it trades only the E-mini S&P 500 index futures), it did so (+18.38%) during February-March, 2018’s “volmageddon”, and now it did it again (+12.98%) during February, 2020’s Covid-19 crisis. (As of this writing, March is up over 21% gross.) There are many names to this strategy: some call it “crisis alpha”, others call it “convex”, “long gamma” or “long vega” (even though no options are involved), “long volatility”, “tail hedge”, or just plain old “trend-following”. Whatever the name or description, it usually enjoys outsize return when there is real panic. (But of course, PAST PERFORMANCE IS NOT NECESSARILY INDICATIVE OF FUTURE RESULTS.) Furthermore, our strategy did so without holding any overnight positions.

Why is a trend-following strategy profitable in a crisis? A simple example will suffice. If a short trade is triggered when the return (from some chosen benchmark) exceeds -1%, then the trade will be very profitable if the market ends up dropping -4%. Vice versa for a long trade. (As recent market actions have demonstrated, prices exhibit both left and right tail movements in a crisis.) The trick, of course, is to find the right benchmark for the entry, and to find the right exit condition.

Naturally, insurance against market crash isn’t completely free. Our goal is to prevent the insurance cost, which is essentially the loss that the strategy suffers during a stretch of bull market, from being too high. After all, if insurance were all we want, we could have just bought put options on the market index, and watched it lost premium every month in “good” times. To prevent the loss of insurance premium requires a dose of market timing, assisted by our machine learning program that utilizes many, many factors to predict whether the market will suffer extreme movements in the next day. In most years, the cost (loss) is negligible despite the long bull market, except in 2019 when we lost 8.13%. That year, which seems a long time ago, the SPY was up 30.9%. (It was in the August of that year that we added the machine learning risk management layer.) But most investors have a substantial long exposure. A proper asset allocation to both Tail Reaper and to a long-only portfolio will smooth out the annual returns and hopefully eliminate any losing year. (Again, PAST PERFORMANCE IS NOT NECESSARILY INDICATIVE OF FUTURE RESULTS.)

But why should we worry about a losing year? Isnt’ total return all investors should care about? Recently, Mark Spitznagel (who co-founded Empirica Capital with Nassim Nicholas Taleb) wrote a series of interesting articles. It argued that even if a tail hedge strategy like ours returns an arithmetic average return of 0%, as long as it provides outsize positive returns during a market crisis, it will be able to significantly improves the compound growth rate of a portfolio that includes both an index fund and the tail hedge strategy. I have previously written a somewhat technical blog post on this mathematical curiosity. The gist of the argument is that the compound growth rate of a portfolio is m-s^2/2, where m is the arithmetic mean return and s is the standard deviation of returns. Hedging tail risk is not just for the psychological comfort of having no losing years - it is mathematically proven to improve long-term compound growth rate overall.

PAST PERFORMANCE IS NOT NECESSARILY INDICATIVE OF FUTURE RESULTS.

For further reading on convex strategies, please see the papers by Paul Jusselin et al “Understanding the Momentum Risk Premium: An In-Depth Journey Through Trend-Following Strategies” and Dao et al “Tail protection for long investors: Trend convexity at work” (Hat tip to Corey Hoffstein for leading me to them!)

43 comments:

Jemma Riddiford said...

Hi Ernie

Very interesting blog post, especially given the current financial climate (but of course that's precisely why you've published the article now).

May I ask, why trend-following strategies and not an option-based strategy like Taleb and Spitznagel would employ? Apparently Universa Investments purchase 30-35% OTM puts on the S&P 500 with expiry in 60-90 days, rolling them out at 30 days until expiry. Wouldn't this be a better way to capitalise on market turmoil, capturing the sudden spike in volatility with the explosive growth in the option premium?

Jemma

Ernie Chan said...

Hi Jemma,
The question isn't whether long puts will make money in a crisis - of course they will! The question is how you can prevent them from losing large amount of premium during peace time. Do you have any idea how Taleb did it?
Ernie

Jemma Riddiford said...

Hi Ernie

Would it matter if they lost premium during peace time? Surely that would be the time to leverage mean-reversion up to the eyeballs, ensuring there's more than enough cash to purchase options that would cover losses in the mean-reversion portfolio during black swan events (and then some, if the anecdotes of Universa's performance are to be believed)?

Jemma

Ernie Chan said...

Hi Jemma,
It depends on how much premium it loses during peace time.
Our strategy is designed so that it loses minimal amount, while capturing about 50% return YTD.
Ernie

Anonymous said...

Hi Ernie,

Very interesting blog post and congratulations on your March performance! Seems Tail Reaper would fit nicely in a traditional 60/40 portfolio. Just to make sure I understand. The Tail Reaper is an intraday trend following system on the S&P 500 e-mini future?

Thanks,
James

Ernie Chan said...

Hi James,
Thanks!
Yes, Tail Reaper is intraday trend following on ES.
Ernie

Anonymous said...

Hi Ernie,

What is the transaction costs of liquid stocks in US stocks market? Many thanks!

Anonymous said...

Hi Ernie,

Is cryptocurrency a good target to trade?
Is there anything we should pay special attantion to? Many thanks.

Ernie Chan said...

Anon,
Liquid stocks have transaction costs of about 2-5bps per trade.
Ernie

Ernie Chan said...

Yes, one can profit from trading crypto at high frequency, and through cross exchange arbitrage. But credit worthiness of the exchanges are critical.
Please read my 3rd book Machine Trading's chapter on Bitcoins for some more nuances.
Ernie

Anonymous said...

Hi Ernie,

Where can we download free long historical tick data of cryptocurrency? Many thanks.

Ernie Chan said...

Hi,
You can download using Gemini exchange's API. Otherwise, please check out the Bitcoin chapter in my 3rd book Machine Trading.
Ernie

Anonymous said...

In response to those talking about Universa, it seems that someone actually posted the interim letter sent to investors online:

https://www.scribd.com/document/455607584/Universa-Letter-April-2020

Says that in 2020 they had made 4144% by April but the life to date return was 239%, so it does seem like they are burning a lot during peaceful times, but still increases CAGR.

Anonymous said...

Great article. However I do have a question for Ernie. How do you know whether intraday momentum strategies can stay profitable? For example after 2004 or so large opening gaps on major indices correlated to their daytime performance that day, especially near the close. But for at least a decade before 2004, this correlation between big gaps and the day returns was negative. Do you know what caused this shift and how you might be able to predict a future regime change? I do not know why

Ernie Chan said...

Hi,
Indeed, there is no guarantee *any* strategy will work forever. Hence one must monitor the market environment and understand whether the fundamentals or market structure have changed. One cannot predict regime change - one can only adapt to those changes once they occur.

The specific phenomenon you observed is due to the increasing volume of gamma hedging by options market maker, rebalancing by levered ETPs (and other levered vehicles), and variance swap delta hedging.

Ernie

Anonymous said...

I took a look at qtscm's returns. They are quite impressive! I've tried to trade intraday strategies myself without as much success, including variants of the index momentum strategies you discussed in your book, Algorithmic Trading. How do you minimize or negate the effects of slippage in your execution? I'm trading Micro-ES futures but I generally incur a non negligible amount of losses due to the spread, which makes my strategy unprofitable if the timescale is too restrictive.

Ernie Chan said...

Thanks!

There is not much one can do about slippage in a momentum strategy except to execute in small batches, and holding for long enough time.

Ernie

Anonymous said...

Ernie,

One more thing I wanted to ask: Are you familiar with the strategy on index funds that goes something like this:

If the index fund had intraday movements in the same direction for two trading days in a row and the more recent one is large (lets just say >1.5%), bet in the opposite direction during the next trading day (enter on open and exit on close).

This is a mean reversion strategy but I've noticed it also generates significant returns during times of crisis, much like your intraday momentum strategies. I am again unsure about why this works but was curious if you had any idea.

Ernie Chan said...

There are all sorts of patterns that work during certain periods. The question is whether they work over long periods. Have you backtested this? What is the Sharpe?

Ernie

Anonymous said...

Here it is on SPY without trade fees, slippage, or short fees: https://ibb.co/7tfmsGZ

Most popular brokerages are commission free now and since its SPY and intraday I doubt you'd incur much slippage or any short fees either. But unfortunately I am not using the auction prices for open and close since the data is too expensive.

The sharpe ratio in sample is ~4.7 but I'm not sure if that really means much since the number of trades is very sparse (only 98 since 2008-05-06). Using 2% as a cutoff instead of 1.5% raises the sharpe ratio too.

Ernie Chan said...

I did a quick backtest, and it wasn't profitable from 2010 to the present.
If you want to compare code/output, please email me.
Ernie

Hari K. said...

Hey Ernie, your blog is excellent, thanks. Can you pls explain one crucial point about Tail Reaper? Clearly there is a trade off between eliminating options bleed and the danger that you will have late entry into a crash because of path dependency issues. Is your system guaranteed to go short given a down move beyond a threshold? Many thanks, Hari K.

Ernie Chan said...

Hi Hari,
We do not "guarantee" a short when the price level drops below the "support" level, due to our machine learning risk management layer.

There is no guarantee that a sharp decline isn't followed by a sharp intraday reversal. We have seen that movie a few times in March-June. Our risk management system is to prevent that from happening.

Ernie

MonsieurMagique said...

Hi Ernie,

If you plot historical overnight close-open returns vs next day open-close returns for SPY, it looks like a ball of points, which indicates no correlation (and also no correlation in the tail). Does your Tail Reaper strategy identify downward momentum which only lasts through a portion of the trading day before a reversal?

Many thanks,
Henry

Ernie Chan said...

Hi Henry,
Indeed, Tail Reaper only benefits from intraday momentum. Nevertherless, that's enough to hedge the tail risk of an overnight position, if a proper hedge ratio is used. This is a point that was discussed in the podcast: https://blog.thinknewfound.com/podcast/s3e5-ernest-chan/

Ernie

Anshul Saxena said...

Good Morning Ernie,

Your strategy looks very interesting to me. Is it possible to implement it on crude oil futures after doing a event study for the reasons behind its historical up and down

Ernie Chan said...

Hi Anshul,
The Tail Reaper strategy may work with crude oil futures - but it depends on the notional size of the derivatives re-balancing tied to crude oil volatility.

Of course, metalabeling is a concept that works on any strategy, even discretionary ones.
Ernie

MonsieurMagique said...

Hi Ernie,

I listened to the podcast episode - great stuff. I have a follow-up question for you on intraday momentum. Do you observe statistically significant intraday momentum following overnight negative returns? Or is this not a requirement for Tail Reaper to perform well?

Many thanks and best regards,
Henry

Ernie Chan said...

Thanks Henry!

Overnight negative return is certainly part of the input to Tail Reaper. If that is large, it does foretell momentum.

Ernie

MonsieurMagique said...

Thanks Ernie. The issue I have is that when I look at historical SPY, I see no statistically significant intraday momentum following large negative overnight returns days - does one just need to trust a small historical edge from a small sample in order to proceed with the meta-labelling? Or does the overnight negative return only 'foretell' when combined with other predictive factors?

Best,
Henry

Ernie Chan said...

Hi Henry,
You do not need to use overnight return alone to predict intraday return. That's just part of the input. There are other conditional factors, as you suggested.

Ernie

Alex said...

Hi Ernie, I listened the podcast yesterday and was hyped to start playing with Random Forest. I have two questions I am hoping you could answer.

1. Do you ever plan to make a tutorial like your "mean rev videos" on how to approach ML, build a Random Forest in the spirit of how Tail Reaper is built?

2. Does Tail Reaper only work in high volatile markets? if so how does your company approach bull markets?

Ernie Chan said...

Hi Alex,
1) Yes, I am planning one for Sept 13 and 19, and it is free! Register here: predictnow.ai/register_workshop.
2) Yes, TR only works in highly volatile market. In a bull market, we long the index.
Ernie

Alex said...

Hi Ernie,

Fantastic! looking forward to the 13th. Regarding the behavior of TR in bull market, do you still use the probability outcome to size the long investment type? I am trying to understand what the typical trades look like in a long only portfolio that employs TR.

Ernie Chan said...

Hi Alex,
Yes, both long and short trades are sized by ML, whether bull / bear market.
Best,
Ernie

Alex said...

Hi Ernie,

Thanks for the reply. I watched your talk "The Peculiarities of Volatility" and I was wondering whether you looked at using volatility as a way to size trades? I understand the TR probability outcome is used today to scope the size. Would it make sense to add volatility as a confirmation layer or simply feed volatility as a feature to the TR model which if important anyway influences the probability and thus the size?

Thanks,

Alex

Ernie Chan said...

Hi Alex,
You can safely assume that volatility is one of the features we use as input to the ML program. One should not, however, use volatility directly to size trades, contrary to what the Turtles Traders recommend. You do not want to lower allocation when volatility is high, because that's exactly when you need the full power of tail hedging.
Best,
Ernie

Henry said...

Hi Ernie,

In your podcast with Corey, you mentioned Fast Fourier Transform. Is Fourier Transform used to remove noise from the indicators as features into TR ? Any useful resources you could point to, to explore the use of Fourier into the making of indicators?

Best,

Henry

Ernie Chan said...

Hi Henry,
While I can't comment on the specific signal processing techniques we applied to TR, you should definitely check out FFT or wavelet transforms. Many other techniques are discussed in signal processing textbooks or in Lopez de Prado's book, but beware of any that involves look-ahead bias.
Ernie

Henry said...

Thanks Ernie. Last question, have you looked at the use of Wavelet Neural Network in forecasting time series? If so, what are some of the downside you foresee using ML in forecasting future frequencies given the assumptions that stock returns contain frequencies.

Best,

Henry

Ernie Chan said...

Hi Henry,
No, I haven't applied wavelet in conjunction with NN. But it sounds like a fine idea to me.
Ernie

Hei said...

Hi Ernie,

Congrats on such a great year for your tail Reaper strategy. I do have a question though. You mentioned that some patterns may last multiple days but your TR strategy is only intraday. Shouldn't TR strategy holds position over night then if a pattern last multiple days ? Are you just referring to the "regular trading sessions" over consecutive days ?

Thanks.

Best Regards,
Hei

Ernie Chan said...

Hi Hei,
The trends that last multiple days have become rarer these days. Those trends would be what typical CTA funds tried to capture, with regrettable results in the last few years.
Ernie