Forums  > Trading  > Index trading strategy  
     
Page 4 of 5Goto to page: 1, 2, 3, [4], 5 Prev Next
Display using:  

goldorak


Total Posts: 1050
Joined: Nov 2004
 
Posted: 2018-04-11 18:27
OK. I get you better with this clarification.

Now, do I understand the following right:

> The probability of having the worst underperformance in a calendar year no worse than -20% is 7.3%.

means that only 7.3% of simulations were above -20% underperformance ?

If you are not living on the edge you are taking up too much space.

Energetic
Forum Captain

Total Posts: 1491
Joined: Jun 2004
 
Posted: 2018-04-11 18:49
Can I still use the excuse that English is not my native? Wink

For each simulation run, I select the worst underperformance in a simulation "year" vs. S&P and store that number. In 7.3% cases this number was greater than -20% and in 92.7% cases it was worse than -20%. On average, this pseudo-DD vs. S&P is 40% with stdev of 18%.


For every complex problem there is an answer that is clear, simple and wrong. - H. L. Mencken

goldorak


Total Posts: 1050
Joined: Nov 2004
 
Posted: 2018-04-12 09:48
Ouch ! That hurts.

We could discuss at length the concepts of average and statistics of max drawdowns and I will not here. But your result is very bad from a risk management point of view. Let's assume your trading signal has been informative up till now. As soon as its informational content will disappear, you will have to potentially navigate through enormous drawdowns without knowing if yes or no you should stop using it. You may lose all your hard earned monies before being able to take a reasonable decision.

However, I must say that I am very surprised with your results and my little thinger tells me that you did not carry out this exercise rightly. You must have an issue somewhere in my opinion.



If you are not living on the edge you are taking up too much space.

Maggette


Total Posts: 1067
Joined: Jun 2007
 
Posted: 2018-04-12 11:01
Seems to be a bit extreme. Maybe a bug in the Implementation? Which technology do you use?

If you want to share the reshuffle code here I could have a look at it.

Ich kam hierher und sah dich und deine Leute lächeln, und sagte mir: Maggette, scheiss auf den small talk, lass lieber deine Fäuste sprechen...

Energetic
Forum Captain

Total Posts: 1491
Joined: Jun 2004
 
Posted: 2018-04-12 18:03
I'll confess to a sloppy implementation right away. This was:

random[1:n] -> tableLookup[1:n]

I can't swear that there isn't a bug somewhere but there is very little opportunity to screw up. I'll look again though.

Meanwhile, please elaborate what bothers you. To recap what was done, the strategy signal was frozen, the sequence of market returns was shuffled. For each simulation run, I had the same sequence of positions for the strategy {P1, P2, ..., Pn} where Pi could be long, short or neutral. These positions act upon a random sequence of returns sampled from the set of observed market returns {R1, R2, ..., Rn}. Along with it runs the randomized S&P which is always long. This is how I understood your idea. We have miscommunicated a couple of times before, so I thought it was worth restating it in my words.

Now, part of that static sequence {..., Pj ... Pm, ...} corresponds to the calendar year 2008 where the strategy was short and neutral a lot of the time. In most simulations, the sequence of random returns that falls between j and m doesn't have majority of negative returns as the real 2008 had. It's a "year" like any other with positive expected average return and due to variance some realizations are very positive. Of course when the random sequence {Rj...Rm} has a lot of positives my strategy sucks while S&P flourishes and of course due to the nature of the distribution of returns it happens often. When it does happen the strategy underperforms by a lot. What is surprising or problematic here?

For every complex problem there is an answer that is clear, simple and wrong. - H. L. Mencken

EspressoLover


Total Posts: 348
Joined: Jan 2015
 
Posted: 2018-04-12 18:08
I'm gonna interject here, and say that this discussion (IMO) is way overemphasizing relative performance to the S&P. The strategy only has a beta of 0.3, it's much much closer to an absolute return product than a smart beta product.

Of course, in some scenarios it's going to have significant underperformance relative to S&P. Simply because it's barely correlated with the index. Just because it trades the S&P doesn't mean that it's the right benchmark, anymore than one should focus on relative performance to the Nikkei, oil, bitcoin, or Seattle real estate.

Scalability

> The main issue with your strategy is that it's a daily strategy with the sharpe of a weekly/monthly strategy. This implies that your return per trade is very low, and that the strategy will not scale well.

This is a fair criticism assuming this was trading single-name equities. However, the strategy would be trading the most liquid instrument in the world. The estimated market impact cost of $100 million in ES is 1.25 bps. Let's make some rough assumptions and assume the signal flips every week on average, future returns roughly match backtest, and market impact scales with the square root of size.

That implies the strategy turns over 100x AUM per annum. At 20% CAGR, that's 20 bps returns per notional traded. The strategy could pay 5 bps per trade before returns degraded to less than 75% of zero-impact performance. That gives the fund the ability to do $1.6 billion per clip. Since each clip is 2X AUM (because it flips), that gives a capacity estimate of $800 million AUM.

Theoretical justification

> I have never found a leading indicator in options (and I have looked), and I could never find a non-hand-waving explanation for why there would be a leading indicator in options.

I think you're framing this in terms of alpha or information. In reality, I think the results are driven by time-variation in the equity risk premium (ERP). Pretty much everyone sane accepts that ERP is real. Even totally uninformed investors can earn excess returns relative to the risk-free rate, by taking on equity beta risk.

Almost as broadly accepted, is that the expected ERP varies over time. It would be silly to expect that ERP stays static forever. Equilibrium shifts, and whatever the process that makes ERP non-zero is almost certainly going to vary over time. Academic finance has done a pretty good job of showing that the sizable majority of day-to-day market variance is driven by changes to the discount rate, not changes to expected cash flows.

The ERP isn't directly observable, but there's no reason to believe that it's adversarially hidden, like alpha. Alpha's extremely rivalrous, and is quickly consumed by the first traders to access it. But the sizable majority of investors are highly inelastic to index prices. If they weren't, market volatility would be significantly lower, since most market movements aren't driven by revisions to expected cash flows.

Consider that market-wide CAPE is a pretty good predictor of 10-year forward returns. This isn't a hard predictor to generate. There's certainly no way you could call it "alpha". Yet the market is highly inelastic to this measure, CAPE does vary substantially over time, and there's no sign of this relationship being arb'd away any time soon.

It seems pretty likely to me that the VIX curve has a pretty deep relationship with the ERP. The curve could be backwardated either because of genuine vol expectations, i.e. maybe 30-day real volatility will be higher than 120-day volatility. In which case it shouldn't have any predictive value. But nearly every well-studied term structure invalidates the expectation hypothesis. This includes VX, where future realized vols have only a very weak relationship to term structure prices. More likely the VX curve is proxying for market-wide risk aversion.

This isn't an endorsement, and I don't know if this is actually the case. But I am saying that there does seem a viable justification for why this thing might work. The one thing that does "smell" to me is that the strat makes money on its short positions (rather than just conserving capital and reducing long-run beta). I do believe that there are periods where the ERP significantly compresses to near-zero, but it's a lot more extraordinary claim to believe that ERP regularly goes negative.

Good questions outrank easy answers. -Paul Samuelson

tradeking


Total Posts: 18
Joined: May 2016
 
Posted: 2018-04-12 18:52
>This is a fair criticism assuming this was trading single-name equities. However, the strategy would be trading the most liquid instrument in the world. The estimated market impact cost of $100 million in ES is 1.25 bps. Let's make some rough assumptions and assume the signal flips every week on average, future returns roughly match backtest, and market impact scales with the square root of size.

The 1.25 bps that you quoted is for a VWAP over the entire day. The strategy in question in this thread requires the intraday data up until the final hour before the close to decide on the position it wants to keep overnight, and it will only have 1 hour to enter into this position. That will probably cut the capacity by at least 80% of your estimate (which would bring it into my original ballpark estimate of ~$10mm pnl a year).

Energetic
Forum Captain

Total Posts: 1491
Joined: Jun 2004
 
Posted: 2018-04-12 19:01
Thank you, EspressoLover.

I don't have a strong opinion on the classification. It seems to me that beta is very much suppressed by the outliers, not sure if it matters. I'll just post a picture which may be worth a 1000 words. Diagonal is added for illustration to separate under- and over performance.


For every complex problem there is an answer that is clear, simple and wrong. - H. L. Mencken

EspressoLover


Total Posts: 348
Joined: Jan 2015
 
Posted: 2018-04-12 19:08
@tradeking

That's 100% a fair criticism of my estimate.

But I believe @energetic mentioned that the signal flips about once per week. That implies a half-life of 2-3 days. Which seems about in line with about how fast the VIX term structure changes. It would seem that an actual fund would have much longer than an hour to scale into positions.

The convention of using the data just prior to the close was to shoehorn a straight-forward interday backtest. My gut feeling is that if you tested this thing intraday, trading at the time-of-day the signal actually changes, and using reasonable VWAP assumptions, that the performance would look very similar to the interday backtest with quite a bit of capacity. But again, just an intuitive gut feeling. At the end of the day it's an empirical question.

Good questions outrank easy answers. -Paul Samuelson

tradeking


Total Posts: 18
Joined: May 2016
 
Posted: 2018-04-12 19:38
>My gut feeling is that if you tested this thing intraday, trading at the time-of-day the signal actually changes, and using reasonable VWAP assumptions, that the performance would look very similar to the interday backtest with quite a bit of capacity. But again, just an intuitive gut feeling. At the end of the day it's an empirical question.

Then it sounds like we both agree on the original point I was making with my first post in this thread:

"If you can find a way to diversify your signal to bump the sharpe up to 2 or so, then it could be something interesting. Or if you can find a way to slow down your signals so that you only trade once a week or month while keeping the same sharpe, that could be interesting too."

goldorak


Total Posts: 1050
Joined: Nov 2004
 
Posted: 2018-04-12 19:51
@ExpressoLover.

I strongly disagree with you and I am very surprised how academic towards the "benchmark" and beta your thinking is in this particular case. We are talking risk management of one single strategy on one underlying taken in isolation here and in such a case beta is exposure, not correlation-adjusted volatilities ratio ( Confused ). We are not in a situation of a large portfolio, and even in such a case just hearing the word "correlation" in the portfolio construction process drives me crazy. It reminds me of people talking of a strategy that can be long or neutral the S&P 500. Under the deceptive pretext that historically the strategy was long only 50% of the time, they now sell it to you as a strategy with half the S&P 500 volatility. This is great until the strategy is long in a period with extremely high volatility.

As you will have noted, as soon as Energetic posted his first message, everybody has been referring to the performance relative to the S&P 500, even him in his table of performances. People are astonished by the performance in 2008, and again later he emphasized his year to date performance relative to the S&P 500.

Positions are taken on the S&P 500, long, short or neutral. The signal is based on the VIX, which is a near parent (probably uncle Scrooge) of the S&P 500. The risk is nothing else but the one of the S&P 500. The beta in this case is the exposure, namely 1, 0 or -1, and not 0.3 which is the result obtained by academics doing regressions. Averages never saved anybody's ass as long as I can remember. The day the S&P 500 has a sequence of returns : +5%, -5%, +5%, -5%, +5%, -5%, and the signal is -1, +1, -1, +1, -1, +1, S&P 500 is down 80bps while the strategy is down 27%.

"But I swear, my beta was 0.3, this is just a black swan" will not help here.

I know, I am an extremist.

@Energetic.

relative_perf = signal * shuffled_sp_rets - shuffled_sp_rets

right? and not

relative_perf = signal * shuffled_sp_rets - sp_rets

If you are not living on the edge you are taking up too much space.

goldorak


Total Posts: 1050
Joined: Nov 2004
 
Posted: 2018-04-12 19:53
@Energetic:

Do the same plot on daily returns.

If you are not living on the edge you are taking up too much space.

Energetic
Forum Captain

Total Posts: 1491
Joined: Jun 2004
 
Posted: 2018-04-12 21:29
Of course:

relative_perf = signal * shuffled_sp_rets - shuffled_sp_rets

The daily plot is, IMO, uninteresting but here you go:


For every complex problem there is an answer that is clear, simple and wrong. - H. L. Mencken

goldorak


Total Posts: 1050
Joined: Nov 2004
 
Posted: 2018-04-13 10:03
> IMO, uninteresting

Not less than the monthly one that artificially diffuse the information. Blush

My (quick) notes and conclusions on your strategy are the following:

Notes:

1) Data snooping. The strategy is most probably a fluke. Its performance is based on an extremely small sample of volatility bursts that count for all of the outperformance. For the strategy's designer it is easier to think it is not because he is convinced of having done his job properly. However, just looking at what happens in 2008 (the data points are pretty evident from the daily returns scatter plot), I am not sure this system will be doing a lot better in the long run than a plain

be long
if (VIX > threshold1) {
then
go to cash
if (VIX > threshold2) {
be short
}
}

2) Decisions. From your simulations and from your backtest, it appears that, would the potential informational content of your signal be zero, the potential drawdowns of your strategy would be so large that any reasonable tactic to allocate the strategy would be difficult to put into place.

3) Unclear trading rules and mechanics. There is very much obscurity on the alignment of your signal and trades. You lose 5% a year just delaying the trade at the close and the open. As I told you, there are plenty of other historical details like the time of the closing price of SPY that can enter the game.

4) Data. You definitely miss perspective. You have only tested very recent periods, and plenty of market configurations, like the 95-00 years, are missing to give better perspective.

5) Context. I fully understand excitement in the present outperformance context. However would we be at a moment when the strategy is in a large dradown, you would have yourself doubts on the matter. Back to my point on drawdowns: when the phuck do I stop using this strategy ?

Conclusions:

1) A single timing strategy on one single underlying with binary positioning is extremely difficult to manage (sizing, retirement of the strategy).

2) This kind of strategy should be mixed with a number of different ones. Different by their source of information, different by their trading horizon and different by their underlyings. Note that this is no different from what a legit CTA would do.

3) You probably need to joint venture with an able CTA out there to make this strategy more robust... or find a gullible institutional (unfortunately plenty available out there Hammertime )




If you are not living on the edge you are taking up too much space.

ronin


Total Posts: 384
Joined: May 2006
 
Posted: 2018-04-13 14:36
At the risk of hijacking the thread:

> Pretty much everyone sane accepts that ERP is real.

Whoa there @el. I'm happy to put myself in the insane part of the population on this one. In my humble opinion, "risk premium" of any sort is the most misused concept in finance.

Before you go off explaining to me how rational investors behave, just pick two random equities.

Say one of them makes 76k deliveries, has revenue 7 bln and loss of 1 bln.
The other has 10 mln deliveries, has revenue 165 bln and profit 10 bln.

3 month ATM volatility is pretty much the same for both of them.

What would the ERP be for either of those?

Because the market values the first one at 51 bln, and the second one at 50 bln. It's TSLA and GM. And it's not a temporary fluke either - they have been there for over a year now.

That would be the exact same market where you think sophisticated investors make detailed ERP calculations and choose to transmit them to the rest of us through their activity in some thin options, where most contracts don't even trade on any given day. Give me a break.


> The 1.25 bps that you quoted is for a VWAP over the entire day.

That actually bothers me the least. Without looking at any numbers, roughly half of the VWAP happens in the hour after the open, and the other half happens in the hour before the close. Mid day contributes maybe 10-20%. So restricting to the last hour it may take off 50-60% from the estimate, but it's not an order of magnitude thing.


"There is a SIX am?" -- Arthur

Energetic
Forum Captain

Total Posts: 1491
Joined: Jun 2004
 
Posted: 2018-04-13 16:49
It was uninteresting in the sense that it was entirely predictable. There's no way it could look otherwise.

1) Data snooping. Yes and it was pretty much unavoidable. With each failed version of the strategy I learned something about data. The only clean part is live. I suspect every developer is guilty of snooping on some level.

2) Agreed on large temporary drawdowns vs. S&P. But we also know what S&P itself is capable of. Allocation tactics is a separate problem. Just because it's not trivial doesn't mean it's not worth considering.

5) I'm still thinking about a better answer.

Thanks again for your comments. Would you know how I could meet gullible institutionals? Wink

For every complex problem there is an answer that is clear, simple and wrong. - H. L. Mencken

goldorak


Total Posts: 1050
Joined: Nov 2004
 
Posted: 2018-04-13 18:17
Did you have a look at empirical distributions of returns conditional on +1, 0 or -1 signal, for both S&P 500 and your strategy? This may be pretty informative.

If you are not living on the edge you are taking up too much space.

Energetic
Forum Captain

Total Posts: 1491
Joined: Jun 2004
 
Posted: 2018-04-13 19:49
I looked but came to no conclusions. Here're e.g. distributions conditional on being short.


For every complex problem there is an answer that is clear, simple and wrong. - H. L. Mencken

goldorak


Total Posts: 1050
Joined: Nov 2004
 
Posted: 2018-04-13 20:18
What piece of software are you using to produce such horrible plots?

I would say that roughly skew is influenced.

If you are not living on the edge you are taking up too much space.

Energetic
Forum Captain

Total Posts: 1491
Joined: Jun 2004
 
Posted: 2018-04-13 21:02
Blue and red bars are obviously symmetric WRT 0 in this case.

For every complex problem there is an answer that is clear, simple and wrong. - H. L. Mencken

nikol


Total Posts: 573
Joined: Jun 2005
 
Posted: 2018-04-23 17:02
collective2

Energetic
Forum Captain

Total Posts: 1491
Joined: Jun 2004
 
Posted: 2018-04-23 17:50
Yes, I've been thinking about this. What bothers me there is that all trades become public so that your strategy could be mined.

For every complex problem there is an answer that is clear, simple and wrong. - H. L. Mencken

NeroTulip


Total Posts: 1016
Joined: May 2004
 
Posted: 2018-04-24 03:17
Or First loss funds

Not a serious recommendation... I don't see why someone would put in the work to manage money and take half of the profits but 100% of the losses - except overconfidence.

"Earth: some bacteria and basic life forms, no sign of intelligent life" (Message from a type III civilization probe sent to the solar system circa 2016)

HitmanH


Total Posts: 465
Joined: Apr 2005
 
Posted: 2018-04-24 07:45
It works for higher leverage to higher sharpe strategies.
Paulson doing it - WOW...

NeroTulip


Total Posts: 1016
Joined: May 2004
 
Posted: 2018-04-24 09:17
At what Sharpe does it start making sense to use first loss funding? Not even sure how to think about this rationally... but definitely not at Paulson’s Sharpe LOL. Seems quite desperate.

Has anyone seen the returns of these first loss funds? Must be pretty good if they take half of the gains and none of the losses on a portfolio of high Sharpe strategies.

"Earth: some bacteria and basic life forms, no sign of intelligent life" (Message from a type III civilization probe sent to the solar system circa 2016)
Previous Thread :: Next Thread 
Page 4 of 5Goto to page: 1, 2, 3, [4], 5 Prev Next