Australian (ASX) Stock Market Forum

How is quantitative system trading profitable?

Thanks for relating back to the thread Mazza, i got a whiff of what you were saying but it was complex :) Im still too new to care about semantics:D
 
Hi tech,
Could you elaborate further? - I've hardly seen any technical analysis, unless you count plotting the underlying price on a 2d plane, used in quant analysis.

It's more akin to what skc describes - corr, auto corr,co-integration, residuals on corr etc , basically relative value metrics.

Did I really say all that? I have no idea on what half those terms means.

But I do conduct most of my trading - technical, fundamental or quants - on relative value metrics. To me that offers higher accuracy and logic than without.
 
I don't see any clear distinctions here of Technical Analysis compared to Quantitative Analysis, so here's my interpretation based on Howard's 10,000+ hours of studying and using both types of analysis. I'll then address the question asked of this thread.


Any Technical Analysis is essentially reduced to finding statistically significant relationships. Quantitative Analysis seeks to quantify how statistically significant those relationships are. QA is aligned with mechanical systems trading as it seeks to execute trades based on precise information and timing. Whereas TA allows for any type of logic, reasoning, rational and irrational, more art than science.


How is quantitative system trading profitable?

Okay I'll give an answer that favours either method.
Take for instance, momentum. Totally quantifiable
and it's use also subject to a viewers interpretation.
Depending on your method of analysis you could
trade and be profitable or unprofitable. Profitability
is not mutually exclusive to analysis.
 
Rocket scientist is developing a cancer identification instrument through protein identification using light refraction,
So I guess you couldn't get further from finance!

That is awesome!!! At least he is putting that great mind to good societal use other than providing liquidity :p:

But I do conduct most of my trading - technical, fundamental or quants - on relative value metrics. To me that offers higher accuracy and logic than without.

+1 e.g. index beta trackers, implied v realized vol, discount arbs, convergence of tenor in futures
 
That is awesome!!! At least he is putting that great mind to good societal use other than providing liquidity :p:



+1 e.g. index beta trackers, implied v realized vol, discount arbs, convergence of tenor in futures

Mazz
Youve tweeked my curiosity
I dont use any of that.

Would you ( or anyone else) mind explaining how you go about this and what you look for.
Then how you apply it in a practical sense to your trading.

Appreciated.
 
Would you ( or anyone else) mind explaining how you go about this and what you look for.
Then how you apply it in a practical sense to your trading.

Appreciated.

Professionally, I've never worked in equities space, but the methods mentioned are fungible/transferable to this market.

Index Beta Trackers
What to look for?: Premise is to compare a custom index of products vs. target that tracks the counterpart well e.g. for index vs. custom basket of stocks. Identification can be based on beta, correlation &/or clustering e.g. k-means

Practical: Once the basket has been identified, you can take the moving average - above would be MA of index and custom basket and trade the crossovers as signals

Implied v Realized vol
When I mention vol = volatility, not volume.
Observe the response of implied vol in the options to changes in realized vol. Realized vol can be modeled with GARCH etc and implied vol data fit with some parametric skew to reflect the smile in the market you are dealing with.

Practical: Involves trading the spread between the two measures. Analogous to pairs trading when correlations diverge/converge.
From a retail perspective, replicate the vols using ATM gamma in the options (short for converging spread, long for diverging spread).

Discount Arbs
The term is a misnomer since its not a pure arbitrage. The relative value being to basically you replicate positions [derivative] cheaper than current market value. From a buy side perspective, you'd have to take on some price &/or vol risk to replicate

Practical: E.g. replicate butterfly spread at less than mv, long the wings with a view to vol increase, then short the body

Convergence of tenors
Relationships between spread of the prices in futures contracts from month to month, being cognizant of the term structure and news released.

Practical: e.g historically max spread between May/Jun futures is $1.00 . As of today the May/Jun spread is $1.50. Play for convergence of spread to $1.00 [assuming flat term structure]
 
mazza,

I'm not quite clear on whether you are saying these are the main strats you trade; that is the implication of post #24;
+1 e.g. index beta trackers, implied v realized vol, discount arbs, convergence of tenor in futures

but this quote confuses me a bit;
Professionally, I've never worked in equities space, but the methods mentioned are fungible/transferable to this market.

so this is saying all your option strats are in anything but equities ie fx, commodities, Treasuries/rates etc, indices(stiil in equity space?), or are you saying here that you stick to options or strategies including options, as opposed to just long/short equities?

in short, if you are trading the 4 strats you list, what space are they in?



mind if I drill down a bit more on the strategies just to be clear?

Implied v Realized vol - From a retail perspective, replicate the vols using ATM gamma in the options (short for converging spread, long for diverging spread).

is this a fair summary;
replicate (or just become exposed to) the IV by using the options, with the simplest way being using the ATM straddle (sell when IV relatively high, buy when IV relatively low)? Then replicate Realised vol by delta hedging the underlying (is there another way)?
Close out position if spread converges by eg IV dropping to meet RV (in case where IV>RV), or run till expiry to harvest the difference if they never converge?
In comparing IV-->RV are you using an 'offset' of some sort to adjust RV to account for the fact that RV if measured in the usual way (eg HV45) usually underestimates potential RV because it uses a small sample which does not usually include a fair representation of the whole population of possible returns?
(ps I am using RV interchangably here with Statistical or Actual or vol, correct me if you mean something different)
If that is what you mean , what sort of securites do you find it works best in?



The relative value being to basically you replicate positions [derivative] cheaper than current market value. Practical: E.g. replicate butterfly spread at less than mv, long the wings with a view to vol increase, then short the body

how do you replicate a butterfly at less than current mv ? obv you can replicate it all sorts of ways but where do you find components out of line pricewise by enough for a retail punter to take advantage before the big boys do?


Practical: e.g historically max spread between May/Jun futures is $1.00 . As of today the May/Jun spread is $1.50. Play for convergence of spread to $1.00 [assuming flat term structure]

as most futures spreads are tied to the cost of carry, again, where do you find spreads out of line by enough to profit such as in the above example? surely any time it gets out of line enough it would be arbed back in pretty quickly, otherwise it would be free money. Either that, or there is a reason for it to be out of line (seasonal factors, maybe actual production or transportion issues?), but then it wouldnt be a relative value trade, rather a punt on whether the markets expectations are right.
The only exception to this that i know of the VIX curve, which i do trade on the premise you have given, as well as the time decay aspects. If there are others about would love to know about them.


thanks:)
 
so this is saying all your option strats are in anything but equities ie fx, commodities, Treasuries/rates etc, indices (stiil in equity space?), or are you saying here that you stick to options or strategies including options, as opposed to just long/short equities?
I've traded equity + index ops in a personal account in the past, however for work, yes, its been anything but equities.

replicate (or just become exposed to) the IV by using the options, with the simplest way being using the ATM straddle (sell when IV relatively high, buy when IV relatively low)? Then replicate Realized vol by delta hedging the underlying (is there another way)?
Close out position if spread converges by e.g. IV dropping to meet RV (in case where IV>RV), or run till expiry to harvest the difference if they never converge?
Yes, but as you know the rub is transaction costs of dynamic hedging. I've found that if vol model signals are very short durations e.g. <7 days, trade the fly for exposure, with wings as static hedges and avoid dynamic hedging.

In comparing IV-->RV are you using an 'offset' of some sort to adjust RV to account for the fact that RV if measured in the usual way (e.g. HV45) usually underestimates potential RV because it uses a small sample which does not usually include a fair representation of the whole population of possible returns?
RV/stat vol would be calculated and analyzed using high frequency data to overcome the sample limitations of daily data w.r.t to vol (greater sample size -> central limit theorem for parameters). iirc correctly we discussed using an alternative rv measure (Parkinsons?) in another thread if you want to work exclusively with daily data.

how do you replicate a butterfly at less than current mv ? obv you can replicate it all sorts of ways but where do you find components out of line pricewise by enough for a retail punter to take advantage before the big boys do?
In my previous post, I mentioned that if you're not on the sell side, for vanilla ops you'd have to take on some price &/or vol risk to replicate - basically legging into the position e.g. for the fly at small debit or credit [equivalence].

So it's strongly contingent on weighing vol &/or price model's edge vs. risk of vega.

as most futures spreads are tied to the cost of carry, again, where do you find spreads out of line by enough to profit such as in the above example? surely any time it gets out of line enough it would be arbed back in pretty quickly, otherwise it would be free money. Either that, or there is a reason for it to be out of line (seasonal factors, maybe actual production or transportation issues?), but then it wouldn't be a relative value trade, rather a punt on whether the markets expectations are right.
It would be a combination of relative value and expectations, you wouldn't blindly trade a convergence/divergence output from your model if there is a valid reason for the [new] spread existing.

I mentioned this with VIX-like products in mind. The other market that the VIX has comparable structures are FI tenors, but the spreading will be on different classes e.g. govt and corporate spreads to converge (bullish play on economy), and matching of short/long legs would be duration (and possibly convexity) based where duration is to delta, and convexity is to gamma in ops for FI securities.
 
mazza, thanks for taking the time to reply to my many questions. that clears up a few things for me
So you work in a bank in quant analysis or something like?

Yes, but as you know the rub is transaction costs of dynamic hedging. I've found that if vol model signals are very short durations e.g. <7 days, trade the fly for exposure, with wings as static hedges and avoid dynamic hedging.

Indeed they are. I have tried diligently to trade vol from the long side as well as the short side for balance.

however bearing in mind the following rule of thumb ;
- when long vol, you have to adjust/hedge frequently, as that is where your profit is coming from most of the time. failure to take a profitable hedge (round trip) is money lost forever and will see time decay (and possibly vega) eat away at you
- when short vol, every adjustment/hedge eats away at your potential profit, so you want to make as few adjustments as you can get away with

when long vol then, transaction costs in dynamic hedging at the optimum frequency are a big killer, meaning you need a much larger edge in the first place to attempt it.
Transaction costs when short vol are less of a killer since we are trying to keep adjustments to a minimum anyway. However I tend to take a hybrid 'two bob each way' approach, adding the wings to say half the position early on, then loosely dynamically hedging the rest when and if required.


RV/stat vol would be calculated and analyzed using high frequency data to overcome the sample limitations of daily data w.r.t to vol (greater sample size -> central limit theorem for parameters). iirc correctly we discussed using an alternative rv measure (Parkinsons?) in another thread if you want to work exclusively with daily data.
i dont like to work with daily data much at all, apart from being the base unit in a spreadsheet. I find the idea of using a small recent sample of daily fluctuations as a basis for the distribution of possible future returns over some longer period to be somewhat silly. It contain 2 flaws which compound each other;
1. in real life the SD of returns over a period longer than 1 day does not match the SD implied by using one day returns multiplied by the sqrt of time
2. the distribution of a small recent sample does not (usually) faily reflect the distribution of all possible outcomes

No method provides a definitive probability distribution, but I prefer to use the distribution of x day returns over a larger sample (i use 10 years) as a starting point, influenced upwards or downwards by recent HV. daily data HV also is useful as an indication of how many hedge adjustments you can expect, to either have to make, or have the opporunity to make, depending on which side youre on.

But i would be interested in hearing more about how the 'using high frequency data' method works
 
So you work in a bank in quant analysis or something like?
Yeah, in the past I was a quant at an IB

- when long vol, you have to adjust/hedge frequently, as that is where your profit is coming from most of the time. failure to take a profitable hedge (round trip) is money lost forever and will see time decay (and possibly vega) eat away at you
- when short vol, every adjustment/hedge eats away at your potential profit, so you want to make as few adjustments as you can get away with

when long vol then, transaction costs in dynamic hedging at the optimum frequency are a big killer, meaning you need a much larger edge in the first place to attempt it.
Transaction costs when short vol are less of a killer since we are trying to keep adjustments to a minimum anyway.
Are you implying there is edge to short gamma side of a position?

What you say is more a truism for a choppy market. In a strongly trending market, you'd rather the intervals between hedges be larger [reducing frequency] for long vol and hedge more frequently for short vol.

No method provides a definitive probability distribution, but I prefer to use the distribution of x day returns over a larger sample (i use 10 years) as a starting point, influenced upwards or downwards by recent HV. daily data HV also is useful as an indication of how many hedge adjustments you can expect, to either have to make, or have the opportunity to make, depending on which side you're on.
What variables distribution are you trying to define here?
I ask because there is a difference between the distribution of asset returns and the sampling distribution of volatility.

Going by what you've got above, you seem to be incorporating a lot of redundant data. Basically all institutions run intra-day data for volatility analysis e.g. for 20 days realized vol, by using tick data, so the sample size increases significantly without changing the calendar time as you have done [10 years].

Question boils down to how you are determining the adjustment factor [rhetorical].

lol, sorry about all the bs. Let us know if you want it to move to another thread, since its moving away from systems trading
 
haha , nothing like having a debate on a public forum with someone more knowledgeable than one's self to expose and focus one's thinking.

Are you implying there is edge to short gamma side of a position?
I am not implying there is any edge to either side of options per se. whether there is or not depends on the price of the particular option in question.
I havent done much with currency options, but from my brief look at them there was no edge anywhere that i could see, so no I am not suggesting any edge to either side. I have never really looked at commodities/FI etc so cant comment but assume it would be the same there
In equity space; equities, paticularly equity indices , and even more particularly the SPX, the mean reverting bias outweighs the fat tails from an EV point of view, and options on these seem to be overpriced more often than they are underpriced.
Then in these cases yes I am implying there is often an edge to short side, although that clearly isnt always the case. My comments in the previous post were in reference to equity/index options only, which I should have made clear.

if there was no edge you couldnt have a picture like this; underlying is SPX
SellVolPerf.jpg

What you say is more a truism for a choppy market. In a strongly trending market, you'd rather the intervals between hedges be larger [reducing frequency] for long vol and hedge more frequently for short vol.

Agree. Again I was thinking of equity/indices which imo are choppy and mean reverting enough to favour hedging less if short vol. In above graph unhedged short vol strategy beats delta hedged short vol strategy.


also agree; apologies for hijacking the thread, we can move it to the derivatives forum if others prefer.
 
What variables distribution are you trying to define here?
I ask because there is a difference between the distribution of asset returns and the sampling distribution of volatility.

with this anaylsis; distribution of asset returns. I acknowledge that this is looking at the end result rather than the path taken to get there, so it is looking at only half the story. I do look at the path to get there as well, using data from daily returns. I see what you are saying in that that method of analysis is not useful in detremining the volatility one might experience between opening and expiry, which is more relevant to a discussion on hedging
now I am going to have to go and read up on sampling distribution of volatility. :)

Going by what you've got above, you seem to be incorporating a lot of redundant data. Basically all institutions run intra-day data for volatility analysis e.g. for 20 days realized vol, by using tick data, so the sample size increases significantly without changing the calendar time as you have done [10 years].

I did not know that. But if its a quieter 20 days than normal isnt the RV using tick data still going to be lower than 'normal'?

Question boils down to how you are determining the adjustment factor [rhetorical].
I dont know to determinine it quantitively, I just know it needs to be (and is) allowed for somehow, i am curious how the pros do it which is why I was asking you.:)
 
In equity space; equities, particularly equity indices and even more particularly the SPX, the mean reverting bias outweighs the fat tails from an EV point of view, and options on these seem to be overpriced more often than they are underpriced.
Are you saying that when calculating expected value, the effect of fat tails is offset and more by the tendency of the index to mean revert? This is how I am reading that statement, and if that is the case why worry about black swan events?

Either way vanilla options are priced with risk neutral valuation. To incorporate jumps in price, empirically there is a risk premium that exists for all options (iv > hv, its about 400 basis points on index) regardless of market e.g. currencies, interest rates, commodities etc.

More often than not there is a reason for the risk premium. If the ops are overpriced, why does a volatility skew exist [in indices]?

if there was no edge you couldnt have a picture like this; underlying is SPX
The draw down in 2008 is quite large, and I'd imagine if using a geometric position sizing model, it would be difficult to recover from those losses.

I didn't see what the Sharpe ratio is for those strategies? I'd imagine that it would be pretty low and wouldn't be justifiable for any of the big boys to run a book based on this strategy.

Studies have shown a random entry for stocks [long only] have positive expectancy over the long run. Edge?

But if its a quieter 20 days than normal isnt the RV using tick data still going to be lower than 'normal'?

No, I think you are confusing long run variance with population variance for n days.

If you are asked what is the "true" variance for the period spanning dates 1/1 - 20/1?
You can take the 20 days returns (sample = 20 observations) and calculate by dividing the sum of squared returns from the mean, divided by n-1 degrees of freedom to obtain a variance measure.

Generally by law of large numbers the greater the sample size, the closer the parameter is to the statistic being estimated, so using tick data, from 1/1 - 20/1 the sample can contain >100 observations rather than 20 as before.

Once this variance is calculated - then the process to contextualize it historically begins. [regime changes (macro-events, earnings etc.), current risk premium levels, forward vol as function of stat vol e.g. what percentile is it vs. the 10 year variance etc]
 
I am enjoying this discussion. Please treat my responses as answers to a test rather than trying to teach granny to suck eggs.:)

Are you saying that when calculating expected value, the effect of fat tails is offset and more by the tendency of the index to mean revert? This is how I am reading that statement, and if that is the case why worry about black swan events?
yes, that is what I am saying, if you calculate the probabilities using the distribution from back data. however back data by definition doesnt account for the possibility of all jumps or black swan events. As I have said before I acknowledge the need to make some allowance for them. Why worry about them? because even if a bet was known to be +EV it doesnt mean the worst possible outcome wont happen more than it 'should' over a small number of trials?

Either way vanilla options are priced with risk neutral valuation. To incorporate jumps in price, empirically there is a risk premium that exists for all options (iv > hv, its about 400 basis points on index) regardless of market e.g. currencies, interest rates, commodities etc.
that would be the offset or 'adjustment factor' we have discussed in previous posts? but how do we know that risk premium is 'correct', since we cant really know the probs or impact of all black swan events? since the risk premium the market puts on it is a bit of a guess , it is quite possible it is under or overstated.

More often than not there is a reason for the risk premium. If the ops are overpriced, why does a volatility skew exist [in indices]?

It exists because the fat tails to the downside have historically meant that OTM puts have 'paid off' far more than would be priced in by using HV or ATM IV. However I am not seeing that that proves anything about whether ATM options are over or under or fairly priced.

I didn't see what the Sharpe ratio is for those strategies? I'd imagine that it would be pretty low and wouldn't be justifiable for any of the big boys to run a book based on this strategy.

yes, that chart would have been more useful if it had included some sort of benchmark to compare it against, such as risk free rate or the underlying. There were no sharpe ratios included in the article i got it from and i am not familiar enough with the information ratio to know where IR=1.1 puts it?

Regardless of whether it could be traded as a system with +EV, would it not be fair to say that if options were always fairly / risk neutral priced then the net result of such a strategy should be random swings ending up at zero over a long term, not an upward sloping equity line? (I will concede here that perhaps 1996 to 2009 is not a long enough period to draw a conclusion, and that it is possible that the positive equity from such a strategy to date is in fact merely advance compensation for the risk of some black swan event that didnt happen yet ).
Or have I got it wrong and it is supposed to be that a short vol strategy 'should' make the risk free rate ( and by corrollary a long vol strat would expect to 'pay' it.?

Furthermore, it is my understanding that the whole BS model is based on the principle that the value of an option should be equivalent to the amount gained or lost by delta hedging (ignoring costs) if RV turned out to be equal to IV paid. If that was in fact true , then the result of an unhedged ATM straddle strategy should be the same as the delta hedged strategy (long or short) ? Yet in the graph I posted that has not been the case, there is divergence between the strategies, which suggests there must be an edge somewhere. If we took the delta hedged equity line as being 'par' or zero EV (just equiv to risk free rate?), then unhedged must have an edge over it at least?

Studies have shown a random entry for stocks [long only] have positive expectancy over the long run. Edge?

i am guessing that the positve expectancy would be equiv to either inflation or a risk free rate or somewhere round there? If we defined an edge as being positive expectancy over and above the risk free rate then no, no edge there. if it was above that, then, well; positive expectancy=edge, is there any other way to see it?


No, I think you are confusing long run variance with population variance for n days.

quite possibly, I am not an expert on stats. i will take your word for it that using tick data is better, but I dont suppose I will ever be in a position get to use it
 
Arrite guys, I have another question. (sorry to interrupt ur superdiscussion)

So we basically have a logical idea we want to trade, back test it, then test it on the out-of-sample data and if its profitable we trade it right.

Isnt this sort of curve fitting? If we do this procedure enough times wont we come across a system which is profitable in the OOS data just by pure chance?

How can we trust our system to have a positive expectation?
 
Hi GoPoncho --

You are correct. There is always a bias. Even systems designed and validated using walk forward out-of-sample testing can appear to be valid, when they are not, due to having examined so many that some will eventually pass the test.

What is the alternative? Using less stringent testing procedures will allow a higher percentage of invalid systems to move from development to trading.

I recommend continuously monitoring the health of every system, applying statistical tests to give a metric of system health, taking any system that appears to be performing poorly offline, and waiting for its performance to return to an acceptable level.

The risk of just watching a good system is lost opportunity. The risk of trading a broken system is lost money.

My newest book, Modeling Trading System Performance, addresses exactly this issue.

Thanks for listening,
Howard
 
Hi Howard,

Thanks for the response.
I now see that creating a system around a trading idea is a tradeoff as you explained. As only a certain frequency of truly profitable systems should be expected it makes sense to have active monitoring.

I think I'm still not sure about where to look for the trading idea itself, but I'm slowly trudging through information :) Getting there; very, very slowly.:)

Reading your Introduction to Amibroker at the moment, and will certainly read Modeling Trading System Performance later down the track.

Thanks!
 
Top