Round Table -- Frank Fahey - The Greeks

Discussion in 'Round Table Presentations' started by Luke, Jul 19, 2017.

  1. Luke

    Luke Well-Known Member

    What did you all get out of today's presentation? I'll make time to watch it again when it's uploaded, but the point that stood out to me was to really take the time and effort to watch one trade and see what the greeks did to help or hurt each component. So, I'm assuming that it would be helpful for me to take apart each position in spreads or singles to monitor how the greeks changed both individually and how that affected the position as a whole.

    Frank mentioned that each day he knows why his position was affected - delta, Vega, gamma, or changes to theta and what it did to the P&L and shape of his curve. I look at the snapshot and risk graph (and project out 7 days) but I haven't broken down each individual option or even the spreads to understand which side got hit and why.

    The continuous theme I see with the guys that are successful is that they are continuously students of the game and documenting each aspect of their trades. My daily screenshots look like K5 stuff and these guys are building PhD dissertations...
  2. PK

    PK Well-Known Member

    When I started options trading I was taught that greeks were for freekies. When I started losing money without knowing why I became interested again in the greeks. It was this painful experience that made me "sense" gama and vega risk in a different way. Once you observe your positions with a reasonable risk graph (OptionVue/ONE) and/or do manual backtesting with this software, it will take you only a few hundreds of trades to "sense & feel" the greeks and, more importantly, get a grasp on how they affect your PnL. For me it was a good lesson to build an Excel sheet with my positions, the corresponding greeks and the PnL. And when the market opened and I got some stable values for price movements and volatility, I returned to the spread sheet and entered my own "predicted" values for the option value and the greeks. And then I opened the options trading platform to figure out the real numbers. After a few years doing that, I developed a good sense for how a complex portfolio's risk graph morphs in response to market conditions. The effort to develop a gut feeling for how market moves affected my positions and whether this could be remediated by manipulating delta, gama or vega paid off in the long run.
    Luke likes this.
  3. DavidF

    DavidF Well-Known Member

    Have watched all of Frank Fahey´s available material and learned a lot from him. On a practical level the most important thing I observed/learned was that even if he´s linked up to OV he uses EIOIO, not the advised variable model. In fact as you may have heard him say, he´s not in the slightest interested in the values provided by the var. model.

    The inaccuracy of the var. model caused me a lot of stress in the beginning, as I didn´t trust my own empirical observations. If the whole underpinning of his talk was knowing your greeks, then this issue (even if it´s been discussed to death) is so fundamental.

    For example say I purchase OV software, am told "problem with EIOIO is it assumes IV remains static etc etc" and start trading iron butterflies on SPX using the var. model thinking it´s more accurate. I set up a 20-lot 2340/2440/2540 Oct fly last night. Var model says I need 6 calls to hedge the neg. 600 deltas (neg. 603). EIOIO says I only need 3 ( neg. 318). If I click "Combine calls and puts" on var. model I need 4 calls (neg 418, the main advantage of this feature is that it reduces the inaccuracy of the var. model). Those differences are wild.

    So heads up to any new SPX traders who are trading iron flies, iron BWBs, or even 60-40-20 with OV, be very consistent in your backesting model and for new trades go through the greeks using the same strikes in puts only and check EIOIO vs. var. to get a ballpark figure of where your greeks are (EIOIO benchmark), otherwise you could be way off. I ended up going over to OTM BWB put structures are I ended up only trusting price.
    Last edited: Jul 20, 2017
    Luke likes this.
  4. Kevin Lee

    Kevin Lee Well-Known Member

    Greeks are indeed important. But the issue is that if the IVs are wrong, greeks are wrong too because greeks are derived from IVs. None of the software I'm familiar with are total accurate with the greeks, not OV, not ONE, not TOS and not IB. The main reasons for the inaccuracies are :

    1. Firstly, the software reads in market price data. From that, the software calculates IV. Now, here's the issue. To calculate IV, the BS formula takes in dividend rate and risk free rate as inputs but there isn't a single source of truth. Each software does its own thing. So we end up with IVs that don't agree among the different softwares.

    Not only that, within the same software, the Put and Call IV of the same strike can be quite far apart, clearly violating the Put Call parity. See SPX IVs in TOS for example.

    The IV skew graphs are clearly wrong as well. You get weird IV curve that looks like this


    So.... when the Put/Call IVs are not in-synch, we need to use "combined Put / Call skew". That's basically a short cut to take the average of Put and Call IVs of every single strike. The end result is a single number for both Put and Call. Not totally accurate but it is often close enough.

    The proper way to ensure Put / Call IVs are in synch is to set the dividend and interest rate accurately. OV allows users to set both of them. How to get the accurate data is another very big topic. I have had quite a lot of conversation with OV about that already. Hopefully they can program the software to set it more accurately on a daily basis.

    2. After the put and call IVs are in synch, that is still not good enough. The second problem is EIOIO. In EIOIO mode, IVs are assumed to be static. That means when the software draws the T+0 line, it assumes IVs to be unchanged across different underlying prices. We know that never happens. IVs go up and down as underlying moves and IVs of individual strikes go up and down at different rate too, thus either steepening or flattening both the vertical and horizontal skew. Therefore, the greeks that are derived from EIOIO are unlikely to be accurate and the T+n lines drawn in EIOIO mode will likely be off too.

    So what does VAR mode do ? The VAR mode tries to predict how much IV will move up or down as the underlying price moves. That can be done by using statistical analysis a to estimate the correlation between IV and Underlying price movements. OV's VAR model tries to do that for both within the same expiration as well as across different expiration. But, OV still does not take into consideration the potential IV skew shift (ie steepening or flattening). And I think that is a big problem. Without incorporating skew shift, we'll forever end up with an unstable T+0 line. None of the analysis software out there I know of does that.

    So which is better? VAR or EIOIO ? Well... technically VAR should be better but that's actually not the right question to ask. Since none of them is totally accurate, experienced traders have learned how to mentally compensate for the inaccuracies. Therefore, it's not which one is more accurate, but rather which one have you learned to work with ? As long as you know how to compensate properly, then that would be the better model.

    Okay... I hope what I've said above isn't too confusing. My point is greeks CANNOT BE taken at face value. None of the software produces 100% accurate greeks. The biggest issue that will remain unresolved for sometime is the ability to model IV skew shift. I believe without that, there is no chance that the greeks and T+0 line can be completely accurate. My personal belief is that the solution to IV and IV skew modeling is with machine learning and AI. But I might be bias. Let's see where future leads us....

    Attached Files:

    Jason A, Luke, SVL and 1 other person like this.
  5. DavidF

    DavidF Well-Known Member

    Hi Kevin, I know you know your stuff and nothing I theoretically disagree with in what you write above.

    However, what I´m referring to is wanting to get a rough estimate of my position delta. If it was a subtle difference, e.g., 325 in EIOIO and 390 in variable I´d be fine and learn to work with it. But 300 vs 500-600 is not workable. No matter how good variable is at predicting changes in greeks if your starting deltas are vastly divergent, you´ve zero idea what your position risk is. Happened to me in real trading and happens in backtesting. Another anomaly:- although I haven´t traded RUT in a long time I found the issue less relevant in RUT. So could be a specific SPX call issue in variable mode and not the theoretical models per se.
    Last edited: Jul 20, 2017
  6. Paul Demers

    Paul Demers Well-Known Member

    Great post Kevin

    I gave up searching for the holy grail of accurately defining risk with the IV and greeks mostly due to watching this:

    If these Noble prize winning mathematicians could not develop models that accurately assessed their risk who am I to think I can.
    SVL likes this.
  7. Kevin Lee

    Kevin Lee Well-Known Member


    My post wasn't written in response to what you said earlier. My intention is to point out the fact that none of the models are accurate and share my thoughts on why that's so.
  8. PK

    PK Well-Known Member

    Maybe that my reply is not up to date, but my very personal conclusion is that reality and real greeks amd fair values are not relevant for trading and that knowing the "real" delta of your put is irrelevant. I remember Jim's story of becoming used to drive a bike with the handle bars turned down. At the end it it does not matter what one may perceive as as reality but rarther how you interpret and react to changes in your environment. And options trading is possibly one of the most instructve cases of converting the gap between expectations and reality into benefits for those who consistently detect this gap.
    SVL likes this.
  9. DavidF

    DavidF Well-Known Member

    Just backtested the last decent spike in IV on May 17th setting up an iron fly on SPX on the 16th. Variable model tells me I´m notionally short $86k (minus 36 deltas) at the current strike and that a drop the next day to 2355 is fine. However the EIOIO model says I´m notionally long $260,000 (plus 110 deltas) dollars and that a drop to 2355 the next day will put me down approx. $8,000.

    24hrs later SPX hits 2357 and I´m $8000 down. Am not cherry picking,you can test mutliple down moves and you´ll find that (in general, haven´t tested every day) the variable model is overly generous to the downside with iron flys. Have done this years ago and came to same conclusion (and posted findings on here). I understand no model is perfect but I´ll go with the one that gives me a much better idea of my downside risk.
  10. Marcas

    Marcas Well-Known Member

    This is so important topic.
    I tried to ignore it for way too long hoping that greeks provided to me by broker are sufficient. It was experience very similar to Peter's (PK), except I don't think I developed good sense for greeks yet, at least my sense is still not 'good enough'.

    While I'm in this business, trading, I see more and more clearly need for most accurate greeks I can afford. The reason is pretty much the same as given by DavidF. To be successful I need good greeks. How do I hedge effectively if I don't know my real delta? Profitability of my trades goes down if I do unnecessary adjustments. If I trade on unreal greeks I expose myself to higher risk than I wish to.
    Thus I strongly disagree with PK statement: “knowing the "real" delta of your put is irrelevant”. (I don't think he really meant it. 'Handle-bar' adjustment is a method to obtain better greeks anyway.)

    There are trades for which greeks are less important; like high risk-high reward, you can even trade them without looking at gereeks at all - just P/L, but for trades I trade, low risk-low profit good calcs are very important. True, I won't be hurt too much making bad move, but I won't earn anything either.

    One can trade based on 'greek signals', like trading M3 with OV (no personal experience). It is not that OV model is better, it is just that M3 was developed and tested on OV. If OV shows such and such number do this, if number change do smth else. Can you trade it on ONE? Sure you can, but whole translation of ONE numbers to specific moves needs to be redone. Signals from OV or ONE are not what I'm after. I want greeks that show me real risk of any trade.

    Paul said he gave up on holy grail, I think that my quest is about to begin. I have some ideas to test. Kevin is working on great thing, imo, using machine learning to get numbers close(r) to reality. Maybe he'll be successful finding right balance between sticky strike and sticky delta every time. Or can say with high precision what model adjustments to use in particular market conditions for specific trade. I wish!

    So the reason for this, a bit longer than planned, post is to invite some thoughts how to go about that. My first planned step will be to find a way to evaluate models, how accurate they are. I wouldn't like to do it via manual testing but rather automate it somehow. I know that many of CD members are working on theirs data, I hope to join them soon. How would you go about testing model that claims to better that one you have?

    PS. I didn't watch Round Table presentation yet.
    Luke likes this.
  11. Kevin Lee

    Kevin Lee Well-Known Member


    You are exactly right. You said the same thing as what I was trying to express.

    In general, trades of which the strikes are tucked well below the market, ie OTM puts, are more stable in their greeks and T+n lines. One would need less mental compensation of what they see on the analyzer. So, for example, BWB or BWC where all the legs are below market, typically will be more stable than a M3 trade where the right leg usually is above the market. The reason is the right leg IV crush which I have explained. Therefore, this is exactly why knowing the trade, the software platform and the modeling setting well is important. Otherwise, you won't know how to compensate for the greek's inaccuracy.

    In addition, knowing the root cause of the inaccuracy is important too. Because different situation will require different greek compensation. It's not as simple as "always add x deltas" but rather "in this situation, add x delta, but in that situation, minus y delta".
    Luke likes this.
  12. DavidF

    DavidF Well-Known Member

    Marcas, the only ´truth´is price, so simply backtesting what each model predicts P/L will be at certain fwd dates, and then seeing how the actually turned out is a good way.

    This topic is an old one:-

    My purpose wasn´t to rehash it (and have another conversation with myself), it was more in relation to Fahey´s comments on EIOIO vs veriable model (although I noticed a long time ago he used EIOIO which in turn gave me some belief that my earlier choice wasn´t mad).

    As you correctly observe there are multiple ways to mitigate the issue but, to use an analogy, not having accurate starting position delta and learning to compensate is like playing golf with a huge slice and compensating by aiming far left.
    Last edited: Jul 21, 2017
  13. Kevin Lee

    Kevin Lee Well-Known Member


    In general, for butterfly trades like M3, EIOIO's delta will be more positive than VAR. When market goes down, it depends on the IV behavior. When the IV curve either moves up vertically or moves up and steepen, the T+0 line will turn anti-clockwise, hence making delta more positive than what it was showing before. In this situation, EIOIO will show accurate prediction than VAR. But if market moves down and the IV curve moves up only gradually and the skew flattens (yes it does happen), VAR will be more accurate than EIOIO. The explanation for this is that when IV skew flattens, the IV of the right leg goes up more than the other legs.

    On the other hand, if market moves up, then EIOIO will be too positive for a M3 type of trade. I'm sure many M3 traders have experienced this - Delta is slightly positive, market moves up, IV drops. You expect P&L to increase - gain from both delta and vega. But P&L dropped.

    Before OV, I was trading with TOS + ONE. The only setting then was EIOIO. I suffered every time market moved up. Overtime, I knew EIOIO's delta was too positive that what it actually is. So, I tried to compensate. But the problem is the amount to compensate is dependent on many factors, including where the right leg is relative to market, DTE, IV skew, expected market move etc... With VAR, I still have to compensate on the upside because even in that mode, the delta is also too positive when market moves up, but the amount that needs to be compensated is more stable.

    The important point is this.... while delta is too positive when market moves up, but at the same time, the delta is too negative if the market decides to move down. Understanding this is key because the T+0 line will shape shift differently depending on what the market does. In VAR mode, the quantity to compensate I find is more stable than EIOIO.

    So again, what I'm trying to say is it's all about familiarity of the software modeling you are using. If you found a better way to adapt to EIOIO, then that is a better model for you. But the determination of which model is better isn't by comparing which one predicts more accurately, because you can find evidence for either model depending on situation. There will be specific situations when one model is better than the other. The key is which model is more "consistent" in how they are wrong and hence allow you to compensate for the inaccuracies more easily in most situations.
    Luke and SVL like this.
  14. DavidF

    DavidF Well-Known Member

    Thanks for your input Kevin.

    Your point is well taken and if I was trading RUT I´d agree 100%. I just checked and backtested some M3s and irons on RUT. As I found before, the differences between EIOIO and varable in RUT are well within an acceptable range that make assessing your immediate notional risk workable. Doesn´t matter if I click combined call/puts, the range of strikes on calls etc, the values are within ballpark of one another. Furthermore there´s thousands of hours of backtesting and validation of var. model with combined put/call box checked so you know you´re not far off.

    However it´s a totally different scenario on SPX, the values are not within reason and the consequences are not subtle. I´d contend that not knowing whether you need 3 of 6 calls to hedge a $55k max loss iron fly is not simply the vagaries of various models that you can work around. Something is way off. My situation was the reverse of yours, started with the OV advised settings and swtiched to EIOIO , but for same reaons.

    Again I´m not convinced it´s the theoretical models per se, I think it´s something else, e.g., a glitch in the SPX call skew curve that OV var. model generates that throws everything off (otherwise changing strike range wouldn´t be so important).
  15. Kevin Lee

    Kevin Lee Well-Known Member

    100% agree, Paul. I too went down the road of trying to learn advance math to understand financial modeling and hopefully be able to gain an edge. After couple years, I realized that was a dead end. Greeks and IVs are useful theoretical constructs, but they are limited and the modeling isn't very accurate. Finance really isn't physics and using the same type of theoretical modeling in hope of the holy grail is difficult.

    I follow Emanuel Derman's work. He's one of the most famous quants and he's also saying the trend now is moving away from trying to push for more accuracy in theoretical modeling. Here's a quick interview of him expressing that view

    I think that we might be able see some light in empirical methods and machine learning. That's the direction I'm going now. I'll likely know if that's another dead end or not hopefully in a couple years.
  16. Paul Demers

    Paul Demers Well-Known Member

    My thinking is that the Citadels and the Goldman's of the world have much more money to direct to AI development and that by the time the rest of us get a handle on it the edge will be gone.
  17. Kevin Lee

    Kevin Lee Well-Known Member

    In general that is true. I agree. But that doesn't mean we can't make use of AI in niches that are too small for them to bother, ie like how we are surviving today. One practical example would be using artificial neural networks to do a better job forecasting IV given the underlying movement. That value will be immediately apparent to our current trades.
    Paul Demers likes this.
  18. SVL

    SVL Well-Known Member

    Let’s imagine that the magical risk-free model is found based on AI / UFO/ etc..
    Then, someone like Citadel or GS could literally put billions of dollars at risk as ” there is no risk” and lever up like LTCM guys did ( 3 billion USD Hedge Fund got 1 trillion USD exposure).
    Once you get that big, there is no one left to be your counterpart in the times of panic. The volatility spikes, the bid /ask spread gets very wide and you can not get filled even at the bid . Your broker steps in and liquidates your position at whatever they could get. The end of the story.
    No financial model is capable to predict that and that’s why Taleb’s hedge fund made 1.0 billion dollar profit in August 2015.
    Last edited: Jul 21, 2017
    Kevin CK likes this.
  19. DavidF

    DavidF Well-Known Member

    So I just spent a lot of time comparing EIOIO vs. VAR model in SPX in 3 historical environments:- one-day volatility spike (May 17th) , collapsing IV with rapid move up (post election), and low IV grind up (recently).

    I used iron-flys on SPX, 75 point wings, 50-80 DTEs with shorts approx. 1% under the money (combined box unticked, large strike range).

    The first part involved calculating postion delta as if greeks were static. So I compared the price of a fly 25 points above and below the selected fly, then extrapolating the difference in price to calculate approx. position delta (e.g., if fly price changes by $7500 on a 1% move it´s approx. 300 deltas with a low-gamma set-up). Variable model carries a lot more neg. deltas than EIOIO in SPX iron flys. To my surprise VAR model was much more precise here, opposite of what I expected.

    The 2nd part involved actual changes in fly price versus predicted changes via modelling. To cut a long story short it came to what Kevin writes, it´s unpredictable. In a collpasing IV environment the VAR model was far too harsh on the upside. In a low. vol. grinding up environment EIOIO wasn´t harsh enough. In a spike VAR wasn´t harsh enough on the downside. But even then there was variability. And to add to that every IV spike/crush is unique so becomes a minefield of unpredictabiliity.

    So to some degree you choose your poison. I realise my dogma & bias towards EIOIO was based on focus on not getting hit on downside with a vol.spike (real-iife experience). However, if we ignore the extremes my conclusion is that the VAR model is more accurate for 1-2% changes in price without major changes in skew. So Kevin, I think you´re correct, no caveats.
    Last edited: Jul 22, 2017
    Luke, SVL and Kevin Lee like this.
  20. Marcas

    Marcas Well-Known Member

    DavidF – good job with your test. That's how you validate your models. Unfortunately your work is useful only to OV users.

    Absolutely right. It That is why I'm interested in developing some kind of method to examine performance of models. The best (well, the easiest for me anyway) would be to work with my own model variants. Goal is to see if existing market conditions combined with trade type can direct us into preferring model A vs model B. It is very interesting and I think it is doable, I mean creating tests. My current tasks take all my time, but hopefully in few months I could be able start typing some code in this direction.

    Also thanks for video link to E.Derman. I didn't know this fellow before. I watched 'Models Behaving Badly” - he has some interesting points (and some I'd dare to question – flags were risen : ). There he states, what is surprising to me, that 'industry' do very little of model testing. Hmmmm…

    Better greeks are desirable. I don't care much how this is achieved; by improving models, by machine learning, by reading tee leafs. As we have only observation to validate results standardized test are needed (badly) or we are doomed to do anecdotal tests like DavidF did, like you did, I did and many others did and many will be doing.

    I guess what you mean is risk free trade. 'Ideal' model will not tell you what will happen with market, it will tell you where your risk is and let you take appropriate actions according to your risk appetite. What happened to, mentioned also by Paul, LTCM, imo, was that they were using their model for trading instead of using brains - opposite to Taleb obviously. Derman said about such situations :'catastrophes strikes when hubris evolves into idolatry”. (Some degree of hubris is indeed needed in situation we are with our models.) And don't be intimidated by Nobel Prices. In news papers you can read another example of nobelist who is so attached to models that he lost common sense. Not the rule, but it happens, especially in economy-finance territory. That failure, LTCM, doesn't mean that we should abandon modeling, we should abandon idolatry.

    So I have another point in my schedule: produce a test to validate options models. I don't quite expect to come up with perfect solution as I have little experience, I will try though. Outcome be, hopefully, a better understanding of 'the root cause of the inaccuracy' and improved P/L.

    Any discussions on this topic, now or in future, are warmly welcome.
    Ice101781, DavidF and Kevin Lee like this.

Share This Page

  1. This site uses cookies to help personalise content, tailor your experience and to keep you logged in if you register.
    By continuing to use this site, you are consenting to our use of cookies.
    Dismiss Notice