More in particular, it's the positive "resolution" itself of the "hedging orders" (in practice, those which would appear like "stops" to an external observer) which provides the possible edge for the whole algorithmic work. That is, the source itself of systematic profitability (in statistical terms, clearly). Look for instance at the picture below. This is the typical result of a simulation of these ideas. Let's see what's in it. You see 4 lines. Those are the PNL and its 3 components: Green (dotted): Realized Gain Cyan: PNL Light cyan dotted: Unrealized Red: (temporarily realized) Losses <img src="http://www.elitetrader.com/vb/attachment.php?s=&postid=3897661" /> Now, due to the "temporary" character of (a portion of) the "losses", which have however a <b>crucial hedging value and function</b> in this context, many of the past losses are continuously converted into realized gains, when the prices passes multiple times over the same levels. This, in the long term, causes an unbalance in favor of the realized gain, and therefore a "drift" is impressed to the PNL. In time, the realized Gains and Losses will both become enormous (with hopefully the gains exceeding the losses ) and the unrealized (current "investment" in new moves and volatility) will have less and less power to affect the PNL. Both the PNL and the unrealized become relatively unimportant in dynamic "tug of war" between the realized gains and the losses. In this regard, a metric which could perhaps be useful to express this "drift" could be the relative difference of slopes (PNL components are considered with their respective signs): <pre> Gain Loss ----- + ----- Time Time Relative Drift = ------------------ * 100 | Loss | ---- Time Gain + Loss which we can simplify as: RD = ----------- * 100 | Loss | </pre> This, imagining a 50% chance for any order to capture a given positive scalp (there would be no reason to assume otherwise), could also be interpreted as a "Loss recovery ratio" expressing the percentage of "temporary" losses which have been resolved into profits through the "player superposition" mechanism, that is the fraction of the 50% losses which have been "converted" into profits by storing the relative information: "Loss recovery ratio" = 0.5 * RD In very practical terms, we have "apparent" stops which will be turned (in a hopefully good proportion) into (delayed) profits (on retracement). What is crucial is (apart notable exceptions) <b>not to allow</b> the exposure of each instrument or set of correlated instruments to go too far, and rather early and promptly hedge against it (following appropriate rules). And a good frequency certainly helps in this regard. The <b>player superposition</b> mechanism will automatically take care of orderly closing most of the players at the due time (each one with its due profit), thus throwing a good % of the "apparent losses" into the realized gain component. Then let patiently the time do the rest. No prediction needed.
Hello there, I think I understand the concept of the realized gains going faster than the losses and (hopefully) fast enough to also pull the unrealized losses. Over time, with enough instruments, enough diversification and enough capital, what you call the positive drift will have taken so much advance that the unrealized losses can not be a danger anymore for the account. Normally, that is⦠The concept is intellectually very tempting, even more when I consider that your system is non-predictive and that there is no need to âbe rightâ on the market direction. And even, even more since most of the time markets donât move in a specific direction. But what if the unrealized losses happen at the beginning of the operation (maybe due to a runaway in the wrong direction, and maybe with a wrong diversification that gives too much weight to one particular market going wrong)? In this case, the account may not be ârich enoughâ and the âinvestmentâ might pull the realized gains low, so very low that the account will break⦠Did I understand right and is this statement correct? While the concept is tempting and makes sense intellectually, does it work in the âreal worldâ? How many accounts are capitalized enough to take the chance? Of course, the trader does not have to shut his brain off and can still observe what is happening and eventually cut the instrument running away, or deciding to enter a hedging trade. As you mention, it is important ânot to allow the exposure to go too farâ. But it seems to me that it speaks a bit against the interest of trading fully automatically. Now the 2nd part of my comment: I think it is (or should be, at least!) a basic of sound trading to not take losses. But waiting that the market reverses can take time⦠and during this time, unrealized will still pull PNL down. So how about, instead of not taking losses, just âcompensating themâ with a specific hedging system? As far as I can see, the motor of your system is the balancing of the folio as to always reduce the risk while the system itself gives to the trading a positive drift. But what if a 2nd motor was to always trade UNDER the risk of collapse? I explain: At any moment, we know the âinvestmentâ, or what I would also call ânegative exposureâ. Therefore, at any moment, it is possible to check whether we still have enough money in the account to counter this exposure by a hedging trade. The last available money on the account should (in this case) always be used for this hedging trade. I posted in attachment a basic drawing that shows there are 2 solutions only: Either the countertrade (hedge) reaches a value of 100 and compensates the âinvestmentâ (open exposure), which I can then cancel (so as to not pull my PNL down more) or the prices reverse before they reach the lower level, and go back to a level where the scalping can take place UNDER the critical level of negative exposure. Then the âhedge tradeâ can be canceled at a value of 0 and the open exposure can maybe even turn back to a profit. Doing so may even allow to always scalp with a strong leverage, and therefore deliver more profit. And the trading action can only be in the following situation: case 1: Either the system scalp normally and the PNL increases case 2: Either the system puts scalp on hold while the open exposure will be compensated then cancelled, then the system can go back to normal case 1 operation case 3: Either the market reverse before case 1 is completed and normal operation can restart like in case 1. Then the hedging trade is cancelled. In any case, the drawdown is limited. I would be interested to have your opinions about this. Thank you!
> I would be interested to have your opinions about this. >But what if the unrealized losses happen at the beginning of the operation (maybe due to a runaway in the wrong direction, and maybe with a wrong diversification that gives too much weight to one particular market going wrong)? In this case, the account may not be ârich enoughâ and the âinvestmentâ might pull the realized gains low, so very low that the account will break Thank you Johnny Ca$h for your articulated and insightful question and suggestions. I hope I can address most of your points. My opinion is that you are completely correct about the idea that in order to trade one needs to be decently capitalized. And there are definitely technical reasons why under a certain threshold one would be better off staying away from the markets. Unfortunately, often people have little resources but still want to somehow dream or hope in some kind of luck. They are almost all wiped out pretty soon. Not even luck may be enough, as your opponent is carefully programmed (by people like me) to rip off the greatest number of participants. As B Lee once famously said, "boards don't fight back". And this is something that can also be applied to people dreaming after some curve fitted useless backtests. The mkt does fight fiercely back and it's a rather efficient machine in doing that. One definitely needs to be equipped with superior resources to pick this fight. Or else it is pretty much like going plain naked and unarmed to fight against a division of armored trucks controlled by smart satellites I am most grateful to this investor which has courageously put at disposal his own money and machines for this new testing round, and has given me the greatest incentive and responsibility to go deeply into this new line of logic of reasoning, much more in sintony with what most investors an real traders want. > But it seems to me that it speaks a bit against the interest of trading fully automatically. Well, every action one takes can always be implemented into an automated procedure. And, once done, this certainly allows a more careful assessment of the statistical performances. When you find yourself repeating the same action again and again and you see it may be helpful, it's definitely time to implement it and go through a deeper and due statistical analysis, to tune it at best or dismiss it if it works out to be an illusion. I usually follow this "bottom-up" approach, where I implement only what is strictly coming from experience and a real and strong need. Never try to go top down in this case: most of the time one will never have enough prior knowledge to do that. >Now the 2nd part of my comment: I think it is (or should be, at least!) a basic of sound trading to not take losses. But waiting that the market reverses can take time⦠and during this time, unrealized will still pull PNL down. Well, we are actually taking ("memory-full") losses here in a massive form (as you can see from the loss curve). The real crucial point is that all entry points (in addition to being organized in a certain scheme, ensuring "internal coherence" of the order cloud) are efficiently kept in memory, and possibly unwind and "resolved" in the future if there is the possibility (without re-opening new positions). So, some losses (with their hedging function) will remain in perpetuity until the price passes again on those entry points. In the meantime we are using the margins freed by taking these temporary losses to make new scalps and diversify. Using the ("memory-less") stop approach, as many people naively do, informed by misleading internet literature, there is absolutely no way to survive in the long term. It's a simple statistical fact that takes little thought to understand. Furthermore, the MMs will be working actively against most of the positions and efficiently screw up all those not adequately equipped to efficiently postpone their gains and do their mathematics and strategic game. > Then the âhedge tradeâ can be canceled at a value of 0 and the open exposure can maybe even turn back to a profit. That's right. Just as you indicate in your picture, the system has now built in an adjustable "exposure control" (I will be refining it in the future updates) which kicks with forced countetrades to "trap" the PNL within given bounds (see screenshot). In addition to that (at single layer level) there is also a mechanism working "across layers", to take into account the global exposure of groups of strongly correlated instrument. [In fact each block of instruments going in the same (or specularly inverse) direction can be seen actually as a "unique instrument".] Many of these are all new features that I hve incorporated in the next update of my application (shipping soon btw). However, it also allows the freedom to override the setting when appropriate, as actual experience shows that for some specific instruments most of the time you can even do well without. It also allows to put layers on hold as you indicate, and many many more possibilities. Simulations shows that too strict continuous hedging up and down may make the desired PNL drift so slow that in practice the commissions may even grow faster than your expected profit. It's a game which has to be carefully balanced, and this can actually be done only resorting to carefully implemented simulation methods: there are no shortcuts. Like a little bird in your hand, you don't want to hold it too loosely or it will definitely fly away, nor too tight or you will kill him <img src="http://www.elitetrader.com/vb/attachment.php?s=&postid=3898025" />
This is an example picture (from a simulation study) showing the forced-hedging-entry mechanism I mentioned in the previous post at work and also envisioned by Johnny. Here, for instance, I have programmed it to "kick in" at a $ 200 drawdown (DD) (and other additional directional conditions about the tickdata microstructure to optimize the entries). (By DD I mean, as usual, the maximum excursion of the PNL from a local maximum.) The hedging entries are those T players (red squares for T Sell) with a yellow frame (the big green arrow is pointing at them). <img src="http://www.elitetrader.com/vb/attachment.php?s=&postid=3898136" /> Looks pretty good, this should contribute decently to keep automatically under check the exposure of a single layer.
In the previous posts we have discussed how to possibly measure the <b>relative drift</b> (<b>RD</b>) caused by recycling the "past losses" into the realized gain component. We mentioned we could use the following: <b>RD := "Relative Drift" := (Gain + Loss ) / | Loss | * 100</b> (where Gain and Loss are taken with their respective signs.) This metric tells however only "one side" of the story. In fact, we have said that the algorithmic procedure can be seen as a "tug of war" between 2 forces: - massive scalping (contributing to "gain" when it realizes positively) VS - "losses" (due to "hedging") + Unrealized (the residual part) The RD metric seen above quantifies the intensity of the positive scalping activity. In order to evaluate the "hedging force" of our procedure we need another metric. I would propose the following one, which can be interpreted as "<b>hedging intensity</b>": <b>HI := "Hedging Intensity" := (Unrealized + Gain) / | Loss | * 100</b> In order to have a good outcome it is necessary to keep under check these 2 values. In fact, it is not enough to have a good "drift", if the "hedging intensity" is insufficient, as the effect of the commissions and the unfavorable volatility could render very slow or totally nullify the action of any existing positive drift. Through some simulations, I have determined that "good" values of these metrics might be as follow (I might revise these values later, after more research): HI >= 110 % RD >= 10-15% (the higher the better, provided the HI threshold is satisfied) This provides a simple but very effective method to supervise and evaluate our algorithmic procedure and make sure the 2 opposite forces are operating in "optimal balance". In addition to these "relative" metrics, we may also be interested in some "absolute" quantity, as for instance the "overall" or "daily" net gain. <img src="http://www.elitetrader.com/vb/attachment.php?s=&postid=3900157" /> Equipped with these metrics we can do a much better job supervising and possibly rebalance the outcome of a trading session, and more clearly see where we are leading in the long term. (I will be adding these metrics in the next update of the application).
After some "theory", let's switch back to "practice" and see how it is going. The current (real $$$) situation is as shown in the picture below. <img src="http://www.elitetrader.com/vb/attachment.php?s=&postid=3902435" /> In particular at this very instant we have about 17K of gain and about 11K of losses. So the we have been creating about 6K of positive exceedance in about a month, for a daily (including holidays) net gain of about $ 170. The <b>relative drift</b> is 51% and the <b>hedging intensity</b> about 81%. So a little unbalanced: we need to increase a good bit the hedging action. At this time the investor has doubled the capital in the account, but I am still using only a fraction. About 80K margins. Commissions are heavy: we have burned over $ 760 in commissions alone, moving over 41,300 shares in almost 1K fills. (With futures, both margins use and commission efficiency will be greatly improved) DD so far has been decently contained considering that I am trading <b>without (memory-less) stops</b> a folio of almost 80 instruments, and the lowest PNL has been around -3K. If you compare with the previous pictures (those from simulations), so far it looks pretty much as expected from simulations. We have massive gains and massive losses <b>part of which are being "recycled" into realized gains, through the "superposition" mechanism</b>. We expect the <b>drift</b> to "pull up" the PNL gradually up to the point that the unfavorable volatility and the unrealized should not be able to hurt anymore. In the meantime I have been refining the scalping/hedging games to adjust to the higher frequency of this test and to enhance the self hedging capabilities within a single layer, as well as across layers (using the information about instrument correlations).
While we wait patiently for our algo to pull up our PNL (beyond the influence of mkt volatility), let's produce some more "theory", also useful for a finer dynamic control of the scalping/ hedging action. In the previous posts we have introduced the "Relative drift" (<b>RD</b>) and "the Hedging Intensity"(<b>HI</b>). A final piece is missing in our puzzle. In fact we know that the performance (RD) is a resultant from the action of two opposite forces we put in action: the intensity of the "investment", on one side, and the intensity of the "hedging" action (HI) on another side. We need, finally, to introduce a metric for the "Investment intensity". To this purpose let me propose the following: <b>II := Investment Intensity := - Unr / | Loss | </b> This completes, for now, our picture, and may also provide in practice a useful metric to govern the algo. In fact, when we detect that the <b>Investment Intensity</b> is a "too large" fraction respect to the <b>Hedging Intensity</b>, we might unleash new "hedging entries" to correct the situation. [I would probably need to define what is meant by "hedging entries" and I am currently relying in your intuition for that. In next posts I will define formally what is meant by "hedging entry". For now just think of them as orders which have a "protective", "defensive" function, at the moment they are issued.] Finally, it may be worth noting what is the formal (and intuitive) relationship which links the the 3 metrics RD, HI, and II, we have been introducing. If we add them up, we get: <b> (100 + RD) = HI + II </b> which seems a nice decomposition, as it hints the "Relative drift" as a (percentage) rate of increase, resulting from the action of 2 opposite forces: "investing" and "hedging".
So let's recap the metrics we have seen so far. For greater accuracy it is appropriate to include also the <b>commissions</b> in the computations, as, for instance, a greater frequency in general can help smoothing the equity curve and exercise a finer control on drawdown, but it will cost more in commissions. So, it is appropriate to include the commissions in the computations, so that they will be taken into consideration in the evaluation. So the metrics more explicitly become: <b>RD</b> := (Gain + Loss - Comms) / | Loss | * 100 <b>HI</b> := (Unr + Gain - Comms) / | Loss | * 100 <b>II</b> := - Unr / | Loss | * 100 In such a way, we have explicitly included the commissions, and the relationship: 100 + RD = HI + II still holds true, as desired. Clearly, it depends on terminology: we could also have originally defined the "Gain" as net of the commissions (NetGain = Gain - Comms), and in that case we would not have to specify the additional Comms term. [By "Comms" here I simply mean the total commissions which have been spent for the orders executed up to the time the metrics are computed.] The above decomposition obviously continues to hold true if we take the respective means over multiple Monte Carlo simulations. So we have: <b>100 + E(RD) = E(HI) + E(II)</b>. [Clearly when comparing the avg metrics of different strategic games, one has to make sure to use always the <b>same specific share identification method</b> throughout the simulations, or else they would not be directly comparable, as they are obviously also functions of the share identification method. Also, for methodological consistency, I would suggest not to use share identification methods based on time (FIFO, LIFO, etc.) which appear to incorporate an additional element of arbitrariness (sorting orders by execution time is pretty meaningless in this context), but I would stick with specific methods which either minimize or maximize either Gains or Losses.]
So do you use other means to hedge, besides your virtual hedging? Options, opposite position in a correlated/cointegrated product, etc.?
Hi JoshDance, sure, one can use multiple means to the purpose of hedging, depending on the situations and personal preferences. Sometimes you can just use a correlated instrument, when there is a good reason to do so (e.g., instrument momentarily not tradable, desire to counterbalance or move the money on an inverse instrument with a stronger drift [daily compounding effects, and so on], etc.). Or you could plain and simple also use the same instrument (hedging entries, on the same layer, or on a separate layer). Also option configurations can be used. About options I generally sense quite a "resistance" from many people who I talk to. There is almost this feeling that the "algorithm" has to do all by itself and options seem, to many people, to require too much thinking (although I have made particularly easy to deal with them and automatically scan entire chains to easily find the most convenient protective structure). Supervision, however should be an important part of the any algorithmic system, and it's good to be able to think about multiple possibilities and tools to solve any situation may need to be addressed. Anyway, one can hedge quite well also just using the various etfs and futures and not using options. The "player superposition" and "layer overlay" can be effective enough to take care of any situation, if carefully programmed. One can decide how much risk to load up, before giving up and starting making "hedging entries" (which contribute to the "temporary" losses). It turns out to be a matter of balance and precision.