 Frank Volec est professeur de l'économie à Stanford University. Il a des fonds de spécialisation dans l'organisation solidaire et la théorie écologique. Il a recentement travaillé sur ces méthodes pour introduire la compétition dans les industries infrastructures comme les communications télécommuniques. L'électricité, l'eau, la delivery et les services postardaires et l'assistant sur les impacts de ces policiers de compétition sur le consommateur et l'économie. Frank est also chair of the Market Surveillance Committee of the California Independent System Operator for Electricity Supply. He is a member of the Emissions Market Advisory Committee for California's Market for Greenhouse Gas Emissions Aller-Once. He has an impressive list of publications, I will not detail here. So today we present work on the role of financial market participants in improving wholesale electricity market performance. Welcome, Frank. Thank you very much, Claude. That was far too long an introduction. I asked for a very short one, but Claude refused. Okay, because I like as much time as possible to talk. Okay, so what I'd like to talk about is essentially a much maligned group of participants in energy markets in general and electricity markets. And these are what we could call purely financial participants. This is sort of why I think they're at least the popular public view of financial market participants or you could say as traders. They essentially, you know, they buy something, they have no intention of consuming and sell something that they can't produce. All they do is, you know, attempt to buy low sell high or they can essentially sell first buy back later, short sales, which are even worse. These are typically, you know, turned at least in the press, speculators, and, you know, the big issue always is that these financial participants are taking money away from producers that produce the product and consumers that purchase the product. But this is the sort of popular view. This is a recent article in the New York Times that essentially talked about how, you know, as the power grid is becoming overworked, there are traders making lots of money. But I think the important point is what I have here in print in italics is that none of these profitable trades that these traders are doing are necessarily riskless. They all involve the traders taking on risk. Here's another one for those of you who might find of interest. This is actually a website that was started by one of the traders that sort of run afoul of the political process and decided that they wanted, they were going to fight it. What they did is they set up this website where they essentially sort of are pleading their case in the court of public opinion because, you know, for those of you who know anything about American sports, you might know about Tom Brady and the whole deflate gate. So in other words, the idea being that it's a similar thing here is that FERC is sort of the, if you like, the commissioner of the NFL and these guys aren't too happy so they are trying to go to the court of public opinion on this. What I really want to try to say is, look, these financial participants are no different, I think, from any other participants in the sense that they are basically doing what we'd like them to do which is trying to make a return on investment, serve the fiduciary responsibility of their shareholders and at least the way I like to think about it is that, look, just like essentially designers of buildings and aircrafts don't need to worry about gravity, we as designers of markets need to worry about essentially the so-called laws of economics which is, you know, think of as just simply the incentive constraint and regulators that deny the existence of this law sort of do so at their own peril. And, you know, I think that oftentimes what we're attributing is these undesirable outcomes are really not, you know, nefarious behavior but in many ways the result of a poorly designed market. A well-design market, and what we're moving towards, at least I think throughout the United States, is a fairly well-designed electricity market is that trade financial participants exploiting these profitable opportunities I think really can deliver benefits and that's the purpose of what I want to talk about today is essentially one such example of this. It's coming from the introduction of what's called convergence bidding or if really I think the more appropriate word but certainly politically incorrect word which is virtual trading and it's virtual trading because you're trading, if you like, virtual megawatts, you're not really trading megawatts that you know you're going to be able to deliver because remember what I told you about traders is they sell something they don't produce and they buy something they don't consume so clearly they're not going to do anything but just trade differences between the various markets and so the first is is that what I want to talk about is really the idea of the within the context of the California electricity market but this is true of a number of other markets introduced there are what are called multi-settlement markets and these markets are essentially where you can buy firm financial commitments day ahead which you can then sell back in real time and in February of 2011 California implemented convergence bidding and so what I'm going to talk about is based on paper on my website that looks at this question so the idea that we really want to look at is this idea that in these kinds of markets where you have risk neutral traders you get this relationship if you have zero transactions cost but in most of these markets and in particular the other thing that's definitely true about these virtual bidding markets is that these are if you like politically unfavored markets and so a lot of costs get layered on to the convergence bidders and so if what you're doing is making an assessment of whether or not this relationship holds there really is both an explicit as well as I think implicit cost of trading and so the real idea of whether or not profitable trading opportunity exist is whether or not if you like the absolute value of the difference between those two things is greater than the round trip transactions cost not necessarily that expectation is zero simply because there are per unit costs and so one of the things that makes this I think an interesting paper is recognizing the existence of trading cost and then what we're going to do is essentially assess whether or not when you introduce this new product virtual bidding did the sort of implicit cost of acting to essentially exploit this difference between day ahead that's the forward price and the short term spot price S did the cost of exploiting that fall and so that's the one thing we want to look at the other thing we want to look at is the idea that one of the things we think about with forward markets and multi-settlement is that the forward market is setting up the spot market to essentially clear if you like in as painless a manner as possible you would expect that the variance of the real time prices should reduce as a result of the convergence bidders I mean the way I like to think about what convergence bidding is doing is really just the classic information aggregation what we're doing is we're allowing all these financial market participants to effectively provide their input as to what they think the least cost way to dispatch the system in real time is whereas before without the financial market participants we were essentially saying those that own generation or serve load can make these guesses but with the introduction of the purely financial participants you're giving them this ability to provide their information their specialized knowledge so we'd expect both the variance real time prices as well as the variance of the difference between day ahead and real time prices because remember in a multi-settlement market if real time looks just like day ahead we should have almost zero variance and then finally the other is just the fact that there are systematic differences in this price difference across time so the other thing we're going to look at is just did this to the auto correlation in that price process and then the other thing we want to look at is essentially the efficiency of market outcomes since one of the things that certainly I will in full disclosure say that I was a shameless advocate of virtual bidding as a member of the market surveillance committee of the California ISO and this was the reason why it's because of this idea that you're giving all participants the opportunity to have an input into what is the least cost way to dispatch the system so the question is is that do we see that in terms of market outcomes meaning do we see a lower cost of dispatch as well as essentially less if you like heat energy being used to provide the electrical energy that actually comes out of the grid so in order to set the stage for this I need to give at least everybody some background on U.S. wholesale electricity market so the two things I'll try to explain at least briefly are essentially what we'll call multi-settlement as well as essentially locational marginal pricing which I know in Europe is absolute evil but in the United States it's considered the sort of market design of choice in fact all U.S. markets are essentially LMP markets at the moment so essentially what happens in the LMP market is you take all of the unit specific bids from all the generation units at all of their locations and you minimize as it says right here the as bid cost to serve demand at all locations in the transmission network subject to all relevant operating constraints transmission ramp rates you name it whatever constraint you want to put in it's solving a large optimization problem the LMP is essentially the if you like the optimized value of the objective function how does that change as a result of withdrawing one more unit of output at that location and what we get from the day ahead market is essentially schedules of generation units and prices and those schedules are as I said firm financial commitments which means you get paid for what you sell in the day ahead market regardless of what you do in real time it's just that it has financial consequences come real time as we'll discuss so then between day ahead in real time in most markets you're able to essentially revise your offer curves based upon what you think might be the changes in system conditions and then you run the process in real time so you essentially determine you do five minute dispatch every five minutes within the hour in real time to essentially what you're again doing the same thing you're solving that same optimization problem you're updating the transmission network for effectively the conditions that exist in each of those intervals just like in the day ahead you used your best guess of what you thought the real time conditions in the grid would be and then what happens is is that the hourly real time prices then are just simply the arithmetic average of the 12 LMPs that occur in that hour ok so and you know as you said the hourly real time prices are certainly more volatile than the day ahead prices simply because of you know limited flexibility but also you know in part market power sorts of issues as well so as I said the way that the settlement works is that you get paid but if you don't cover it with what you scheduled as a generator you're on the hook so if I only produce 30 I gotta buy 10 out of the real time market to replace the schedule that I had of 40 from the day ahead market similarly for the load serving entity alright so and so what happens is that you know what I can do is essentially as a generator supplier is I can play day ahead versus real time differences and this is what was at least initially called implicit virtual bidding if I as a supplier think that the day ahead price is lower than the real time price then what I'm gonna do is probably you know bid higher in the day ahead price or just withhold my output not even bid it in to the day ahead market and then sell it in real time or similarly if I'm a load serving entity and I think the day ahead price is more expensive than real time then what I'll do is I may not schedule all my energy I may not bid it all in in the day ahead and just purchase in real time and you know what this did was create some significant reliability consequences which essentially you know there would be times certainly in a number of markets where you know a significant amount of energy would be delayed until real time the operators would get a little nervous so virtual bidding was essentially introduced as a way to essentially say look if you expect that those two prices are gonna be equal you should just bid your energy in to the day ahead market because you know there's no advantage to you delaying which market you bid into virtual bidding is a way to essentially achieve that at the first stage so how does virtual bidding work well what you do is you as a virtual bidder you can submit a generation bid just like a physical generation unit you can submit that at any location in the grid that you want and you'll be treated in the day ahead market just like a generator but if you're taken in the day ahead market then there is a financial commitment that you will be a price taker in the real time market to essentially purchase that power you know back in the real time market similarly if you submit a demand bid in the day ahead market you'll be treated just like a load serving entity but what you're then going to be required to do for the amount that you sell is you're going to sell that back in real time and so what you can see is if I if I put a debt bid in the day ahead market I'm increasing demand at that location which is likely to increase the price if I put an ink bid in the day ahead market I'm going to increase supply at that location which is going to presumably depress the price both of and you know for the reason that I'm doing that is because I think I want to essentially sell at the real I want to exploit that price sorry exploit that price difference so the other thing about explicit virtual bids is they are identified to the system operator as such so in other words I as the ISO I see these virtual bids come in when they clear I know that they're virtual bids and I know what's going to happen to them in real time and the other is is that financial participants can virtual bid anyone that wants to can virtual bid there's some problems in most of the US markets in the sense that the public utilities commissions that effectively regulate the retail side of the market tend to frown on the retailers that are serving load to virtual bid which has some problems but for the most part if you want to set up shop as a virtual bidder and trader you post your bond and you can do it ok so why would we expect virtual bidding to reduce these sorts of trading cost improve price convergence well the first thing is essentially I can only implicit virtual bid as a generator over the range of my generation unit ok and they can only implicitly virtual bid where I own a generation unit ok similarly the way it works in California is that load serving entities can only bid within the range of their expected demand I can't bid into the day I had market a demand level that's far above my demand the system operator will certainly prohibit me from doing that or at least and so you know effectively the other thing that's true is that in California is you as a utility you only bid at your service territory level rather than when you bid at the node level and so what this really does is prevents load serving entities at least in California from implicitly virtually bidding at load nodes so in other words generators could implicit virtual bid at generation nodes in the early days before explicit virtual bidding but load nodes couldn't so we might expect that if you like the cost the transactions cost associated with virtual bidding at load nodes would be higher traditionally rather than the transactions cost associated with bidding at the generation nodes initially and we'll investigate that one empirically as well ok so the other is is that why would we expect that it would improve market performance and this just simply has to do with the fact that what we discuss which is if there could be long start units I want to make sure that long start unit gets on I can essentially to make sure that long start unit gets committed as a generator I could submit a deck bid at that location to essentially make sure that the demand level is high enough so that long start unit gets committed as opposed to a short start unit that effectively would excuse me be more expensive to run in real time set a higher real time price and so what we could expect is and as well it's just the fact that I think the important thing to remember you know in an LMP market there are thousands of constraints you know and one of the things that we think markets do a very good job of is solve very very complex optimization problems and believe me this is one of the one of the most complex optimization problems I can think of which is essentially how to most efficiently dispatch a system in real time and we could think of that you know what the virtual bidders are doing is through if you like just their desire to exploit this location or price differences they are presumably yielding a dispatch that is lower cost than would be the case if they weren't trying to do that ok so the two things we're going to investigate is essentially this idea of closing the gap reducing this trading cost and you know as I said it's not equal to zero but it's equal to the value less than this trading cost and so we'd expect that you know more accurate scheduling the other thing and we would expect from that is both this implicit trading cost that we talk about to fall as well as we'd expect this variance of the real time and the day ahead prices to fall then the other thing we'd expect is that this sort of dynamic to be working is that the virtual bidders that essentially are good at essentially anticipating where real time demand and supply is needed they're going to be rewarded with profits and they're just going to create this incentive for them to essentially produce the least cost mix of demands so we're going to look at both this improved price convergence and market efficiency for California so what do we do we use the hourly prices and first we'll just do it at the lap level because essentially there are three laps in California these are essentially the service territories of the utilities and each lap price is computed as essentially the nodal load weighted average of the LMPs in each of the firm service territory and then I'm going to present some summary results for the nodal levels just to show you that what we did at the lap level is very representative of what's true at the nodal level and as I said all generation units are paid their nodal price there's you know over 5000 nodes in California so and this is essentially the lap levels of each of the sing so the different colors San Diego Gas and Electric Southern California Edison SCE Pacific Gas and Electric these are the various locations of the nodes and so think of it as what you're doing to get the lap price for PG&E as you're taking the weighted average of all load nodes in their service territory for SCE you're doing the same thing so what we've done here is essentially computed the in each of these graphs just to show you is for each hour of the day we've computed the average sample average over that two year period a year before the implementation of virtual bidding versus and then a year after virtual bidding we've computed the hourly average before virtual bidding and then the point wise 95% confidence interval on that price difference so right here is the price difference for the first hour, the second hour and so on and then over here is the same thing and the thing you can see is that there are certainly large deviations of the 95% confidence interval over here they get a little better this does it for Southern California Edison you can see it gets a little better for them as well San Diego you can see it gets a little better that's pre versus post explicit virtual bidding but if you did the test of whether or not all these means are zero in other words just the standard 24 dimensional check true the statistic would be lower but you know I think this really points to the fact that that's the wrong way to think about it because there are explicit trading cost as well as implicit trading cost associated with essentially trading this difference and so the right way we think about it is to say okay let's think of it in terms of essentially an economic model and the economic model goes like this as we say okay there's a trader and he's got access to these 24 assets so think of this as the vector of price differences across hours of the day and what he does is he essentially he has expected trading cost which is essentially the portfolio weights times the mean of those price differences those portfolio weights can be positive or negative positive means buy one sell the other negative means sell one buy the other in terms of real time versus day ahead and then what we have is essentially the trading cost the trading costs are assessed on the absolute value of the position that you take and then C if you like sort of is the trading cost and then what we have is essentially just the fact that we're going to impose the normalization that the sum of the absolute positions that you take is equal to one and therefore what the trader wants to do is essentially maximize expected profit so the idea that we're about is to say okay let essentially this guy a star equal to the expected profit maximizing portfolio given the vector of weights and what we're interested in is saying is that is essentially the different between that expected profit maximizing portfolio C greater than zero versus the alternative that it's not okay and so this you might look just like a sort of application of the delta method but the problem is is that this function is not differentiable so essentially the conventional delta method doesn't work but what it turns out to be is essentially directionally differentiable and there's some recent work by that essentially shows how to compute the at least an estimate of the asymptotic distribution of that and what we do is essentially employ their method which was essentially turned into this essentially numerical derivative based approach by Hong and Lee so what we come up with is essentially two values which is what we call is we use this distribution to say okay what's the smallest value of this implicit trading cost that causes us to reject the null hypothesis that a profitable trading charge exists the other one is that we can compute is the largest value of C that causes not rejection sorry of the null hypothesis that no profitable trading strategy exists right so in other words if this inequality is satisfied then no profitable trading strategy exists and so what we come up with is essentially those values using this procedure the other thing we do is as you can see we consider a pretty simple trading strategy which is based only on essentially the sort of mean of the price difference vector but the interesting thing is that we then look to see if there's auto correlation so one of the things important to bear in mind is that remember when you're submitting your bids for the next day in the day ahead market the real time market for today is operating so essentially there is first order auto correlation in the price difference vector meaning the price difference between the day ahead price and the real time price for today and the price difference for the real time price the day ahead price for say tomorrow there is first order auto correlation that you can't exploit simply because when you're submitting your bids that still has not been resolved and so essentially what we expect is there to be first order auto correlation but beyond that we want to see whether or not there's any remaining auto correlation to be exploited and so we have this test of whether we go out to in our case we go out to lag 10 and we essentially find there's very little evidence against the null hypothesis that essentially once you get past the first order auto correlation that you know should be there because you can't exploit it there is no exploitable auto correlation in either before or after in sort of the lap prices which makes sense given that pre essentially the implementation of this load we're able to effectively implicit virtual bid at the lap level you might expect that they were doing quite a bit of it ok so what we do is we as we said we compute these implicit trading costs that we talked about here and the only thing that I think it's important to see is that you can see both before versus after virtual bidding they fall meaning that if you like a product certainly reduced the implicit cost of trading and arbitrage in these price differences this just shows what the two distributions look like the purple being before the whatever that green color is being the after then the other thing we look at is we say ok let's look at the distribution of if you like the difference between pre versus post for the cost so in other words we can use those two distributions that we got the distribution of the trading charge before the distribution of the trading charge after and we can compute the distribution of the difference between seed pre minus post and what we can do is then construct a test of whether or not the null hypothesis that essentially the trading charge before minus post is greater than zero and what we find is essentially for San Diego Gas and Electric and Verus CE we certainly reject the null that essentially this is true and we don't reject this null which is and in both cases if you like in PG&E we don't reject either which would say that at least we have no evidence against the null that trading charges fell after the introduction of convergence bidding ok so then the other thing we looked at we talked about is this idea of the variance so what we do is we say ok we've got the variance covariance matrix of the day ahead versus real time price differences and our hypothesis test is that the difference between pre that variance covariance minus post is a positive semi-definite matrix essentially this just amounts to a multivariate non-linear inequality constraints test of the elements of the 24 eigenvalues and what we do is we is essentially moving block bootstraps approach to do that and what you can see both ways so we first do it pre minus post the positive definite matrix post minus pre is a positive definite matrix and these are essentially the p values and so what you can see is for pre minus post in all cases for the price difference we can't reject the null that essentially pre minus post is a positive semi-definite matrix and for real time price we also can't reject it and then for essentially the other way all of them we can reject it at a 0.05 level except again for San Diego gas and electric there appears to be the p value is certainly bigger than 0.05 so it certainly is consistent with explicit virtual bidding reducing price volatility ok so the other thing we do is this idea that we talked about we do things at the nodal level and at the nodal level you can see that we say ok post explicit virtual bidder falls after the introduction of virtual bids the generation node indicator negative meaning that we said remember we said that at the generation node we expect the trading cost to be smaller and then the other thing is we expect is that after we introduce explicit virtual bidding you can do it at any node so that the gap should close and what we see is essentially exactly what we'd expect where you can see that post explicit virtual bidding the trading charge falls that's the minus 2.527 and then after you implement it you can see that the gen node and the post explicit virtual bid coefficient are not statistically different their sum is not statistically different from 0 which means that after you implement explicit virtual bidding there's no systematic difference between the cost at gen nodes versus load nodes ok so in the interest of getting to this I'll skip to here so what we did here is we're looking at market performance so what we did is for each hour we compute the total variable cost to essentially operating all California ISO generation units using the price of natural gas the spot price of natural gas for that day we took the total amount of heat energy that was used to produce that power including the cost to starting up the units in terms of heat energy and then we had total starts and then we said ok we're essentially the composition of how the output is being produced so we're controlling for in-state renewables imports, natural gas prices and then we run that sort of non parametrically so everything in the Z is essentially done non parametrically and then we have essentially in the W's just things that include our day and month of year fixed effects and then essentially in some specifications X is just this indicator that says pre versus post explicit virtual bidding and we used essentially Robinson semi parametric estimation procedure with the crossvalidating to choose the age and what we get is roughly about for the same composition of output, roughly 6% less energy is used pre versus both explicit virtual bidding and then total cost is roughly the same the interesting thing is that this is not small small potatoes roughly about 30 million dollars annually in terms of savings of total variable cost of operating and a significant amount of greenhouse gas emissions so traders are not only cost efficient but they're environmentally friendly so in this this shows how essentially the effect differs across hours the day and the big thing that I think is interesting is that certainly one of the things that the ISO operators talked about was essentially this idea of managing the so-called morning ramp and what you can see is that is that the big savings appear to be coming during those early morning hours of the day relative to the later hours of the day so this is basically the end is to say that what we've got is essentially looking at how has financial market participants benefited market performance essentially it's by essentially reducing that cost of implicit trading cost associated with trading the difference between day ahead and real-time prices which has then led to essentially this lower volatility in real-time lower volatility in the difference between day ahead and real-time and then moreover this reduced cost of dispatching the system reduced amount of heat energy needed to essentially produce the same composition of output in California so financial market participants can benefit market efficiency I think in a market design to allow them to do that ok thanks