The Porto gasoline market:
Collusion and asymmetric pricing
Master Thesis Econometrics
The Porto gasoline market:
Collusion and asymmetric pricing
Abstract
Contents
1 Market structure and data description 6
1.1 Market structure . . . 6 1.2 Geographical area and relevant market . . . 9 1.3 Data description . . . 10
2 Pricing behavior 13
2.1 Pricing behavior . . . 14 2.2 Price fixing . . . 17
3 Vertical relationship and collusion 22
A Chapter 1 48
B Chapter 2 50
Acknowledgment
Introduction
In this thesis we use a disaggregate data set of weekly retail and ex-refinery (Platts) prices to examine the gasoline market in Porto (Portugal). The main aim is to investigate if the market shows any sign of anti-competitive behavior. To this end, we examine the pricing behavior of the stations and we construct a theoretical model that can help us distinguish between competitive and anti-competitive behavior.
To examine the pricing behavior of the retailers and wholesalers, we deter-mine for each brand the so-called recommended price, which is the price that an oil company recommends its stations to use. We find that a large share of stations follow the recommended price or use a price that is very close to the recommended price. Further, there is not much difference between the recom-mended prices of the oil companies. This results in a retail market with a very low level of price dispersion. Further, we try to determine if there is any sign of price-fixing agreement among the retailers or wholesalers.
et al. (1997) describe in a theoretical model how price coordination can induce asymmetric price adjustments, i.e. a rapid response of the prices to increases in cost and a slow response to decreases in cost. We show that price coordi-nation among the wholesalers can also induce asymmetric price adjustments in the vertical structure. Since we do not generally expect asymmetric price adjustments in a competitive market, we can use this prediction to test if there is any sign of collusive behavior.
There is a comprehensive list of applied econometric papers that try to figure out if the retail price adjusts asymmetrically to upstream prices, see Geweke (2004) for a literature overview. A frequently used model is derived in Borenstein et al. (1997) and reviewed in Bachmeier and Griffin (2003). Their model is an Error Correction Model (ECM) that incorporates the long-term equilibrium between the retail and upstream price. We use their model to determine the relationship between the Porto retail price and the ex-refinery price.
Chapter 1
Market structure and data
description
In this chapter we examine how the market is organised. We describe the upstream and retail market by examining market concentration, gasoline dis-tribution and the contracts between the firms. Further, we give a description of the geographical area and the data.
1.1
Market structure
The Portuguese gasoline market has three layers: refinery, wholesale and retail. The major oil companies that are active in Portugal are Firm 1, Firm 2, Firm 3, Firm 4, Firm 5 and Firm 61 In November 2007, it was made public that Firm
1 will buy the Firm 6 stations in Portugal and Spain.
Crude oil There is no oil production in Portugal, thus crude oil has to be imported from abroad. The type of crude oil that is mostly imported is of the Brent type, which is extracted from the North Sea, and is traded on the IntercontinentalExchange (ICE) market.
1
Refinery There are two oil refineries in Portugal. Firm 1, the incumbent oil company, is the sole owner of both refineries and imports the crude oil. It is possible to import refined oil from Spain, but this does not happen often: the production of the Firm 1’s refineries accounts for more than 90% of the internal gasoline consumption2. The reason for this is that one has to pay transportation
costs to import refined gasoline from Spain. Further, the price at which the refineries sell the gasoline is called the ex-refinery price, and it is set by Firm 1.
Wholesale Each wholesaler buys gasoline from Firm 1’s refineries or termi-nals and distributes it to the service stations that operate under its banner. The companies that are active on this market are the major oil companies and two independent oil companies. Typically, a wholesaler, say Firm 3, sells its gasoline to Firm 3 stations. The major oil companies account for about 90% of wholesales3.
Retail The retailers are the service stations that sell the gasoline to the con-sumers. There are three kinds of service stations: major brand stations, inde-pendent stations and white pumps. Major brand stations sell the brand of a major oil company and operate under the following contracts:
• Company-owned and company-operated (COCO): The oil company com-pletely owns the service station and has Firm 4 control of the station, including the retail price.
• Company-owned and dealer-operated (CODO): The oil company owns the
2
AdC.
3
land and the infrastructure, and the dealer is a self-appointed manager of the service station. In this case, it is the dealer that sets the retail price.
• Dealer-owned and dealer operated (DODO): The dealer completely owns the service station and sets the retail price. Further, he has a franchise contract with an oil company.
The independent stations sell gasoline under their own brands and the white pumps sell brand-less gasoline. White pumps are located at hypermarkets, for example Carrefour. The independent stations account for 5% to 10% of the retail market and the white pumps account for less than 5%. The rest of the market is served by the major brand stations4.
There is a strong concentration over the major oil companies in Portugal. In Table 1.1 one can find the number and percentage of stations that serve a major brand (separated by type)5.
Number of stations Percentage of the retail market (rounded)
Firm 1 840 38%
Firm 2 380 17%
Firm 4 300 13%
Firm 3 290 12%
Table 1.1: Distribution of major brand stations in Portugal.
Thus about 80% of the stations in Portugal operate under the Firm 1, Firm 2, Firm 4 or Firm 3 banner.
In our local market, i.e. the Porto county, there are 44 stations active. All these stations serve major brands, thus the number of independent stations and white pumps is zero. The brand distribution is given in Table 1.26.
4AdC. 5
AdC.
Number of stations Percentage of the retail market (rounded) Firm 1 18 41% Firm 2 10 23% Firm 4 7 16% Firm 3 7 16% Firm 5 2 5%
Table 1.2: Distribution of the stations in the Porto county.
We observe that the brand distribution in Porto is similar to the nation-wide brand distribution.
1.2
Geographical area and relevant market
Portugal is divided in so-called NUTs and these NUTs are divided in counties. In Figure A.1 (Appendix A) one can find a map of Portugal, where the different NUTs are shown with different colors.
The county under investigation is located in the Great Porto NUT, which is indicated by Square 1. A more detailed map of Square 1 is given in Figure A.2. The Porto county is indicated by the square in Figure A.2. Further, we have indicated the Great Lisbon NUT by Square 2 in Figure A.1.
Since a station solely competes with its neighbouring stations, we consider the Porto county as the relevant market for the stations that are operating in that county.
However, as said in the introduction, each major oil company has a policy to advise all its stations to use a certain price. The relevant market for this price policy is the national market, as these oil companies compete with each other on the national level. These two market definitions are in accordance with the findings of the European Commission (EC)7.
7
1.3
Data description
The AdC has provided us with price data that corresponds to different market levels. For each market level, we will state the relevant price data.
Crude oil The Brent crude oil price was provided by the AdC. The series consists of weekly averages in/liter.
Refinery The ex-refinery price, i.e. the price at which Firm 1 sells the refined gasoline, is not publicly available. However, the AdC has informed us the Portuguese ex-refinery price follows the so-called Platts spot price in a very specific manner. The Platts spot price is the international price of gasoline on the spot market.
The main variable component of the ex-refinery price follows the Platts as follows: let p(t) be the Platts at week t and r(t) be the main variable component of the ex-refinery price at week t, both in $/liter. We have that r(t) = pt−1+pt−2
2 .
The main variable component of the ex-refinery price, i.e. r(t), is also called the reference Platts in the gasoline market. The AdC has provided us the reference Platts series and has scaled it to /liter by multiplying it by the average exchange rate of week t − 1.
Wholesale To examine the retail markup we need to know how much each retailer pays its wholesaler, which typically is a major oil company, for gasoline supplies. Alas, the wholesale prices are not available to us. However, we can
use the ex-refinery price as a proxy for the retail margins. We now explain how. For a COCO station, which is owned by an oil company, a margin based on the retail price minus the ex-refinery price is a good proxy for the actual retail margin, as in this case the oil company does not face any wholesale cost, as she is her own wholesaler.
But in the case of CODO and DODO stations, a margin based on the ex-refinery price will not contain the complete price that a retailer has to pay to its oil company. However, this margin can act as a proxy, as the wholesale price will depend on the ex-refinery price. Further, CODO and DODO stations are also co-owned by the oil companies, which means that this margin is suitable to measure their share of the retail profits.
Retail The stations’ prices are only measured on Thursday’s afternoon. We do prefer weekly average prices, as upstream level prices would then match the station prices. But the problem is less grave than it seems, because the AdC has found that the stations change their prices about once a week, see AdC (2008). The Thursday’s prices do not exactly correspond to the weekly average prices, but we will use them as an approximation of the weekly average prices. However, we have to keep in mind that this can affect the randomness of our data set.
10 Firm 2 stations, 5 Firm 4 stations and 1 Firm 5 station. Further, all prices are net of taxes and in/liter.
In our thesis each week starts on Monday and ends on Sunday, and we denote the week by its Monday. All the above price series are from week 16/02/2004 till week 16/10/2006. Thus we have weekly prices for a period of 140 weeks.
On the average, each oil company recommends its stations to use a certain price every week, the so-called recommended price, see AdC (2008). From the AdC we have also received retail prices of the service stations that are located in the Great Lisbon NUT and the Great Porto NUT, where the latter includes the Porto county, as said before. We will use this data set to determine the recommended prices by comparing for each week and for each brand the prices that were used in the Great Lisbon and Great Porto NUT. For each week and brand, the prices that are the most common between the two NUTS will be selected as recommended prices.
Chapter 2
Pricing behavior
In this chapter we will empirically examine the retail market. The examination can be divided in two parts: in the first part we examine the pricing behavior of the retailers and wholesalers, and in the second part we examine if there is any evidence of a simple price-fixing scheme among the wholesalers or retailers. We start by analyzing the retail price, reference Platts and average markup, which we define as the difference between the average retail price and the refer-ence Platts. In Table 2.1 one can find some descriptive statistics of the series. Further, to examine how the series develop over time, we have plotted the series in Figure B.1 (Appendix B). From this graph we observe that both the retail price and the Platts are steadily increasing. Further, we observe that there is a week to week variation in the markup, but in the long-run it does seem to stabilize around 14 cents per liter.
Reference Platts Average retail price Average markup
Mean 31.945 45.646 13.701
Median 31.200 46.992 13.973
Maximum 48.390 56.731 21.065
Minimum 19.1600 29.508 6.811
Std. Dev. 7.660 8.000 3.677
To examine if the Porto retail market behaves similarly to a typical Por-tuguese retail market, we have compared the Porto average retail price to the nationwide average retail price by plotting the two series in Figure B.2 (Ap-pendix B)1. This graph clearly shows that the Porto retail price follows the
average price. However, the Porto price is shifted upwards for a large number of consecutive weeks. This suggest that there are additional forces active in the Porto retail market (compared to the average Portuguese retail market).
2.1
Pricing behavior
90% of the gasoline that is sold in Portugal is produced at two Firm 1 re-fineries. Each oil company may add some additive to their branded gasoline, but this does not significantly change the product. Thus gasoline is nearly a homogeneous product.
To measure the price dispersion between the stations we estimate the fol-lowing OLS model2:
Rit= ct+ eit, (2.1)
where Rit is the retail price of station i in week t, ct is a constant that varies
with the week and eitis the residual of firm i in week t. This equation measures
for each station its deviation from the weekly average, as the OLS estimate of ctis the average retail price in week t.
The Firm 4 number of observation is 1960 and the R2 is 0.982. In Figure B.3 (Appendix B) we give the density distribution of the residuals. From the
1
The average national retail price can be found at: www.ec.europa.eu/energy/oil.
histogram we observe that the residuals are tightly distributed around zero. For example, around 80% of all the prices have a deviation from the mean that is smaller than 1 cent and around 60% of all the prices have a devation that is smaller than 0.5 cent. Thus we can conclude that the level of price dispersion is low.
As said before, on average, every week an oil company advises all its sta-tions to use its recommended price for the coming week. A station is free to deviate from that price, however it is suspected that a large share of stations use their recommended prices. In Table 2.2 we give, for each brand, the average difference between its recommended price and the average recommended price. From this table we can conclude that the level of price dispersion between the recommended prices is low3.
Firm 1 Firm 2 Firm 4 Firm 3 Mean -0.033 -0.029 0.077 -0.015
Maximum 1.198 1.365 1.604 1.736
Minimum -1.282 -2.562 -1.756 -1.384 Std. Dev. 0.424 0.610 0.607 0.560
Table 2.2: Recommended prices: Average price differentials (cents/liter).
To determine for each brand the extent to which its stations use its rec-ommended price, we have plotted for each brand a histogram (Figure B.4, Ap-pendix B) that shows for every week how many stations use its recommended price. We observe that, on average, the majority of the Firm 1, Firm 4 and Firm 3 stations use their corresponding recommended prices. The Firm 2 stations are an exception, as the stations do not often use the Firm 2 recommended
3For example, consider the average price difference between buying at Firm 1 and Firm 4:
price.
From the histograms we also observe that the number of stations that use their corresponding recommended prices do vary over the whole period, but there is a high degree of persistence from week to week. The Firm 2 stations are again an exception, as Figure B.4 (Appendix B) shows that there is no persistence in the number of stations that use the Firm 2 recommended price.
To determine the distribution of the stations’s prices around their recom-mended prices, we estimate for each brand the following model:
Rit= At+ eit, (2.2)
where Ritis the retail price of station i in week t, Atis its recommended price in
week t and eitis the residual of firm i in week t. Thus eitmeasures the deviation
of station i’s price from its recommended price. In Figure B.5 (Appendix B) we give for each brand the density of the deviations of its stations’ prices from its recommended price. For the Firm 1 stations, about 90% of all their prices have a deviation that is less than 0.5 cents. For the Firm 2 stations, the percentage is 55%, for the Firm 4 stations it is 71% and for the Firm 3 stations it is 95%. Thus, for each brand, except Firm 2, the prices of its stations are tightly distributed around its recommended price.
Some striking observations from the density distributions:
• Firm 1: despite the prices of its stations being tightly distributed around its recommended price, there are instances where stations use a relatively much higher or lower price, but they show no persistence.
the other brands. However, still about 90% of all its stations’ prices have a deviation that is smaller than 2 cents.
• Firm 4: there are no large deviations from the recommended price. Fur-ther, all significant deviations are results from the stations prices being lower than the recommended price. Furthermore, in contrast with the Firm 1 stations, there is some persistence in the deviations.
• Firm 3: in contrast to the Firm 4 stations, most deviations are results from the stations prices being higher than the recommended price and they do not show much persistence.
Thus we can conclude that a large share of stations follow the recommended price or use a price that is very close to the recommended price. Further, there is not much difference between the recommended prices of the oil companies. This results in a retail market with a low level of price dispersion.
2.2
Price fixing
not necessarily mean that there is a price-fixing scheme, however it does indicate that an in-depth investigation may be desirable.
Wholesalers As noted in Section 1.3, on average, every week an oil company recommends its stations to use a certain price for the coming week. A station is free to deviate from that price. In this section we are going to investigate whether the weekly recommended prices of the oil companies show any sign of the existence of price-fixing schemes. In Table 2.3 we give, for each wholesaler, the number of weeks in which its recommended price was equal to the recom-mended price of another wholesaler. Further, we also give the largest number of consecutive weeks in which they had equal recommended prices.
Stations Equal recommended prices Largest number of weeks
Firm 1-Firm 3 11 3 Firm 1- Firm 4 27 5 Firm 1-Rep 3 1 Firm 1-Firm 5 0 -Firm 3--Firm 4 14 4 Firm 3-Firm 2 10 4 Firm 3-Firm 5 0 -Firm 4-Rep 6 3 Firm 4-Firm 5 0 -Firm 2--Firm 5 0
-Table 2.3: The number of weeks (of a Firm 4 of 140 weeks) that two brands had equal recommended prices and the largest number of consecutive weeks in which they had equal recommended prices.
respectively.
To examine the spread of the weeks in which the oil companies had equal recommended prices, we give in Figure B.6 (Appendix B) the weeks in which the pairs Firm 1-Firm 4, Firm 3-Firm 4, Firm 1-Firm 3 and Firm 3-Firm 2 had equal recommended prices. We observe that there is a large spread in the weeks in which they had equal recommended prices. We are of the opinion that, if there was a price-fixing scheme in place, we would observe less spread in the weeks in which they had equal recommended prices. We can conclude that it is not likely that, during the relevant period, there was a simple price-fixing scheme in place where the oil companies use the same recommended prices.
Retailers We will now determine if there has been a price-fixing scheme among the retail stations of different brands. From the Canadian price-fixing case (MG (2008)) it is clear that a price-fixing scheme among stations of differ-ent brands implies that the stations that take part of this scheme use the same retail price for a long period. As stated before, finding that stations of different brands have used the same price does not necessarily mean that there was a price-fixing scheme, as it could be the case that this is due normal local retail competition. Additionally, in this case, it could also be the case that some sta-tions of different brands had the same price because their recommended prices were the same.
had the same prices and the same recommended prices.
Stations Equal prices Equal recommended prices Largest number of weeks
Firm 1-Firm 2 30 3 4 Firm 1-Firm 4 33 27 5 Firm 1-Firm 3 11 11 3 Firm 1-Firm 5 0 - -Firm 2--Firm 4 28 6 4 Firm 2-Firm 3 18 10 4 Firm 2-Firm 5 1 0 1 Firm 4-Firm 3 23 14 4 Firm 4-Firm 5 0 - -Firm 3--Firm 5 0 -
-Table 2.4: The number of weeks that stations of different brands had the same price, the number of weeks that two brands had the same recommended prices and the largest number of consecutive of weeks in which they had equal recom-mended prices.
From the table we observe that there have been many weeks in which sta-tions of different brands had the same prices, except for the Firm 5 station. Further, we observe that most stations of different brands had the same prices, because their recommended prices were the same. However, there are two ex-ceptions: stations from Firm 1 and Firm 2 and from Firm 2 and Firm 4. We will now examine these two cases in greater detail.
In the case of the Firm 1 and Firm 2 stations, further examination showed us that in all 30 ‘possibly collusive’ weeks the ‘collusive’ price was equal to Firm 1’s recommended price, which indicates that, in those weeks, some Firm 2 stations used Firm 1’s recommended price. As a matter of fact, in those weeks, on average, two Firm 2 stations used Firm 1’s recommended price.
Chapter 3
Vertical relationship and
collusion
In this chapter we construct a theoretical model of the vertical relationship between wholesalers and retailers. In gasoline markets it is common that each wholesaler, which, in our market, is an integrated oil company, provides gaso-line to the retailers that operate under her banner. Following Tirole (1988), we model this situation in a two-stage game where in the first stage the whole-salers set their wholesale prices and in the second stage the retailers set their retail prices. At the end, we use our model to determine the nature of a possi-ble collusion among the wholesalers and relate our model to asymmetric price adjustments.
3.1
Wholesaler with competing retailers
for a depiction of this simple relationship1. Note that this model describes the
relation between a wholesaler and CODO and/or DODO stations. Since COCO stations are owned and controlled by the wholesaler, they are not included in this model. However, we will show that adding a COCO station in the model does not modify the main point of the model.
Figure 3.1: Vertical structure: One wholesaler and two competing retailers.
The Firm 4 demand is given by q = D(p), where p denotes the price that the retailer charges and q denotes the quantity that is bought by the consumers. The wholesale price at which the wholesaler sells her product to the retailers is denoted by pw. We suppose that the wholesaler has a marginal cost of c > 0
and that the retailer has a retail cost of γ > 0 per unit of sales2.
Assuming that the retailers are identical, the demand function for the first retailer is given by:
Dr(p1|p2) = D(p1) if p1< p2; 0 if p1> p2; 1 2D(p1) if p1= p2, (3.1)
where p1 and p2 denote the price of the first and second retailer, respectively.
1
Trivially, we can extend this model to more than two retailers. However, to ease the exposition and analyses, we only consider two retailers.
2
By replacing p1 with p2 and vice versa in (3.1), we find the demand function
Dr(p2|p1) for the second retailer. The profits functions of the first and second
retailer are given by:
Πr1(p1) = (p1− pw− γ)Dr(p1|p2);
Πr2(p2) = (p2− pw− γ)Dr(p2|p1).
The profits function of the wholesaler is given by:
Πw(pw) = (pw− c)(Dr(p1|p2) + Dr(p2|p1)), (3.2)
as the Firm 4 demand for her product is the sum of the demand for the first and second retailer.
Before we examine the above model, consider the case where the wholesaler and retailers are vertically integrated, i.e. the wholesaler has complete control over the retailers and, thus, sells directly to the consumers. In this case, the only target of the integrated structure is the retail price p. The profits function of the integrated structure, denoted by Π(p), is given by:
Π(p) = (p − c − γ)D(p),
as the integrated structure faces a Firm 4 marginal cost of c+γ. By maximizing Π(p), we get the following First Order Condition (FOC) for the monopoly price pm:
(pm− c − γ)D0(pm) + D(pm) = 0. (3.3)
Consider now the decentralized structure. It is clear that, for any wholesale price pw, price competition will induce the retailers to price at their Firm 4
marginal cost ( i.e. wholesale and retail cost) pw+ γ. Thus the wholesaler has
the ability to set the retail price through the wholesale price pw, which means
that, similar to the integrated structure, the wholesaler has complete control over the retailers. This implies that we can rewrite the wholesaler’s profits function (3.2) into:
Πw(pw) = (pw− c)D(pw+ γ). (3.4)
The FOC of the above equation is:
(pw− c)D0(pw+ γ) + D(pw+ γ) = 0. (3.5)
From the FOC of the integrated structure given in (3.3), it follows that (3.5) is solved by pw = pm− γ. Since we have assumed that the SOC is met in the
integrated structure model, we can conclude that pw = pm− γ maximizes the
profits of the wholesaler.
Thus, in stage 1, the wholesaler chooses wholesale price pw = pm− γ and
earns the integrated-structure profits Πm (as Πw(pm− γ) = Πm). In stage 2,
the retailers use the monopoly price (as pw+ γ = pm) and earn zero profits.
R2, adding a COCO station to retailers has no effect on the equilibrium nor
the profits of the wholesaler. Thus, in this model, the COCO stations can be considered as a ‘standard’ retailer.
3.2
Competing wholesalers with competing retailers
In this section we extend the previous model by adding a competing wholesaler with two competing retailers, see Figure 3.2 for a depiction of this vertical structure. We assume that the wholesalers simultaneously set their wholesale prices, and, subsequently, each retailer learns its wholesale price and sets its retail price. The wholesalers are denoted by W1 and W2, the retailers of W1are
Figure 3.2: Vertical structure: Two wholesalers and four competing retailers.
denoted by R11 and R12, and the retailers of W2 are denoted by R21 and R22.
Further, the retail price of Rij is denoted by pij (i, j = 1, 2), and the wholesale
prices of W1 and W2 are denoted by pw1 and pw2, respectively. Since each
Dr(p11|p12, p21, p22) =
D(p11) if p11 is lower than the prices of the other retailers; 1
2D(p11) if p11 is equal to the price of one retailer
and lower than the prices of the remaining retailers;
1
3D(p11) if p11 is equal to the prices of two retailers
and lower than the price of the remaining retailer;
1
4D(p11) if p11 is equal to the prices of the other retailers.
(3.6)
Using the above demand function we can define the profits function of R11:
Πr11(p11) = (p11− pw1− γ)Dr(p11|p12, p21, p22), (3.7)
as the the Firm 4 marginal cost of the retailer consists of the wholesale price pw1 and retail cost γ. In a similar manner, we can find the demand and profits
function of R12, R21 and R22.
Using the above, we can now define the profits functions of the wholesalers. The profits function of W1 is given by:
Πw1(pw1) = (pw1 − c)(Dr(p11|p12, p21, p22) + Dr(p12|p11, p21, p22)). (3.8)
as the wholesaler’s Firm 4 demand is the demand of retailers R11and R12, and
its marginal cost is c. Similarly, the profits function of W2 is given by:
Πw2(pw2) = (pw2 − c)(Dr(p21|p11, p12, p22) + Dr(p22|p11, p12, p21)).
For this model we propose the following equilibrium:
Proposition. The wholesalers set their wholesale prices equal to the marginal cost: pw1 = pw2 = c. Next, the retailers price at their marginal cost: p11 =
p12 = p21 = p22 = c + γ. In this case, the retailers and wholesalers earn zero
Proof. Consider in stage 1 the incentives of, say, wholesaler W1. From the
equilibrium it follows that in stage 1 W2 sets pw2 = c. We are now going to
show, that, given pw2 = c, wholesaler W1 has no incentive to deviate from the
equilibrium.
If pw1 > c, then R11 and R12 have a marginal cost of pw + γ, and will
be undercut by R21 and R22, as pw + γ > c + γ. In this case we have that
Πw1(pw1) = 0, as its Firm 4 demand is zero.
And, if pw1 < c, then R11 and R12 are able to undercut the other retailers,
as pw+γ < c+γ. However, in this case, it follows from (3.8) that Πw1(pw1) < 0.
Thus we can conclude that wholesaler W1 has no incentive to deviate from
the equilibrium price c. Since the model is symmetric, we have that the same hold for W2. Further, given that the wholesalers are using the equilibrium
price, it is clear that price competition will pressure the retailers to price at their marginal cost c + γ.
Thus, in stage 1, the wholesalers price at their marginal cost c and earn zero profits. And, in stage 2, the retailers price at their Firm 4 marginal cost c + γ and also earn zero profits.
COCO station Let us now examine the effects of having COCO stations in the above model by replacing one retailer of each wholesaler by a COCO retailer. Say that R12 and R22 are COCO retailers. We will now denote the
former retailer by R1c and the latter retailer by R2c.
cost of a COCO retailer is equal to the Firm 4 marginal cost of a ‘standard’ retailer, i.e. it is equal to her wholesale price plus retail cost γ.
In this case, the profits function of W1 is given by:
Πw1(pw1) = (pw1− c)Dr(p11|p1c, p21, p2c)
+(p1c− c − γ)Dr(p1c|p11, p21, p2c), (3.9)
and the profits function of W2 is given by:
Πw2(pw2) = (pw2 − c)Dr(p21|p11, p1c, p2c) + (p2c− c − γ)Dr(p2c|p11, p1c, p21).
For this model we propose the following equilibrium:
Proposition. The wholesalers set their wholesale prices at their marginal cost: pw1 = pw2 = c. Further, they set their COCO retailers’ prices at their
marginal cost: p1c= p2c= c + γ. Subsequently, retailers R11 and R21 set their
prices at their marginal cost: p11= p12 = c + γ. In this equilibrium, each firm
earns zero profits.
Proof. We are now going to show, that, given pw2 = c and p2c = c + γ,
wholesaler W1 has no incentive to deviate from the equilibrium. If pw1 > c,
then R11 has a marginal cost of pw+ γ, and is unable to compete against R21
and R2c, as pw+ γ > c + γ. However, R1c is able to compete against R21 and
R2c, as its marginal cost is c + γ. Since price competition will pressure retailers
R1c, R21 and R2c to price at c + γ, it follows from (3.9) that Πw1(pw1) = 0.
And, if pw1 < c, then R11is able to undercut the other retailers, as pw+ γ <
Thus we can conclude that wholesaler W1 has no incentive to deviate from
the equilibrium. Further, given that the wholesalers are using their equilibrium prices, it is clear that price competition will pressure the retailers to price at their marginal cost c + γ.
Since the wholesale and retail prices are the same in the ‘standard model’ and ‘COCO model’, we can conclude that it is irrelevant for the model if a retailer is a ‘standard’ or COCO retailer.
3.3
Collusion
In this section we consider a special case of a cartel where the wholesalers earn higher profits than the competitive profits by coordinating their wholesale prices pw1 and pw2. Based on the trigger-price model of Tirole (1988) and Green and
Porter (1984), Borenstein et al. (1992) examine price coordination through a simple oligopoly model with two price-setting firms. They show that after a decrease in cost the firms may find it profitable to maintain their prior prices until demand conditions force a change. In this case, we will observe a rapid response of the prices to increases in cost and a slow response to decreases in cost, i.e. asymmetric price adjustments. Using this model, we modify our ver-tical structure model to examine if price coordination between the wholesalers can induce asymmetric price adjustments.
ran-dom variable. Let ˆDr(p11), ˆDr(p12), ˆDr(p21) and ˆDr(p22) denote the stochastic
demand for retailer R11, R12, R21 and R22, respectively, and define the
follow-ing: ˆ Drt11(pt11) = θ1tDr(pt11|pt12, pt21, pt22) ˆ Drt12(pt12) = θ2tDr(pt12|pt11, pt21, pt22) ˆ Drt21(pt21) = θ3tDr(pt21|pt11, pt12, pt22) ˆ Drt22(pt22) = θ4tDr(pt22|pt11, pt12, pt21),
where θ1t, θt2, θt3 and θ4t (t = 1, 2, . . .) are identical and independent random variables. In this case, in period t, the demand function of wholesaler W1 and
W2 are given by ˆDtr11(p t 11) + ˆDrt12(p t 12) and ˆDtr21(p t 21) + ˆDrt22(p t 22), respectively.
Further, we assume that the retailers do not know the value of θ1t, θt2, θt3 and θ4t (t = 1, 2, . . .).
Following Borenstein et al. (1992), we also suppose that each wholesaler has a trigger strategy: a wholesaler return to the competitive price if her demand falls below threshold ¯q > 0.
The demand of a wholesaler decreases if her competitor undercuts the co-ordinated price or if there is a decrease in the consumer demand. We assume that a firm cannot distinguish between the two causes3.
Say that the wholesalers had in the previous period (t = 0) a marginal cost of c0 and used their competitive price: pw0 = c0. Suppose now that in
the current period (t = 1) their cost decreases from c0 to c1. We assume that
hereafter the cost is constant at c1. If, in the current period, both wholesalers
3
use prior price c0, the current expected profits of W1 and W2 are: Πw1 := (c0− c1)E( ˆD 1 r11(c0) + ˆD 1 r12(c0)) and Πw2 := (c0− c1)E( ˆD 1 r21(c0) + ˆD 1 r22(c0)),
respectively. In the next period (t = 2), there is a probability that the demand of a wholesaler is below the threshold, which would induce the wholesalers to price competitively and, thus, earn zero profits. Otherwise, in the next period, the wholesaler faces the same situation again4. Thus the expected current and
discounted future profits of W1 are given by:
V (c0) = Πw1+ δ(1 − α)V (c0), (3.10)
where 0 ≤ δ ≤ 1 is the discount factor and 0 ≤ α ≤ 1 is the probability that the demand of W1 or W2 at price c0 is lower than threshold ¯q, i.e.
α = Pr{ ˆDtr11(c0) + ˆD t r12(c0) < ¯q or ˆD t r21(c0) + ˆD t r22(c0) < ¯q}.
Note that, since θt
1, θt2, θt3 and θ4t (t = 1, 2, . . .) are identical and independent
random variables, we have that:
Pr{ ˆDtr11(c0) + ˆD t r12(c0) < ¯q or ˆD t r21(c0) + ˆD t r22(c0) < ¯q} = Pr{ ˆDt+1r11(c0) + ˆD t+1 r12(c0) < ¯q or ˆD t+1 r21(c0) + ˆD t+1 r22 (c0) < ¯q}
Thus the probability that the demand of a wholesaler is below the threshold is the same in each period.
Solving (3.10) for V (c0) gives:
Vw1(c0) =
Πw1
1 − δ(1 − α).
4
As θt1and θt2(t = 1, 2, . . .) are identical and independent random variables and we assume
By symmetry, we have that the expected profits of W2 is given by:
Vw2(c0) =
Πw2
1 − δ(1 − α).
We will now show that W1 has no incentive to deviate from the equilibrium
price c0, given that in the current period W2is using the equilibrium price c0and
the cost for both wholesalers is c1. Let us determine the gain of undercutting. If
W1uses a price that is slightly lower than equilibrium price c0, then her retailers
have a marginal cost that is lower than the marginal cost of W2’s retailers. This
means that the retailers of W1 are able to undercut the retailers of W2and serve
the entire market. In this case, the profit of W1 is equal to 2Πw1. However,
undercutting W2induces her demand to be zero, which triggers the wholesalers
to use the competitive price infinitely. Thus W1 has no incentive to undercut
W2 if5:
Πw1
1 − δ(1 − α) ≥ 2Πw1.
By rewriting the above, we find that W1 has no incentive to undercut W2 if
δ ≥ 2(1−α)1 . As Borenstein et al. (1992) note, given the high frequency of
adjustment in the gasoline market, it is reasonable to assume that δ is close to one. Hence, W1 has no incentive to undercut W2 if α is not too large. We do
expect the latter, since a large α means a high probability that the wholesalers use the trigger strategies, and, thus, low expected collusive profits.
Additionally, if c0is an equilibrium, W1should not have the incentive to use
a price that is higher than c0. If W1using a higher price, then the retailers of W2
5
are able to undercut the retailers of W1and serve the entire market. In this case,
W1earns zero profits. Thus we can conclude that W1has no incentive to deviate
from equilibrium price c0. Due to symmetry, we have that the same holds for
W2. This oligopolistic coordination implies that the wholesalers may maintain
their prior prices when the cost decreases until a demand shock occurs that triggers the wholesalers to use their current competitive price. However, when the cost increases, the wholesalers will surely increase their prices immediately, as the wholesalers would make a loss if they keep charging the prior price (which is lower than their current cost).
The above model implies that if the wholesalers are coordinating their prices, we will observe a rapid response of the prices to increases in cost and a slow response to decreases in cost, as in the latter case they will try to maintain their prior prices. Thus we will observe asymmetric price adjustments on the wholesale level if the wholesalers are coordinating on their prices.
Since the prices of the retailers directly depend on the wholesale prices that the wholesalers set, having asymmetric price adjustments on the wholesale level will induce asymmetric price adjustments on the retail level. Thus finding asymmetric price adjustments on the retail level, indicate that there is evidence of price coordination among the wholesalers.
For a general overview of possible explanations of asymmetric price adjustments, see Meyer and Cramon-Taubadel (2004).
Chapter 4
Empirical testing
In Section 3.3 we have shown how a particular form of price coordination among the wholesalers will cause asymmetric price adjustments. Since we do not gen-erally expect this behavior in a competitive market, we can use this prediction to test if there is any sign of collusive behavior.
4.1
The data
Note that in our wholesale price coordination model we did not consider the fact that each wholesaler recommends its retailers to use a certain price. We find it plausible that the recommended prices help the retailers to coordinate their prices if there is heterogeneity among the retailers (e.g. it may be the case that the retailers have different retail costs)1.
Since price coordination among the wholesalers induce asymmetric adjust-ments between the ex-refinery price and the wholesale price, we conjecture that, if there is price coordination among the wholesalers, the recommended price also adjust asymmetrically to the ex-refinery price. We find the latter plausi-ble, because we expect that the recommended price is based on the wholesale price.
This raises the question if we could also use the recommended price to determine if there is price coordination among the wholesalers. But, since most stations follow the recommended price or a price that is very close to the recommended price, the recommended price and retail price are practically the same. Further, we will find that using the retail price or the recommended price to estimate the econometric models gives the same estimation results.
Further, in Chapter 1 we have argued that we can use the reference Platts as a substitute for the ex-refinery price (which corresponds to the ‘marginal cost’ of the wholesalers in Chapter 3), as we do not have any data on the ex-refinery price.
1
4.2
The model
The ECM is built on the long-run relationship between the retail price and reference Platts (from now on we will write Platts in place of reference Platts). We will firstly determine the long-run relationship, and, subsequently, we will use this long-run relationship to construct the ECM. Before we begin with the analysis of the long-run relationship between the retail price and the Platts, we are going to examine if the series are non-stationary. If we find that the series are non-stationary, then we should be careful with directly applying OLS and should determine if there is a cointegrating relation between the series. To test for non-stationarity have conducted the Augmented Dickey-Fuller Unit Root Test with intercept. Its output is:
t-Statistic MacKinnon p-value
Retail price -1.962 0.303
Platts -2.264 0.185
Since the p-values are higher than 5%, we cannot reject the null-hypothesis that the series are non-stationary.
the Platts2.
Since there is a cointegration relation between the the retail price and the Platts, we will proceed with the estimation of a linear relation between the series with OLS, as, in this case, OLS is a consistent estimator, see, e.g. Green (2000) p.795. Consider the following linear relation for the long-run equilibrium, as suggested by Borenstein et al. (1997):
Rt = c + β Pt+ 11
X
j=1
DjMj+ et (4.1)
where Rt is the retail price in week t, Pt is the Platts in week t, c is a
constant, P11
j=1DiMj are the month dummies (where there is an indicator
variable for each month except June) and et is the error term.
Notice that the equation is in levels and not in logarithms. However, it could be the case that the markup, i.e. the difference, between the retail price and Platts varies with the Platts, which would indicate that we should take the logarithm of the variables. In Figure B.1 we can find a plot of the retail price markup and the Platts. In the long-run, the graph shows that the Platts increases steadily, but the markups do not increase in the long-run. However, due to the shortness of the sample period, we can not robustly test this assertion. Further, there could be seasonality in the relationship, as gasoline demand may vary with the seasons. To control for this, we have included month dum-mies in the equation. The OLS estimation of (4.1) is given in Table 4.1.
The Durbin-Watson value of the estimation is significantly lower than 1.5,
2
Dependent var:Rt
Retail price on Platts Equation Coefficient T-statistic
c 11.812 4.05 Pt 1.024 15.261 M1 3.483 1.652 M2 2.129 1.011 M3 1.833 0.722 M4 -0.099 -0.049 M5 -1.061 -0.521 M7 -2.371 -1.278 M8 -0.568 -0.317 M9 1.331 0.509 M10 2.951 1.551 M11 4.574 2.616 M12 5.637 2.681 R2 0.862 Durbin-Watson 0.238 Jarque-Bera 4.833
Method OLS, Newey-West T-statistic
Observation 140
Table 4.1: The OLS estimate of the long-run equilibrium.
which indicates that the residuals are positively autocorrelated. Because of this autocorrelation, we have used the Newey-West method to estimate the T value, see, for example, Green (2000) p. 537.
To check if the residuals are normally distributed, we have conducted the Jarque-Bera Test. At the 5% level, we cannot reject the null-hypothesis that the residuals are normally distributed.
increases and decreases in their gasoline costs, and a pass through-rate that is higher than one means that the retailers amplify the increases and decreases in their costs.
The only month dummies that are significant are for the months November and December. It could be the case that the Portuguese gasoline demand cause this variation. However, as we do not have any data on the Portuguese gasoline demand, we cannot examine this possible seasonality in greater detail.
Error Correction Model Now that we have determined the long-run equi-librium between the retail price and the Platts, we can construct the ECM. We want to incorporate in the ECM the reversion of the retail price to its long-run equilibrium and the contemporaneous and lagged changes in Platts and its own lagged changes, as it is likely that they influence the retail price. This results in the following ECM for the retail price, as suggested by Borenstein et al. (1997):
∆Rt = m X k=0 βk∆Pt−k+ n X k=1 θk ∆Rt−k + λ zt−1+ et, (4.2)
where ∆Rt denotes the variation in the retail price, Pmk=0∆Pt−k denotes the
contemporaneous and past variations in Platts,Pn
k=1∆Rt−k denotes the past
variations in the retail price and zt−1=:
Rt−1− c − βPt−1−P11j=1D1jMj
denotes the reversion of the retail price to its long-run equilibrium.
To examine asymmetric price responses, we modify the above model to:
+λ+ (zt−1)+
+λ− (zt−1)−+ et, (4.3)
where (∆Pt−k)+ (k = 1, . . . , m) denotes the increases in the Platts and
(∆Pt−k)− (k = 1, . . . , m) denotes the decreases. These variables are the most
interesting, as differences between their coefficients i.e. between βk+ and βk−,
will imply that the retail price reacts differently to increases and decreases in the Platts. We have also included (∆Rt−k)+ (k = 1, . . . , n), which denotes
the increases in the retail price, and (∆Rt−k)− (k = 1, . . . , n), which denotes
the decreases. These variables make the model richer and differences in their coefficients will imply that lagged increases and decreases in the retail price affect the contemporaneous retail price differently.
We divide the equilibrium errors in a similar way in positive and negative deviations and denote them by (zt−1)+ and (zt−1)−, respectively. Borenstein
et al. (1997) do not divide the equilibrium errors, but we find it interesting to examine if there is also asymmetry in the speed that the retail price reverts to its long-run equilibrium. Further, m and n are the levels of lag that we still have to determine.
Dependent var: ∆Rt Dependent var:∆At
Equation Coefficient T-statistic Coefficient T-statistic
(zt−1)+ -0.070 -1.438 -0.079 -1.229 (zt−1)− -0.086 -2.178 -0.060 -1.304 (∆Pt)+ 0.360 4.485 0.528 3.516 (∆Pt)− 0.207 2.924 0.147 1.936 (∆Rt−1)+ 0.083 0.576 (∆Rt−1)− 0.253 1.205 (∆At−1)+ -0.034 -0.205 (∆At−1)− 0.308 1.432 R2 0.263 0.295
Lagrange multiplier test 1.601 1.456
Observations 140
Method OLS, White Heteroskedasticity-Consistent
Table 4.2: The OLS estimate of the Error Correction Model.
model; they linearize the model, and then they estimate it with OLS. Their estimation method is not a standard method and little is known about its properties (Bachmeier and Griffin (2003)). Since the two-step method is a well-known and easy method to estimate an ECM, we have used this method to estimate (4.3).
The estimations are given in the first column of Table 4.2. We have chosen the lag levels through the Akaike Information Criteria. The estimation has a R2 of 0.263. Since our dependent variable is differenced, we do not expect to
obtain a high R-squared value (e.g. compared to a model in levels). To test for autocorrelations, we have performed the Lagrange multiplier (LM) test up to lag order 4. Since its p-value is 0.178, we can conclude that the residuals are not autocorrelated. Further, the residuals are asymptotically normally distributed, as we have a long sample period (140 weekly observations).
coefficients. To be more exact, we have conducted a Wald-test under the null-hypothesis that there is no asymmetry. Since the test has a p-value of 0.507, we accept the null-hypothesis that there is no asymmetry.
As we have noted in the previous section, we do not expect that there is much difference in the estimation if we use the retail or recommended prices as the dependent variable. To be sure, we have re-estimated the above model with the recommended price as dependent variable. Its result is given in the second column of Table 4.2, where we have denoted the recommended price by At. We
observe that there is indeed not much difference between the two estimations. Further, we have conducted a Wald-test under the null-hypothesis that there is no asymmetry. Since its p-value is 0.107, we accept the null-hypothesis that there is no asymmetry.
Endogeneity It could be the case that there is a third unobserved variable that is correlated with the retail price and Platts. In this case, ∆Pt and zt−1
are endogenous regrFirm 5rs in (4.2). To test for this possible endogeneity, we have identified three possible instruments. For ∆Pt we have identified the
lagged changes in the Brent crude oil price (denoted by ∆Bt−1). For zt−1 we
have identified the lagged Brent crude oil price (denoted by Bt−1), as zt−1 is
a function of Pt−1. Further, since ∆Rt−1 is exogenous, it can be used as an
instrument for itself3.
Since the Brent oil is sourced from the North-sea and its price is determined by the world market, it is very likely that it is exogenous to (4.2). In addition to
being exogenous, our instruments should also be correlated with the variables being instrumented. Following Cameron and Trivedi (2005) (p.104), we use Shea’s partial R2 to determine if instruments Bt−1 and ∆Bt−1 are correlated
with regrFirm 5rs zt−1 and ∆Pt, respectively4.
For zt−1we received a partial R2 of 0.251 and for ∆Pt we received a partial
R2 of 0.451. We can conclude that both instruments are reasonably correlated with their respective regrFirm 5rs.
Using these instruments, we conducted a Hausman test for endogeneity of regrFirm 5rs to determine if there is endogeniety in (4.2). See for example Cameron and Trivedi (2005) (p.276) for a description of the Hausman test. Under the null-hypothesis of no endogeneity, we received a p-value is 0.447, which indicates that we can accept the null-hypothesis of no endogeinity. Thus, our standard OLS estimation of the ECM is well defined.
4.3
Discussion
Let us compare our finding with Bachmeier and Griffin (2003) and Borenstein et al. (1997). The former did not find any evidence in support of asymmetric price adjustments, however the latter did find evidence of asymmetric price adjustments. We have to note that there are some differences between our model and theirs: Bachmeier and Griffin (2003) examine the response of the daily retail price to daily crude oil price and Borenstein et al. (1997) examine the response of the semi-monthly retail price to the semi-monthly wholesale and crude oil price and use, as said before, a different estimation method.
4
However, it is unclear to us whether the choice of data frequency effects the measurement of asymmetric price responses. In any case, we can say that our study reinforces the finding of Bachmeier and Griffin (2003), i.e., similarly to the study of Bachmeier and Griffin (2003) of the American gasoline market, we have not found any evidence in support of asymmetric price adjustments in the Porto market.
Chapter 5
Conclusion
In this thesis we have examined the Porto gasoline market. We found that most stations follow their recommended prices and that there is not much difference between the recommended prices. This results in a retail market with a low level of price dispersion. Further, we found no evidence of a simple price-fixing scheme among the retailers or wholesalers.
Subsequently, we showed, in a theoretical model, how price coordination among the wholesalers, where they maintain their prior prices even when the refinery price decreases, induces asymmetric price adjustments between the ex-refinery and retail/wholesale price. Thus our model underpins the relationship between asymmetric price adjustments and collusion.
Appendix A
Chapter 1
Appendix B
Chapter 2
Appendix C
Chapter 3
Bibliography
AdC (2008). Relat´orio sobre o Mercado dos Combust´ıveis em Portugal. Au-toridade da Concurrˆencia, report to Ministro da Economia e da Inova¸c˜ao.
Bachmeier, L. and J. Griffin (2003). New evidence on asymmetric gasoline price responses. The Review of Economics and Statistics 85, 772–776.
Borenstein, S., C. Cameron, and R Gilbert (1992). Do gasoline prices respond asymmetrically to crude oil price changes? NBER Working Paper .
Borenstein, S., C. Cameron, and R Gilbert (1997). Do gasoline prices respond asymmetrically to crude oil price changes? The Quarterly Journal of Eco-nomics 112, 305–39.
Cameron, A. and P. Trivedi (2005). Microeconometrics. Cambridge.
EC (1996). Case IV/M.727-BP/Mobil. Commission of the European Commu-nities.
Engle, R. and C. Granger (1987). Co-integration and error correction: Repre-sentation estimation and testing. Econometrica 55, 251–276.
Geweke, J. (2004). Issues in the ”Rocket and Feathers” gasoline price litera-ture. http://www.ftc.gov/bc/gasconf/comments2/gewecke2.pdf. Report to the Federal Trade Commission.
Green, E. and R. Porter (1984). Noncooperative collusion under imperfect price information. Econometrica 52, 87–100.
Green, W. (2000). Econometric Analysis (fourth ed.). Prentice Hall.
Hosken, D., R. McMillan, and C. Taylor (2008). Retail gasoline pricing: What do we know? Forthcoming in International Journal of Industrial Organiza-tion. Federal Trade Commission Working Paper.
Kovenock, D. and K. Widdows (1998). Price leadership and asymmetric price rigidity. European Journal of Political Economy 14, 367–187.
Lewis, M. (2004). Asymmetric price adjustment and consumer search: An examination of the retail gasoline market. Competition Policy Center, Uni-versity of California, Berkeley, Working Paper .
Meyer, J. and S. Cramon-Taubadel (2004). Asymmetric price transmission: A survey. Journal of Agricultural Economics 55, 581–621.
MG (2008). Competition Bureau reveals gas price fixing scam in Quebec. http: //www.montrealgazette.com. The Montreal Gazette.
Shea, J. (1997). Instrument relevance in multivariate linear models: A simple measure. The Review of Economics and Statistics 79, 348–352.
