Price settings in online markets: Basic facts, international comparisons, and cross border integration *

Transcription

1 Preliminary and incomplete Comments welcome Do not cite without authors permission Price settings in online markets: Basic facts, international comparisons, and cross border integration * Yuriy Gorodnichenko UC Berkeley Oleksandr Talavera Durham University Abstract This paper documents basic facts about prices in online markets in the U.S. and Canada which is a rapidly growing segment of the retail sector. While less rigid than prices in brick-and-mortar stores, online price exhibit significant stickiness (price spells can be as high as one month of average). Furthermore, there is significant dispersion of prices even for identical goods within and across countries. The paper shows that for online markets price differentials across countries are adjusted faster (half-life of about 4 months) and the pass-through of nominal exchange rate is larger (approximately 35%) than the corresponding counterparts of goods sold in regular stores. Degree of competition, stickiness of prices, synchronization of price changes, and returns to search effort are important determinants of pass-through and speed of price adjustment for international price differentials. JEL: E3, F3, F40, F41 Keywords: Online market, prices, pass-through, border effects, law of one price * We thank Google Inc. for financial support (Google Research Award N.). Gorodnichenko also thanks NSF for financial support. We thank Viacheslav (Slavik) Sheremirov for excellent research assistance. Contacts: Gorodnichenko, ygorodni@econ.berkeley.edu; Talavera, oleksandr.talavera@googl .com.

2 1. Introduction E-commerce is a rapidly increasing segment of the retail market. For example, the US Census Bureau estimated that total e-commerce sales for 2009 were $134.9 billion (or approximately 4 percent of total sales in the US economy). 1 This represents an increase of 2 percent from 2008 during a period in which total retail sales declined by 7 percent. Not only have online sales risen during the past recession, but they have historically grown much faster (10 or more percent) than sales of brick-and-mortar stores. Furthermore, Forrester research, an independent technology and market research company, predicts that by 2014 online sales will account for more than 8 percent of total retail sales. 2 However, while online prices account for a significant and rapidly expanding share of sales and can shed new light on the sources of price stickiness and the characteristics of highly integrated market, the properties of online prices are still relatively understudied. This paper is aimed to document and understand properties of internet prices in the U.S. and Canada for a broad array of products including printed media, movies/music, electronics, tools, furniture, toys, recreational goods, automotive, and others. Our objectives are threefold: i) systematic analysis of price dispersion; ii) analysis of properties of price spells and price changes; iii) international price comparisons. The first objective (price dispersion) is important for a number of reasons. For example, a conventional view is that internet markets are frictionless (or efficient) and by observing behaviour of these markets one can make predictions about how sellers and prices will behave when various trade frictions are eliminated. In these environments consumers are aware of product price and quality, so that internet sellers should set one price and make zero profit. However, anecdotal evidence and some previous studies reveal substantial variations in internet prices in any given point in time. Our contribution would be to provide a systematic approach that covers many precisely identified goods in a wide range of product categories which contrast with previous literature where price dispersion was documented for special types of goods (e.g., books) or heterogeneous goods within a relatively narrow product category (e.g., office staplers rather than unique product codes). We document that even for very narrowly defined goods price dispersion is substantial and persistent. For example, the average standard deviation of log prices in a given week for a precisely defined good (bar-code level) is approximately We also show that dispersion of prices is systematically related to proxies of returns to search intensity and to proxies of competition as well as reputation of sellers. The second objective of this paper is to report the frequency and size of price changes in e- commerce and thus complement influential studies by Bils and Klenow (2004) and Nakamura and Steinsson (2008) presenting the same information for regular brick-and-mortar stores. Note that in contrast to regular stores there is really no physical menu cost to price changes 1 Source: 2 These patterns are very similar in other developed countries. For example, according to the Centre for Retail Research, online retail sales in Europe jumped 20 percent this year, far outstripping the 1.4 percent growth in store-based sales. Furthermore, the share of online sales in total sales is larger in Europe than in the USA. For instance, the share is 9.5 percent in the U.K.

3 and therefore internet prices could be fluctuating every instant (minute, day, week) in response to shocks to demand and supply conditions. Thus by comparing the duration of price spells and the properties of price changes with their counterparts from brick-and-mortar stores, one can infer the importance of physical costs of price changes. We find that the size of price changes in online stores (approximately 5%) is at most half of the size of price changes in regular stores (approximately 10%). We also find that price changes occur much more frequently in online stores (approximately once every 4 weeks or less) than in regular stores (once every 4 months or more). This evidence is consistent with the view that online prices are much more flexible than prices in regular stores. However, the fact that we still observe some rigidity in online prices suggests that the costs of changing prices are more complex than just physical menu costs and instead are likely to involve costs of gathering and processing information as well as (potentially) coordinating price changes with other sellers. For the third objective, by comparing identical goods across two countries, we can study the properties of the prices (and potentially purchases) when consumers and sellers are located in different countries. For example, with online prices, we can shed new light on why passthrough of exchange rate fluctuations and reversion to the law-of-one-price are very weak in international data and thus constitute one of the central puzzles in international economics (Obstfeld and Rogoff 2000). The key insight again is that since changes in online prices are unlikely to have any physical costs and goods in online trade are easy to ship and consequently the physical location of the seller is not important, the pass-through could be quick and nearly complete while it is slow and partial in the prices of regular stores because of the frictions associated with trade flows and mobility of buyers. We find that the passthrough in online markets is incomplete but at the same time is considerably larger than in regular markets and stands at about 40 percent. There is, however, significant heterogeneity in pass-through across goods and we document that for goods with certain characteristics (e.g., big-price-tag goods) the pass-through can be close to 100 percent. The speed of price adjustment is also substantially faster in online markets (half-life is about 4-5 month) than in regular markets (half life varies from 3 quarters to a few years). We also document that the size of the pass-through and the speed of price adjustment are systematically associated with the degree of price stickiness, turnover of sellers, returns to search (specifically, the level of prices), and the degree of competition as well as dominance of large online sellers. These results can help reconcile the heterogeneity of estimated pass-throughs and speeds of adjustment across studies. Thus, in summary the stylized facts we document about prices in online stores relative to prices in regular, brick-and-mortar stores are: 1) the dispersion of prices is comparable; 2) the duration of price spells is short; 3) the size of price change is smaller; 4) pass-through is larger and the speed of price adjustment is faster. Most of these characteristics are consistent with reduced frictions in online markets and much great integration of markets than is documented for regular stores and thus can inform policymakers and researchers about what one may expect to observe when frictions in regular markets are reduced. To document and study the properties of online prices, we have constructed a unique data set of price quotes. Specifically, for over three years we have been scraping prices and other 2

4 relevant information from a leading price comparison website. The data include good s unique identifier (similar to barcodes in the scanner price data), good s description, prices for each seller, seller's unique identifier, number of seller reviews, ranking of seller, reviews of goods, etc. The data cover a broad range of goods sold online. Information for more than 100,000 goods and more than 10 million price quotes has been collected. There are several advantages of using our data. First, the time span (more than three years) is considerably larger than the time span usually available for researchers studying online prices (typically less than a year). This dimension is important when we study dynamic properties of prices such as duration of price spells, speed of price convergence, and pass-through. Second, the coverage of goods is much broader than in previous analyses of online prices which typically focused on books and CDs. The latter types of goods are easy to compare across sellers or countries but they also have a number of unusual properties which make generalizations tricky. Our data set is heavily populated by durable goods which tend to be under-represented in typical scanner price data and which are much more likely to be traded and moved across even distant locations. Third, we have been collecting prices for identical goods in the U.S. and Canada and thus comparison of prices is direct and simple. Thus, we can avoid a number of pitfalls associated with comparing goods that are only broadly similar or comparing price indexes. Fourth, we have been collecting information on important attributes such as the quality of sellers and goods as revealed by ratings of sellers and products. We can use these attributes to explore the determinants of properties of online prices. In contrast, previous research on basic properties of prices had only very limited (if any) information about properties of goods for which prices were available. Fifth, our data include many sellers rather than one retail chain and therefore we can assess the relative importance of different sources of price variation (e.g., store-level, country-level, etc.). This multi-seller dimension is also important because branches of a single seller are less likely to engage in competition between each other than branches of different sellers. Finally, the high frequency of our data allows us to time reactions of prices to other high frequency events such as changes in the exchange rate or natural experiments thus making identification more clear-cut. This paper is related to several strands of research. The first strand is focused on assessing whether the law of one price (or a milder version of it known as purchasing power parity (PPP) hypothesis) holds and how quickly deviations from the law of one price are eliminated. The early generation of this literature could use only price indexes collected at the country or regional level, which lead to a number of practical and conceptual issues with the interpretation of the results. Rogoff (1996) summarizes this literature as documenting that PPP is likely to hold in the long run but it takes a long time for prices to converge to the PPP (half-life is routinely estimated to be over a year and in most cases multiple years). This literature also found that deviations from PPP can be quite large and heterogeneous across countries and time (e.g., Takhtamanova 2008, Campa and Goldberg 2005, Barhoumi 2005) which can be only partially explained by sticky prices and exchange rate regimes, which 3

5 constitutes the PPP puzzle. Campa and Goldberg (2005) report that the pass-through for the U.S. is about 46% over one quarter and 64% over one year. 3 Data limitations of the first literature lead to the second generation of studies which focused on using micro-level price data to measure more precisely pass-through and the speed of price adjustment. Imbs et al. (2005), Cruchini and Shintani (2008), Broda and Weinstein (2008), and others showed that the pass-through and the speed of price adjustment are higher when individual, narrowly-defined goods are considered. For example, the half-life of price adjustment falls to about a year. Furthermore, these papers demonstrate that the PPP puzzle observed in price indexes can be explained at least to some extent to aggregation biases. Because the micro-level price data are richer and easier to compare across countries, the second generation studies can also provide a more detailed account of the relative importance of producer-currency pricing (PCP) versus pricing-to-market (PTM) hypotheses. For example, Gopinath et al. (2001) find support for the PTM hypothesis in explaining incomplete pass-through in the weekly scanner (barcode level) price data for a retailer selling goods in both the U.S. and Canada. We contribute to this literature by examining the behaviour of prices at the level of precisely defined goods sold by multiple stores in different countries. Easier access to micro-level price data also allowed to explore the determinants of passthrough and speed of price adjustment. For example, Menon (1996), Cardasz and Stollery (2001), Gaulier, Lahreche-Revil, and Mejean (2006), Goldberg and Hellerstein (2007), Mayoral and Gardea (forthcoming) relate market structure, market power (including adjustment of mark-ups), tariffs, presence of multinationals and importance of non-traded inputs for price stickiness of final goods and the size of the pass-through. We contribute to this literature by exploring the determinants of pass-through and speed of price adjustment for online markets. The third strand of research we contribute to is focused on documenting price rigidities at the micro-level which could be used later to calibrate macroeconomic models. Recent examples of this literature are Bils and Klenow (2004) and Nakamura and Steinsson (2008). Studies in this literature concentrate almost exclusively on prices collected in regular, brick-and-mortar stores. In contrast, we focus on online prices which describe a rapidly growing part of consumer shopping. One may anticipate that online prices will play an increasingly important role in the future and, thus, macroeconomists should incorporate properties of a broader set of goods, including goods sold online, when they characterize micro-foundations of their macroeconomic models. The fourth strand documents basic facts about properties of online prices. In a study representative of this literature, Brynjolfsson and Smith (2000) compare online and conventional stores prices on books and CDs. They find that online prices are 9-16% smaller, the changes in online prices are much smaller for online prices, yet quotes of internet prices are quite dispersed. Much of the subsequent literature tried to (mostly theoretically) 3 This literature also related trade frictions to the level of price dispersion. See Engel (1993), Engel and Rogers (1996) and Gorodnichenko and Tesar (2009). 4

6 explain dramatic dispersion of prices in online markets (e.g., Baye and Morgan 2001, Baye and Morgan 2004, Morgan, Orzen, and Sefton 2006, Baye and Morgan 2009) by information frictions (e.g., bounded rationality), sellers ability to discriminate consumers (e.g., based on what sellers know about customers), and differences in advertisement (e.g., investment in building brand, reputation, etc.). 4 We provide evidence that considerable price dispersion in online markets applies to a very broad set of goods. The two previous studies most relevant for our paper are Lunnemann and Wintr (2011) and Boivin et al. (2012). Lunnemann and Wintr (2011) document stickiness of online prices in the U.S. and large European markets (Germany, France, Italy, and the U.K.). They find that internet prices change less often in the U.S. than in Europe (the opposite is true for conventional stores). Online prices are more flexible than their offline counterparts with half of the spells ending within a month. However, the result is not universal as it does not hold for a small subset of goods, nor it is homogeneous across shops, which can exhibit significant differences. In contrast, Boivin et al. (2012) focus on the dynamics of online price differences across three book sellers: Amazon.com (and Amazon.ca), BN.com (Barnes & Noble website), and Chapters.ca. They find that price differentials (or relative quantities) for books react to fluctuations in the relative price of foreign competition following exchange rate movement which is consistent with extensive market segmentation and pervasive violations of the law of one price. We merge these two lines by exploring a much larger set of goods, using longer time series, exploiting significant movements in the nominal exchange rate, and investigating determinants of observed pass-through and speed of price adjustment. The rest of the paper is structured as follows. In the next section we describe the data set and how it was collected. In this section, we also document the basic properties of online prices. In Section 3, we estimate the pass-through and the speed of price adjustment for online prices. In addition, we explore the determinants of the size and speed of price adjustment in response to changes in the nominal exchange rate. In Section 4, we validate our results using a natural disaster which exogenous reduced the supply of hard drives and thus affected the costs of laptops, which are major goods sold online. In Section 5, we discuss our results and make concluding remarks. 2. Data Description A. Data collection This study uses data collected from a price comparison site which provides price quote data for two countries: USA (.com domain) and Canada (.ca domain). This website is one of the leading price comparison websites in both countries. The U.S. part of the website was among 4 For example, Deck and Wilson (2006) study discrimination based on search history. Search history can provide important information on how informed a customer is about prices for a given good. They show that discriminating and non-discriminating firms charge their customers the same prices in aggregate. However, discriminating firms set lower prices for informed customers and higher prices for uninformed ones, which leads to observed price dispersion. 5

7 top 10 Web portals based on total unique visitors in January Styles of pages with price quotes are similar across countries which simplifies data extraction and identification of exactly the same products listed by Canadian and U.S. sellers. Identifiers for goods listed on the website are similar to barcodes used in the analysis of scanner price data. For example, manufacturing product number (MPN) 0S03110 uniquely identifies Hitachi Touro Mobile Pro Portable External 750 GB 2.5" Hard Drive. Although the price comparison platform we use has similar websites in other countries, we limit the set of countries to U.S. and Canada. This choice is motivated by several reasons. First, the link between U.S. and Canadian websites greatly simplifies linking goods across countries. Second, trade flows are more likely to be affected by transatlantic shipping costs, language differences, etc. Finally, we want to study countries with strong trade ties and relatively small trade frictions such tariffs, regulations, etc. In contrast to a few comparable studies that study properties of online prices and typically have up to one year of data (e.g. Lunnemann and Wintr 2011), our data cover period of more than three years. The data collection was launched on November 16, 2008 and continues until present. Importantly, this timeframe includes a period of significant appreciation of the Canadian dollar from 1.30 in the end of 2008 to 0.95 in the middle of 2011 (see Figure 1). Longer time series combined with significant changes in the exchange rate will help us to obtain precise estimates. Every Saturday at midnight, a Tcl/python script has been triggered to collect webpages with price information. The script has several stages. First, it collects information on the universe of goods available for a given type of goods on the comparison website. For each good, there exists a link to a unique webpage with price quotes. The script constructs a dictionary of goods and associated links. Second, the script follows the links and downloads webpages with price quotes. It usually takes about 24 to 48 hours to download a complete set of pages for all goods in targeted categories. Third, after the webpages are downloaded, the python part of the script extracts good description, unique manufacturing product number (MPN), prices for each seller, and sellers' unique ids from every webpage. In the end, we have obtained information for more than 140,000 goods and 11 million goodseller-week-country quotes. Our price data covers 56 types of goods in four main categories: computers (20 types, e.g., laptops), electronics (14 types, e.g., GPS), software (12 type, e.g., computer games), and cameras (10 types, e.g., digital cameras). Table 1 present a list of categories and types of goods in our sample. This selection of goods, the length of the time sample, and variation in exchange rates in our time sample provides us with a number of advantages relative to what researchers used in previous studies. First, our data set covers a diverse set of goods while the vast majority of papers on online prices (e.g., Brynjolfsson and Smith 2000, Boivin et al. 2012) almost exclusively focused on books or CDs for which it was relatively easy to ensure that the same good is compared across sellers. Prices of these goods have, however, a number of unusual properties such as 5 Comscore, January

8 very long spells of constant prices. Furthermore, the market for books and CDs is dominated by a handful of major sellers such as Amazon.com and Barnes&Noble. Thus, it may be hard to generalize results beyond books and CDs. The diversity of goods in our sample will be essential when we study determinants of the size of exchange rate pass-through and the speed of adjustment in relative prices. Second, a great deal of research studying the law of one price has been using data on goods for which transaction costs for cross-border purchases are likely to outweigh even large departures from the law of one price. For example, consumers are unlikely to take advantage of arbitrage opportunities in grocery products which are typically available in the scanner type price data or cost-of-living surveys (e.g., Economist Intelligence Unit). In contrast, we focus on goods for which transaction costs are small and consumers are essentially free to exploit even small arbitrage opportunities. Goods in our sample are durable, standardized, and easy to ship. Also, most goods in our sample are produced outside U.S. or Canada and marginal cost shocks can be differenced out when we take ratio of Canadian and U.S. prices. Third, goods in our data are precisely defined and therefore one can be more certain that he or she compares prices of the same good when he or she contemplates a purchase. This contrasts with previous research using price indexes or prices for broadly defined goods (e.g., toothpaste). Fourth, our data set collects price quotes from multiple sellers while previous research typically used micro-level price data from one seller (e.g., because scanner price data are supplied by one retail chain). This aspect is important because branches of the same seller in different countries (e.g., Amazon.com and Amazon.ca) are less likely to compete with each other than outlets of different sellers (e.g., Amazon and BHphotovideo). Finally, data are collected at weekly frequency which allows us to study responses of prices at relatively high frequency and makes identification cleaner. B. Data filters Because price data are extraordinarily heterogeneous in our sample, we apply a series of filters to minimize the effects of missing values, extreme observations and the like. Specifically, for the sample we use to study international price differentials we constrain the sample of goods only to products which are sold in both the U.S. and Canada. We winsorize prices of goods at top and bottom 1 percent. For time series analyses focused on dynamic responses or control for lagged values of variables, we keep only goods with at least twenty consecutive observations. The treatment of missing values in time series of price quotes for a given seller/good combination can influence the estimated degree of price rigidity (e.g., duration of price spells, size and timing of pass-through). We consider three commonly used procedures to deal with missing values (see Coibion, Gorodnichenko and Hong (2011) and Hong and Li (2011) for a discussion). The first procedure ( A ) assumes that a price spell ends when there is a price change or a missing value. The second procedure ( B ) combines spells on both sides of a 7

9 missing spell provided the price before and after the missing spell is unchanged. Suppose we observe a price of $1 during weeks 2 to 3 and the price for weeks 4 to 6 are missing, but we observe a price of $1 for week 7 followed by $1.5 for week 8 and $1.4 for week 9. The length of the ($1) spell is 2+1=3 weeks. The third procedure ( C ) imputes the previously observed price to all missing values. In the example above, this means that we include weeks 4 to 6, resulting in a ($1) spell length of = 6. We will take calculations based on procedure C as the benchmark. 3. Basic facts about price setting in online markets Table 2 shows basic descriptive statistics for our data. The average log price of a good in our sample is approximately 5 (or approximately 150 dollars). This magnitude is significantly larger than the level of prices considered in previous studies (e.g., with scanner price data or online prices of books and CDs) where goods are routinely have prices below 10 dollars. It is also not unusual in our sample to observe prices of goods above $600 (approximately 75 th percentile) or $1400 dollars (approximately 90 th percentile). Since we will focus on how quickly cross-border arbitrage opportunities dissipate, the level of prices is important since search effort is likely to be larger for big-price-tag items. Also note that the level of prices is approximately the same in the U.S. and Canada. Goods routinely have multiple sellers in our data. The average number of sellers is approximately 2.8 in Canada and 3.4 in the U.S. This is consistent with the notion that U.S. market is larger than the Canadian market but the difference is not as striking as one can observe in the numbers of regular brick-and-mortar stores in two countries. In part this difference in the number of stores is smaller because online markets tend to be more concentrated. In line with this observation, the share of large online stores in Canada (0.112) is more than double of share of large online stores in the U.S. 6 We also observe that the stability of sellers we define stability as the ratio of the number of stores selling a good in a given week to the number of stores ever selling this good in the month which covers the given week is larger in Canada (0.48) than in the U.S. (0.43). One can interpret this statistic as suggesting that the turnover of sellers is lower in Canada than in the U.S., which is consistent with more market power of Canadian online stores. Similar to previous studies of online prices (e.g., Brynjolfsson and Smith 2000, Baye et al. 2006), we observe dramatic dispersion of prices for the same good in any given point in time (Figure 2). On average across goods and time periods, the standard deviation of log prices within a country is approximately , which is significant but smaller than one can observe for the dispersion of prices across regular stores. 7 Thus it is common to observe price differentials of 30 percent or more across sellers in a country. Given that the levels of prices are large in our sample, these price differentials correspond to significant dollar amounts. In some cases, the differences between cheapest and most expensive prices are in multiple 6 We define a large store as a store that sells 75 or more percent of goods in the universe of goods sold online. 7 For example, Coibion et al. (2011) report that the standard deviation in the log price for a given unique product code (UPC), a given market (metro area) and a given week is 28% on average across periods, markets and UPCs. 8

10 hundreds of dollars. These price differentials are surprising given that search for the best prices is very easy in online markets. However, we do observe that the size of price differentials is negatively correlated with the level of prices. That is, more expensive goods tend to have smaller price dispersion. We also find that the cross-sectional dispersion of prices in any given market is fairly persistent. The series correlation of log or level of standard deviation of log prices in a given week is routinely above The duration of price spells depends on the treatment of missing values in price quotes. The mean price spell is the shortest with procedure A (approximately weeks) and the longest with procedure C (approximately weeks). The duration of price spells is a little longer in the U.S. than in Canada (see Figure 3). Overall, the duration of price spells is considerably shorter for online prices than for prices observed in the scanner price data (e.g., Kehoe and Midrigan 2008), in government surveys of prices (e.g., Bils and Klenow 2004, Nakamura and Steinsson 2008), or in online prices for books (e.g., Boivin et al. 2012) where mean duration of price spells is at least several months. The duration of price spells we document, however, is consistent with Lunnemann and Wintr (2011) who analyze a similar set of goods but have data only for one year. Price increases and decreases are equally likely in our data (Figure 4). The average price change however is slightly negative which captures the fact that goods in our sample are subject to rapid technical improvements over time and thus prices of existing goods tend to depreciate with the age of goods. The median absolute log price change is approximately 4% in both countries, which is again similar to facts documents in Lunnemann and Wintr (2011). This magnitude is approximately a half of the counterpart previous studies document for brick-and-mortar stores (e.g., Nakamura and Steinsson 2008). By and large, shorter durations of price spells and smaller sizes of price changes for online prices than for prices in regular stores are consistent with menu costs being smaller for online sellers than for regular stores. However, the non-negligible rigidity of online prices suggests that there are other costs associated with prices changes (e.g., collecting and processing information) which makes prices somewhat inflexible. As a final measure of price stickiness at the aggregate level, we compute synchronization of price changes across sellers. Specifically, we compute synchronization in a given week and for a given good as the fraction of price quotes with a price change conditional on a least one price change at this point in time. The average fraction is about 30 percent. Given that there are on average three stores for a given good in a given week, the magnitude of this fraction points to little synchronization of price changes across sellers. However, there is heterogeneity in synchronization across goods. We present selected statistics by category of goods in Appendix A. 9

11 4. International price differentials A. Stylized facts Figure 5 presents descriptive statistics for log and log /EX, the relative and real exchange rate respectively. We use two measures of prices to compute international price differentials: P and P are median or mean prices of good i at time t in Canada and U.S. is the nominal exchange rate. There is considerable dispersion of price differentials. [to be completed] B. Basic specifications The focus of our empirical analyses is to estimate the depth of pass-through and the speed of adjustment to a new level of relative prices in response to movements in the nominal exchange rate between U.S. and Canada. We also want to study the determinants of the depth and speed of price adjustment. Our empirical approach builds on previous studies. We employ two basic econometric specifications. The first specification measures the responsiveness of relative prices to changes in the nominal exchange rate: log (1) where i and t index goods and weeks, is the median price (in Canadian dollars) of good i in week t in Canada, is the median price (in U.S. dollars) of good i in week t in the U.S., is the exchange rate (CAD/USD) in week t, and is a set of control variables. The law of one price predicts that should be equal to one. Larger values of correspond to smaller departures from the law of one prices. Note that this specification estimates the longrun pass-through (Goldber and Knetter 1997, Campa and Goldberg 2005). Our second specification is aimed to measure the speed of price adjustment: log log, log,,, (2) where is the first difference operator. 8 Specification (2) is set in the errorcorrection/cointegration form where quantifies how quickly the deviation from equilibrium is eliminated. More negative values of mean faster adjustment. In Specification (2), equilibrium relationship between relative and the exchange rate are determined according to Specification (1). Thus, while the equilibrium relationship nests the law of one price, it also 8 We use BIC to select the number of lags for log, allow for more lags., and. Results are similar when we 10

12 allows deviations from the law of one price. To simplify estimation, we use the fact that an estimate of in Specification (1) is super-consistent. Hence, we can first estimate Specification (1) and then use to construct the deviation from equilibrium relationship in Specification (2). Table 4 reports estimated specifications (1) and (2) on pooled data. To account for the fact that error terms in Specification (1) and (2) can be correlated across time, goods, and countries as well as the fact that is common across goods and countries, we use the Driscoll and Kraay (1998) standard errors. Note that for Specification (2) we have fewer observations because we restrict the sample only to price quotes with at least one spell of twenty consecutive observations without imputations. The estimated exchange rate passthrough is about 35% which is similar to estimates reported in other studies based on prices collected from regular stores (see Barhoumi 2005, Campa and Goldberg 2005, Gaulier, Lahrèche-Révil, and Méjean 2006, Kardasz and Stollery 2001, Menon 1996, Takhtamanova 2008). However, this magnitude appears small given that in online markets i) prices are remarkably flexible, ii) competition is fierce, iii) consumers can easily buys goods from U.S. or Canada, iv) distribution/non-tradable costs are small, and v) most goods are produced overseas so that the costs are similar across countries. In other words, the prior for many economists and observers could have been that pass-through should have been larger. The estimate of suggests correction of prices toward a long-run equilibrium. If we abstract from the short-run dynamics in Specification (2), the estimate of implies that 3.8% of the gap from the long-run relationship is closed in a week (or correspondingly about 15% of the gap is closed in a month and 40% in a quarter) which implies the half-life of approximately 4.5 months. This speed of adjustment is considerably larger than the speed estimated on price indexes (e.g., Rogoff (1996) estimates the half-life of 3 to 5 years) or scanner price data where prices of exact same goods are compared across countries (e.g., Broda and Weinstein (2008) estimate the half-life of 2.9 quarters). This speed of price adjustment, however, would probably not surprise observers of the online markets. For example, Baye, Gatti, Kattuman, and Morgan (2007) emphasize that i) online customers compare prices within goods, not within stores; ii) the number of sellers and prices changes frequently; and iii) firms need to constantly monitor prices of their rivals. All of these factors are likely to accelerate price adjustment. Yet, the speed of adjustment we find is much higher than the speed estimated by Boivin et al. (2012) for online prices of books. This discrepancy in the results is likely to reflect the specifics of book prices which tend to have much stickier prices and more concentrated market power of sellers. C. Determinants of pass through and speed of price adjustment A key question is what factors determine the size of the pass-through and the speed of price adjustment. Usually, it is hard to answer this question because the data are available only at the aggregate level or little is known about the properties of goods. Fortunately, our data set 11

13 contains useful information about a number of potentially important determinants which we will use in our analyses. First, there is evidence (e.g., Deck and Wilson 2006) that the return to search effort is higher for expensive goods. For example, consumers are more likely to search for better deals on computers and plasma TVs than on toothpaste or beer. A higher search intensity should put a larger pressure on price convergence across sellers and countries. Thus, one may expect that more expensive goods should exhibit a larger pass-through and faster speed of price adjustment. Our data set has a wide distribution of goods in terms of their prices and we can exploit this variation to examine and quantify this channel. Specifically, we can use the log median price over the sample to proxy for returns on search. Second, the number of sellers should be indicative of the degree of competition. With more sellers, one should expect a larger pass-through and the speed of adjustment. Since online markets tend to be more concentrated (Baye et al. 2007), the role of large sellers could be particularly important for determination of prices. On the one hand, if the market for a good is dominated by large sellers, it is more likely that these sellers will exercise their market power and thus make prices more rigid which in turn may mean lower pass-through and slower speed of price adjustment. On the other hand, large sellers may have an agreement with manufactures which would guarantee the lowest prices for large sellers ( find a cheaper price and we ll match it effect). Thus, a lower price offered by a smaller seller can automatically trigger price adjustment for large sellers and thus accelerate and deepen price adjustment. We will use the log of the number of sellers as well as the number of large sellers to proxy for these forces. Third, an easier entry into selling a good is likely to make competition stronger (e.g., hit-andrun strategy) and, as a result, make pass-through larger and price adjustment faster. A stronger turnover of sellers is likely to be indicative of how easy it is to start selling a given good. We proxy for the turnover using our stability measure (more stable set of sellers means lower turnover) and hence we should expect a negative correlation between stability and pass-through and between stability and speed of price adjustment. Fourth, a number of authors (Imbs et al 2005, Mayoral and Gardea forthcoming, Rogoff 1996, Takhtamanova 2008) suggested that price stickiness could be an important force in determining how deviations from the law of one price are eliminated. With flexible prices, adjustment could be deep and quick. In contrast, sticky prices could delay price adjustment and make it incomplete. We can measure the degree of price stickiness using the median duration of price spells for a given good in our sample. Longer durations should be associated with smaller pass-through and slower price adjustment. To test these predictions, we estimate specifications (1) and (2) for each good separately and then regress estimated and on the factors we describe above. In particular we estimate the following two specifications: log log log (3) 12

14 log log log (4) where i indexes goods, is the median price of good i in Canada, is the median duration of price spells in Canada (procedure C), 9 is the number of sellers of good i in the U.S. and Canada, is the number of large sellers of good i in the U.S. and Canada, is the stability of sellers in the U.S. and Canada, is a set of fixed effects for periods over which and are estimated, is a set of fixed effects for categories of goods. Table 5 reports estimated coefficients for Specifications (3) and (4) by least squares. Consider first results for the pass-through (Panel A). We find consistently across columns that higher prices of goods are associated with larger pass-throughs. Increasing the median price by 100 log points raises pass-through by 25 to 40 percent. This finding implies that for goods with prices of about 1,000 dollars (i.e., approximately two standard deviations above the median price in our sample), the pass-through approaches 100 percent. Thus, the law of one price holds approximately true for expensive items. Once we control for category fixed effects (columns 3 and 4), there is a negative relationship between the size of the pass-through and price stickiness. Specifically, a 100 log points increase in the duration of price spells (e.g., increasing the duration of spells from 4 weeks to 12 weeks) decreases pass-through by percent. Note that although price stickiness is likely to be a short-run phenomenon, it is also likely to capture some long-run properties of price adjustment. In particular, apart from larger menu costs generating longer price spells and making non-adjustment of prices an equilibrium outcome, longer durations of price spells are likely to be positively correlated with the market power of sellers. Indeed, one of the basic premises of models where sellers have sticky prices is that these sellers have market power to set prices. Furthermore, a steeper marginal revenue, which can stem from a low elasticity of demand (high market power), is likely to make prices less sensitive to shocks and thus make prices stickier. One may reason that the negative correlation on the duration of price spells is likely to capture the market power of sellers. Consistent with this logic, we also find that the number of sellers is positively associated with the size of the pass-through. A 100 log points increase in the number of sellers raises the pass-through by at least 30 percent. Therefore, increasing the number of sellers is likely to make the law of one price to hold. Presence of larger sellers tends to decrease the size of the pass-through. Likewise, stability of sellers (i.e., low turnover of sellers) tends to weakly decrease the size of the pass-through. These results can also explain why the pass-through estimated on data from a single seller (especially for branded goods of this seller) is likely to violate the law of one price. These results are also consistent with findings for pass-through in regular markets (e.g., Gaulier et al. 2006). 9 We find similar results when we use procedures A and B to calculate the duration of price spells. Likewise, we find similar results when we use duration of price spells for the U.S. 13

15 The speed of price adjustment can also be related to these factors (Panel B). An increase in the level of prices weakly decreases the speed of price adjustment. This result is not consistent with more intensive search which is likely to be stimulated by higher levels of prices. However, the relationship between the level of prices and the speed of adjustment is not statistically significant when category and time fixed effects are included. The duration of price spells is strongly associated with the speed of price adjustment. Increasing the duration of price spells by log points is quantitatively big enough to essentially eliminate error correction to deviations from equilibrium. Again, several forces could be at work here. First, stickier prices mechanically affect the speed at which prices can change. Second, stickier prices probably also capture the degree of market power which can slow down adjustment of prices. The estimates of suggest that after controlling for fixed effects (columns 2, 3 and 4) a larger number of sellers accelerates converges to long-run equilibrium relationship between prices and exchange rate. This effect is consistent with the view that more competition brings prices faster to equilibrium. Interestingly, an increase in the number of large sellers also speeds up adjustment of prices. Thus in contrast to results for pass-through, effects of more sellers and of more large sellers are pointing in the same direction. Although we do not know the exact mechanism behind this congruence, it seems plausible that large sellers accelerate adjustment because they have match lower prices policies or because they can coordinate price changes. In any case, this difference between pass-through and speed of price adjustment is consistent with experimental evidence in Deck and Wilson (2003) showing that match lower prices clauses are likely to make adjustment faster but also make collusive behaviour easier and thus raises prices (in our case correspondingly lower pass-through). 10 The stability of sellers has a significant effect on the speed of price adjustment. Specifically, a lower turnover of sellers (higher stability) reduces the speed (i.e., becomes larger and closer to zero). This finding is consistent with the view that easy entry into a market and limited time-horizons for sellers (this limits the scope for collusion) are likely to eliminate arbitrage opportunities and mis-pricing of goods faster. The quantitative effect of seller stability is large. A one standard deviation increase in stability (approximately 0.2) would reduce the speed () by at least 0.04 which is again enough to severely limit convergence to equilibrium. In short, results in Table 5 can be summarized as follows. First, higher prices of goods, lower price stickiness, larger number of sellers and smaller number of large sellers increase the size of the pass-through. Second, lower price stickiness, larger number of sellers and large sellers, and higher turnover of sellers are likely to accelerate the speed of price adjustment. 10 Deck and Wilson (2003) examine the effect of various automated pricing strategies. In particular, they consider undercutting, low-price matching, and trigger pricing (Their example of the trigger pricing strategy is as follows. A firm sets price p 0. When the smallest price available drops below threshold p high, a firm triggers n rounds of price p low (p low <p high ), after which it returns to p high.). In an experimental study, they find that undercutting leads to competitive prices, low-price matching produces collusion outcomes with prices higher than competitive ones, while trigger pricing leads to lower price than in the competitive case. Overall, automated prices can increase joint profit maximization by firms. 14

16 When estimating coefficients in Table 5, we assign equal weight to all observations. However, the precision of estimates for and will vary across goods since the number of observations used to estimate and varies across goods. Given considerable heterogeneity of and estimates, one may want to weigh observations based on the precision of and. Although using standard errors of and may be natural choice for weights, in finite samples and and their standard errors may be correlated (see e.g. Abowd and Card (1989) and Altonji and Segal (1996)) and thus these weights may be inappropriate. A simple alternative is to use the number of observations used in estimation. Results for estimation of Specifications (3) and (4) with these weights are reported in Table 6. By and large, results are similar to those reported in Table 5. As a final check on our results, we also explore the effect of controlling for the degree of synchronization in price changes. This control could be important because pass-through and speed of price adjustment could be affected not only by the degree of price stickiness at the level of individual seller but also to what extent price setting is staggered. In standard macroeconomic models, higher synchronization leads to stronger and faster reactions of prices. Consistent with this theoretical prediction, Table 7 shows that increased synchronization is associated with larger pass-throughs and faster price adjustment. 5. Natural experiment In addition to the evidence on the size of exchange rate pass-through and the speed of adjustment, our dataset can provide additional insights on how online prices react to changes in economic fundamentals. In particular, we study the effect of an exogenous cost shock that originates from abroad on online prices in Canada and the U.S. This natural experiment is the flood in Thailand that adversely affected production of hard-drives and, hence, their prices, as well as prices for laptops, for which a hard-drive is an essential and a rather costly component. The advantages of using such a natural experiment are as follows. First, it is clearly an exogenous shock due to natural forces. Exogenous variation helps us deal with reverse causality problem, as well as exemplify the source of variation in price changes. Second, the shock originates far away from the countries of interest (Canada and the U.S.) and clearly does not affect economic conditions in these markets in any other way. Besides, without market frictions it is clear that the cost pass-through into the U.S. and Canadian prices should be the same. Hence, any differences are likely to be associated with such frictions. Third, the scale of the disruption is sufficiently big to have a tangible effect on prices. Fourth, our dataset spans both periods before and after the shock and covers a large variety of models and producers within laptop and hard-drive categories. We proceed with a summary of the events in Thailand and their effect on markets of hard-drives and laptops followed by a description of our empirical strategy. 15

17 The floods in Thailand started in late July By mid-october, they reached the capital, Bangkok, and did not recede until as late as January As of December 2011, the World Bank had estimated US$ 45 billion in damages for the Thai economy, mostly due to disruptions in manufacturing (US$ 32 billion). More than 90% of all losses are borne by private owners. 11 As Thailand hosts major hard-drive producers, the floods took their toll on hard-drives production and prices. For example, Western Digital (WD), one of the leading manufacturers, has over 60% of their capacities in the affected region. Toshiba employs almost four thousand workers in Thailand, with 50% of its production capacities being in the country. Both firms had to shut down their factories during the floods. Figure 8 shows an extent of damages to a WD factory producing hard drives. Nidec, which produces 75% of hard drive motors an essential part of hard drives had to shut down too. 12 As the disruption to hard-drive production led to substantial shortages of hard-drives, we treat the events in Thailand as exogenous adverse supply shock, which, according to economic theory and experts opinion in technology sector, should lead to higher prices. 13 Major laptop producers such as Dell, Acer, HP, and Toshiba are all customers of the affected hard-drive manufacturers to some extent, the floods were bound to affect laptop costs too. Our dataset allows us to track online price quotes for a specific hard-drive/laptop model in the U.S. and Canada. Since the supply side is dominated by the rising costs due to the floods in Thailand, by taking the price ratio of the Canadian to U.S. prices, we can differentiate it out, to see whether incomplete pass-through is due to cost differences or unrelated reasons. [To be completed] 6. Discussion and concluding remarks [to be completed] 11 Source: N/0,,contentMDK: ~pagePK:141137~piPK:141127~theSitePK:333296,00.html 12 Source: ch_firm.html 13 Source: 16

18 References Abowd, J. M. and Card, D. 1989, On the Covariance Structure of Earnings and Hours Changes, Econometrica 57(2), Altonji, J. and Segal, L. 1996, Small-Sample Bias in GMM estimation of Covariance Structures, Journal of Business and Economic Statistics 14(3), Barhoumi, Karim Long Run Exchange Rate Pass-Through Into Import Prices In Developing Countries: An Homogeneous or Heterogeneous Phenomenon?" Economics Bulletin 6 (14):1-12. Baye, Michael R. and John Morgan Information Gatekeepers on the Internet and the Competitiveness of Homogeneous Product Markets." American Economic Review 91 (3): Baye, Michael R. and John Morgan Price Dispersion in the Lab and on the Internet: Theory and Evidence." RAND Journal of Economics 35 (3): Baye, Michael R. and John Morgan Brand and Price Advertising in Online Markets." Management Science 55 (7): Baye, Michael R., J. Rupert J. Gatti, Paul Kattuman, and John Morgan Did the Euro Foster Online Price Competition? Evidence from an International Price Comparison Site." Economic Inquiry 44 (2): Baye, Michael R., J. Rupert J. Gatti, Paul Kattuman, and John Morgan A Dashboard for Online Pricing." California Management Review 50 (1): Baye, Michael R., J. Rupert J. Gatti, Paul Kattuman, and John Morgan Clicks, Discontinuities, and Firm Demand Online." Journal of Economics & Management Strategy 18 (4): Bils, Mark, and Peter Klenow, Some Evidence on the Importance of Sticky Prices. Journal of Political Economy 112(5): Boivin, Jean, Robert Clark, Nicolas Vincent, Virtual borders, Journal of International Economics 86(2): Broda, Christian and David E. Weinstein Understanding International Price Differences Using Barcode Data." NBER Working Papers 14017, National Bureau of Economic Research, Inc. Brynjolfsson, Erik and Michael D. Smith Frictionless Commerce? A Comparison of Internet and Conventional Retailers." Management Science 46 (4): Campa, Jose Manuel and Linda S. Goldberg Exchange Rate Pass-Through into Import Prices." Review of Economics and Statistics 87 (4): Cavallo, Alberto Scraped Data and Sticky Prices: Frequency, Hazards, and Synchronization." Working paper, Harvard University. Coibion, Olivier, Yuriy Gorodnichenko, and Gee Hee Hong, The Cyclicality of Effective Prices: Sales, Store-Switching and the Mismeasurement of Inflation, U.C. Berkeley, manuscript. Crucini, Mario J. and Mototsugu Shintani Persistence in Law of One Price Deviations: Evidence from Micro-Data." Journal of Monetary Economics 55 (3): Deck, Cary A. and Bart J. Wilson Automated Pricing Rules in Electronic Posted Offer Markets." Economic Inquiry 41 (2): Deck, Cary A. and Bart J. Wilson Tracking Customer Search to Price Discriminate." Economic Inquiry 44 (2): Dellarocas, C The Digitization of World of Mouth: Promise and Challenges of Online Feedback Mechanisms, Management Science 49(10):

19 Driscoll, J.C., A.C. Kraay, Consistent Covariance Matrix Estimation With Spatially Dependent Panel Data, Review of Economics and Statistics 80(4), Engel, Charles Real Exchange Rates and Relative Prices: An Empirical Investigation." Journal of Monetary Economics 32 (1): Gaulier, Guillaume, Amina Lahreche-Revil, and Isabelle Mejean Structural Determinants of the Exchange-Rate Pass-Through." Working Papers , CEPII research center. Goldberg, Pinelopi K. and Rebecca Hellerstein A Structural Approach to Identifying the Sources of Local-Currency Price Stability." NBER Working Papers 13183, National Bureau of Economic Research, Inc. Goldberg, Pinelopi K., and Michael M. Knetter Goods Prices and Exchange Rates: What Have We Learned?" Journal of Economic Literature 35 (3): Gorodnichenko, Yuriy and Linda L. Tesar Border Effect or Country Effect? Seattle May Not Be So Far from Vancouver After All." American Economic Journal: Macroeconomics 1 (1): Hong, Gee Hee, and Nicholas Li, Vertical Integration and Retail Pricing Facts for Macroeconomists: Private Label vs. National Brand, U.C. Berkeley, manuscript. Imbs, Jean, Haroon Mumtaz, Morten Ravn, and Helene Rey One TV, One Price?" Scandinavian Journal of Economics 112 (4): Imbs, Jean, Haroon Mumtaz, Morten Ravn, and Helene Rey PPP Strikes Back: Aggregation and the Real Exchange Rate." Quarterly Journal of Economics 120(1):1-43. Kardasz, Stanley W. and Kenneth R. Stollery Exchange Rate Pass-Through and Its Determinants in Canadian Manufacturing Industries." Canadian Journal of Economics 34 (3): Lunnemann, Patrick and Ladislav Wintr Price Stickiness in the US and Europe Revisited: Evidence from Internet Prices." Oxford Bulletin of Economics and Statistics 73 (5): Mayoral, Laura and Maria Dolores Gadea Aggregate Real Exchange Rate Persistence through the Lens of Sectoral Data." Journal of Monetary Economics 58 (3): Menon, Jayant The Degree and Determinants of Exchange Rate Pass-Through: Market Structure, Non-Tariff Barriers and Multinational Corporations." Economic Journal 106 (435): Morgan, John, Henrik Orzen, and Martin Sefton An Experimental Study of Price Dispersion." Games and Economic Behavior 54 (1): Nakamura, Emi, and Jon Steinsson Five Facts About Prices: A Reevaluation of Menu Cost Models. Quarterly Journal of Economics 123(4): Nakamura, Emi, and Jon Steinsson Price Setting in Forward-Looking Customer Markets. Obstfeld, Maurice, and Kenneth Rogoff, The Six Major Puzzles in International Macroeconomics. Is There a Common Cause, in NBER Macroeconomics Annual 2000, B.S. Bernanke and K. Rogoff, eds., Cambridge, MA: MIT Press. Rogoff, Kenneth The Purchasing Power Parity Puzzle." Journal of Economic Literature 34 (2): Syverson, Chad Prices, Spatial Competition and Heterogeneous Producers: An Empirical Test." Journal of Industrial Economics 55 (2): Takhtamanova, Yelena F Understanding Changes in Exchange Rate Pass-Through." Journal of Macroeconomics 32 (4):

20 Figure 1. Time series of CAD/USD exchange rate. Canadian Dollars to One U.S. Dollar, Daily Start of data collection 01jan jan jan jan jan2012 Notes: Source: Board of Governors. 19