NBER WORKING PAPER SERIES BEST PRICES Judith A. Chevalier Anil K Kashyap Working Paper 16680 http://www.nber.org/papers/w16680 NATIONAL BUREAU OF ECONOMIC RESEARCH 1050 Massachusetts Avenue Cambridge, MA 02138 January 2011 The views expressed here are our own and not necessarily those of any institutions with which we are affiliated, nor of the National Bureau of Economic Research. Kashyap thanks the Chicago Booth Initiative on Global Markets for research support. We thank Cecilia Gamba, Aaron Jones, and Ashish Shenoy for outstanding research assistance. We thank numerous seminar participants for helpful comments. We are grateful to the Kilts Marketing Center and SymphonyIRI Group for the data. As a condition of use, SymphonyIRI reviews all papers using their data to check that the data are not described in a misleading fashion. However, all analyses in this paper based on SymphonyIRI Group, Inc. data are the work of Chevalier and Kashyap, not SymphonyIRI Group, Inc. NBER working papers are circulated for discussion and comment purposes. They have not been peer- reviewed or been subject to the review by the NBER Board of Directors that accompanies official NBER publications. © 2011 by Judith A. Chevalier and Anil K Kashyap. All rights reserved. Short sections of text, not to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including © notice, is given to the source. Best Prices Judith A. Chevalier and Anil K Kashyap NBER Working Paper No. 16680 January 2011, Revised March 2011, Revised August 2012 JEL No. E3,E31,L11,L16 ABSTRACT We explore the role of strategic price-discrimination by retailers for price determination and inflation dynamics. We model two types of customers, "loyals" who buy only one brand and do not strategically time purchases, and "shoppers" who seek out low-priced products both across brands and across time. Shoppers always pay the lowest price available, the "best price". Retailers in this setting optimally choose long periods of constant regular prices punctuated by frequent temporary sales. Supermarket scanner data confirm the model's predictions: the average price paid is closely approximated by a weighted average of the fixed weight average list price and the "best price". In contrast to standard menu cost models, our model implies that sales are an essential part of the price plan and the number and frequency of sales may be an important mechanism for adjustment to shocks. We conclude that our "best price" construct provides a tractable input for constructing price series. Judith A. Chevalier Yale School of Management 135 Prospect Street New Haven, CT 06520 and NBER [email protected] Anil K Kashyap Booth School of Business The University of Chicago 5807 S. Woodlawn Avenue Chicago, IL 60637 and NBER [email protected] 1 Introduction Weexploretheoreticallyandempiricallytheimportanceofretailerpricedis- crimination strategies for descriptions of price dynamics. We depart from theusualapproachinthemacroliteraturethatignoresconsumerheterogene- ity in the price-setting problem and instead make this heterogeneity central to the analysis. In particular, we posit that that some consumers are ac- tive shoppers who chase discounts, substitute across products in a narrowly defined product category, and potentially use storage to maintain smooth consumption whilst concentrating their purchases in sale weeks. Other cus- tomers are passive, and retailers will employ strategies to charge these two groups different prices. Due to the actions of these strategic consumers, we find that weighted average prices paid differ substantially from posted prices. Themodelthatweproposetoaccountforthisbehaviorisamodelwhich does not necessarily speak to nominal price rigidities. But the model does suggest that conscious price discrimination rather than restrictions on the ability of firms to change prices (such as menu costs) is critical for under- standing price dynamics. In the context of our framework, the most natural way to understand nominal rigidities would be to look to the consumer side of the pricing problem rather than the firm side. But even without taking a stand on this issue, the evidence we develop does speak to a pair of debates that have emerged amongst macroeconomists looking at micro data. For example, there is a measurement question of whether (and poten- tially how) intermittent price discounts, or “sales”, should be incorporated in price index construction. Since episodic sales occur quite frequently in some sectors, price series that ignore sale prices display infrequent price changes, while price series that contain sales prices display very frequent price changes. Nakumura and Steinsson (2008) estimate a median price du- ration of approximately 5 months including sales, and approximately 8 to 11 months excluding them. Klenow and Kryvtsov (2008) find that consid- ering only regular (e.g., non-sale) prices raises the estimated median price duration in their dataset from 3.7 months to 7.2 months. In general, many researchers have documented that the price series for a given product from retailers such as grocery stores exhibit long periods of a constant regular price punctuated by occasional sales. These measurement issues come to the fore when macroeconomists must decide how to calibrate models that presume that prices are sticky. Our empirical analysis argues strongly for keeping track of sale prices because they play a major role in governing stores’ profits and our model suggests that their depth and/or frequency 2 could be altered to adjust to shocks. Likewise, a largely theoretical literature has emerged to look at the micro-foundations for why sales exist. For example, Matejka (2010) builds on the literature of rational inattention (see Sims (1998, 2003), Mankiw and Reis(2002,2006), for example) to explore its implications for price setting. In his model firms do not know the exact cost of the good each is selling and the firms find it expensive to keep track of all the information that would be necessary to perfectly deduce this cost. Thus, firms must decide which cost signals to monitor and because of their finite processing ability will set prices based on the estimated distribution of cost. In order to economize on information, the seller chooses signals that lead to a small number of distinct prices. Sales in this framework corresponding to situations in which a signal suggests that costs were low. There is a complementary long literature, beginning with Varian (1980) andSobel(1984)thatrationalizeprice-discriminationstrategiesthatinclude temporary discounts. Guimaraes and Sheedy (2011) were the first to bring this perspective to the macroeconomic literature on price-setting. Like us, they assume heterogeneous consumers. They have two types of consumers, bargain hunters and loyal shoppers, who are price insensitive; they explore how price discrimination plays out when retailers compete to sell to these customers. The key finding in their analysis is that competition bounds the importance of sales as a mechanism to adjust to shocks. Interestingly, the theoretical results in Guimaraes and Sheedy (2011) contrast the empirical results of Klenow and Willis (2007), who find that sale prices are just as responsive to macroeconomic conditions as regular prices. That is, Klenow and Willis find that the depth of sales decreases when recent inflation has been high and that sales prices give way to higher regular prices when recent inflation has been high. Our contribution is largely empirical, but to guide the analysis we begin by introducing a model in the spirit of Varian (1980), Sobel (1984), and Pe- sendorfer (2002) that allows for the possibility that pricing strategies might involve variation in the usage of discount prices. In contrast to the previous literature, we focus explicitly on a retailer controlling the prices of multiple products. The robust implication of this kind of model is that store pricing patterns ought to reflect the presence of different consumers and be strate- gically coordinated across products. The model can account for frequent temporary sales and long periods of constant regular prices. Furthermore, evenwithunchangingcostsanddemand, optimalmarkupsarenotheldcon- stant across items or across time. In contrast to Guimaraes and Sheedy, our model also implies that changing the frequency and depth of sales is the 3 optimal response to cost shocks and to certain kinds of demand shocks. To quantify the importance of consumer heterogeneity we turn to de- tailed microeconomic data. Some of our data are for particular supermar- ket products collected over parts of seven years at Dominick’s Finer Foods (DFF), a supermarket chain in the Chicago area. We also analyze data for a laterseven-yearperiod, usingadatasetprovidedbySymphonyIRI.TheIRI dataset covers stores in 47 markets around the country and we (randomly) selectedonestoreineachoftheninecensusregionsforevaluation. Pricesfor individual products at DFF and at the IRI stores display the now-familiar pattern of very infrequent regular price changes combined with frequent temporary sales. But upon more careful scrutiny, it appears that purchase patterns associated with the sales clearly reflect the important role played by bargain hunters. In particular, we show three sets of results. First, the model suggests that average prices paid for goods within a bundleofclosesubstitutesaretherelevantpriceforconsumersandretailers. Because retailers time sales strategically, the model suggests, and the data show, that sales for close substitute products tend not to occur in the same weeks. Consumers, we show, chase sales, and thus, actual prices paid are substantially lower than regular prices, and even measurably below average posted prices. Second, we introduce the concept of a “best price”, defined as the lowest price charged for any good in the narrow product category during a short multi-week time window. The model predicts that “best prices” should be the relevant prices for sales-chasing consumers. We show that the actual price paid tracks the “best price” and is well-approximated by an average of the best price and the fixed weight price index. The data match the structural form of our model. Third, the data exhibit strong spillovers in quantities purchased due to pricechangesforclosesubstitutes. So,forexample,whenthepriceofMinute Maid orange juice is reduced for a temporary discount, sales of Tropicana orange juice plunge (provided the Tropicana price does not change). In previous literature, large quantity variation along with constant prices has leadtoadiagnosisofsubstantialdemandvariability. However, weshowthat much of this apparent demand volatility derives from choices made by the retailer in setting prices for substitute products. For example, we show that total ounces sold of all products within a narrow product category are much less volatile than total ounces sold of any individual product suggesting that while customers actively shift between items when they go on sale, making individual demand quite volatile, the total demand across all the relevant items is much more stable. 4 We read the evidence as suggesting that the models of price setting that emphasize consumers taste, cognitive processing, or information-gathering deserve more attention. For instance, in our model, occasional sales provide a tool for retailers to effectively charge different consumers different prices andpricingdecisionsaredrivenbythedifferentconsumergroups’reservation prices. Thus, how consumers update reservation prices for individual goods becomes a critical factor affecting inflation. This mechanism has been much lessexploredthanhavethosewhichemphasizefirmscostsofchangingprices. Similarly, price index construction may need to be revised to account for the importance of bargain hunters. We provide some preliminary thoughts on how this might be done. Our paper proceeds as follows. Section 2 provides the simple model of a price-discriminating retailer and highlights empirical predictions. Section 3 describes the data. Section 4 establishes a number of new facts about pricing and purchase patterns that are consistent with the model. Section 5 discusses how to use our model to measure and summarize price series. Section 6 concludes. 2 A Model of Price Discriminating Retailers and Heterogeneous Consumers We begin by presenting a simple model that is similar in spirit to, and bor- rows significantly from, Varian (1980), Sobel (1984), or Pesendorfer (2002). The baseline version of our model takes consumer heterogeneity as its prim- itive. The firm knows about this heterogeneity and accounts for it in price setting. In this model, the firm bears no menu cost of changing prices. Nonetheless, we will show that the firm will iterate between a small number of prices, even in the face of some cost changes and some types of demand changes. The “regular” price will change infrequently but “sales” will be utilized. The model suggests that it is possible that the retailer will respond to a nominal shock by changing the frequency of sales while holding the “regular” price fixed. 2.1 Model Assumptions Consider a retailer selling two substitute differentiated products, A and B. We will focus on a single retailer for simplicity. However, we note that it would be fairly straightforward to embed our model into a model of two retailers competing in geographic space. In such a model (see, for example, 5 LalandMatutes(1994),Pesendorfer(2002)andHoskenandReiffen(2007)), consumer reservation prices would be determined by the price that would trigger consumer travel to another store. Thus, for tractability, we focus on a single retailer, but a monopoly assumption is not necessary. We discuss the competitive case more below. Assume that all customers have unit demand in each period but are differentiated in their preferences for the two substitute goods. A share α/2 of the customers value product A at VH and product B at VL, where VH > VL. We call these consumers the high A types. For convenience, we consider the symmetric case where a share α/2 of the customers, the high B types, valueproductBatVH andproductAatVL. Theremainingshareof consumers (1−α), the “bargain hunters”, value both products at VL. We normalize the total number of consumers to be 1 and consider N shopping periods.1 The seller has a constant returns to scale technology of producing A and B and the marginal cost of producing either is c. Each period, customers arrive at the retailer to shop. Consider the choices for the high A types (which will be symmetric for the high B types). If the price is less than or equal to their reservation value (PA < VH), then they buy their preferred good, A. If PA > VH and the price for product B is less than or equal to their reservation price for B (PB < VL), the high A type customers substitute to good B. If PA > VH and PB > VL, then the high A types make no purchase. For simplicity we also assume that, if the high types do not buy, their demand for the period is extinguished so that next period there are no implications of them having been out of the market.2 Next consider the choices made by the bargain hunters. If PA < VL and/or PB < VL, they will buy whichever product is cheaper.3 If PA > VL andPB > VL,thebargainhuntersdonotbuy. However,wecapturetheidea of shoppers being willing to engage in intertemporal storage by assuming that their demand partially accumulates to successive periods, deteriorating at rate 1-ρ. Thus, for example, if they made a purchase in period t − 1, 1In our model, VH and VL are real. However, it is possible that one or both types of consumer have nominal illusion. This would influence the retailer’s optimal response to changes in monetary policy. We leave exploration of this set of issues to future work. 2In our model, as long as costs are less than VH, the firm will always set price so that the high types purchase. We could make an assumption about high type demand accumulating, but it wouldn’t have any important implications for the model. 3Wewillseethatinequilibriumitwillnotbeprofitmaximizingtoputbothgoodson sale in the same week. 6 but PA = PB = VH in period t so that no purchases are made, then their total demand entering period t+1 will be (1−α)+(1−α)ρ. Similarly, if a good was available at a price of VL in period t−1, but PA = PB = VH in periods t, t+1, ..., t+(k−1), total demand from the bargain hunters in period k will equal: (1−ρk) (1−α)(1+ρ+ρ2+···+ρk) = (1−α) if 0 < ρ < 1. 1−ρ Notethen,thatweareassumingthatthehightypesareinactiveshoppers —they do not wait for and/or stock up during bargains, while the low types do. In this sense, our model reflects well the empirical facts described in Aguiar and Hurst (2007), in which they document that some consumers in a local area pay systematically lower prices for the same goods as other consumers. That is, some consumers are strategic in bargain-hunting, and others are not.4 Total profits for the retailer depend on the total amount of A and B sold. The retailer has three basic choices: (i) the retailer can charge high prices and service only the high types, foregoing any potential margins to be earned on the low types. (ii) the retailer can charge low prices and serve both types of customers, thus foregoing the extra willingness to pay that could have been extracted from the high types. Or (iii) the retailer can strategically iterate between high and low prices in an attempt to capture some of the potential demand from the bargain-hunters while exploiting some of the extra willingness to pay of the high types. 2.2 Model Results Depending on the parameters, strategy (i), (ii), or (iii), described above can be optimal. We will explain when each is optimal, with particular attention to parameter values under which (iii) is optimal, since the pricing behavior associated with (iii) is roughly consistent with our empirical observation of occasional sales at supermarkets. We characterize the retailer’s behavior in several steps. Proposition 1: As long as VL > c, the optimal price in any period is either VL or VH. 4AsinPesendorfer(2002),wecombine“bargain-hunting”behaviorwithlowwillingness to pay. We could provide a more detailed model with more types—brand loyals who are willing to intertemporally substitute purchases, brand loyals who do not intertemporally substitute,non-loyalswhoarewillingtointertemporallysubstituteandnon-loyalswhodo not intertemporally substitute. We think most of the interesting implications are evident with these 4 types collapsed into the two extremes. 7 Sketch of Proof: Choosing a price between VH and VL would reduce margins on the high types but would produce no offsetting demand increase for the bargain hunters. Thus, the optimal price is either VH or VL. Proposition 2: It is never optimal to charge VL for both good A and good B in the same period. Sketch of Proof: Charging VL for the second good leads to a loss of margins on the high types that prefer that good, but produces no offsetting demand increase for the bargain hunters. The bargain hunters’ demand is satisfied by charging VL for either good. This delivers one of the prediction that we will test empirically, namely that discount periods or “sales” for products within a product category are not synchronized. Thus, we have shown that the retailer will charge VH for at least one good in every period, and may charge VL for one good in some periods. Below, we show the conditions under which the retailer will charge VL for one good in every period, conditions under which the retailer will never charge VL for either good, and conditions under which the retailer will charge VL for one good in some periods but not all periods (intermittent sales). In order to derive these results, we first provide two straightforward intermediate results. Proposition3: Iftheretailerchoosestoholdoneandonlyonesaleduring the N periods, then the optimal time to hold it will be in the Nth (final) period. Sketch of Proof: If the retailer charges VH for both products during N −1 periods, and VL for one product in 1 period, profits will be: α α (N −1)α(VH −c)+ (VL−c)+ (VH −c)+ 2 2 (1−ρk) (1−α)(VL−c) (1) 1−ρ where k represents the period in which the sale is held. The first term of (1) is the profits from the high types in all of the non-sales periods, the second and third term represent the profits from the high types in the sale period (if the sale is on A, the high A types pay VL, but the high B types still pay VH). The fourth term of (1) represents the profits from the bargain hunters. Note that the first three terms are invariant to the timing of the single sale. The fourth term is maximized when k = N (as long as ρ > 0). Proposition 4; For a retailer who chooses to hold j sales (i.e. charge VL in j periods), the optimal strategy is to hold evenly spaced sales every k periods, where k = N/j. 8 Sketch of Proof: The prior proposition covered the case of j = 1 and k = N. Note that the logic underlying Proposition 3 for the entire period carries forward straightforwardly to sub-periods, so within each sub-period it makes sense to delay the sale as long as possible. Hence sales will be equally spaced. With the results of Propositions 3 and 4 in hand, we can characterize the remaining decision about when it pays to have any sales at all. We consider a retailer who holds a sale every kth period. If k = 1, the retailer holds a sale every period. If k = N, the retailer holds the minimum positive number of sales—one sale at the end to “sweep up” the low demanders. If k > N, the retailer never holds sales. We will focus on interior solutions where 1 < k < N. We consider the profits of a retailer who charges PA = PB = VH every period except during a “sale” and holds a sale at P = VL for one or the other good every kth period. Assuming no discounting, total profits for this retailer over all N periods are: k−1 N α N α N α(VH −c)+ (VL−c)+ (VH −c)+ k k 2 k 2 N (1−ρk) (1−α)(VL−c) (2) k 1−ρ The four terms in (2) are very intuitive. The first piece represents the profits from selling to the high types only, which will occur during all the non-sale periods. The second term is the profits from the high types during the periods where they are able to buy their preferred good on sale; during these sale periods the other high type still pays VH so that explains the third term. The last term is the profits from the bargain hunters. Note that, the larger ρ is, the less the bargain hunters’ demand has depreciated by the time the sale is held. Proposition 5: The retailer will find it optimal to hold some sales if: α 1−ρN (VH −VL) < (1−α) (VL−c) ⇔ 2 1−ρ 1−α1−ρN VH < VL+2 (VL−c) (3) α 1−ρ Sketch of Proof: The left hand side of first expression shows the loss from allowing the high types to pay less than they are willing to pay by offering a single sale. The right hand side of the first expression gives the profits from selling to the bargain hunters in an optimally timed single sale. 9