Quantifying Bid Shading in Online Auctions via A Functional Approach Galit Shmueli, Wolfgang Jank University of Maryland Ravi Bapna University of Connecticut 1. Bid shading and the winner’s curse Winner’s curse is a phenomenon that occurs in auctions for common-value items. These are auctions where the item being auctioned is valued the same by each participant, although the participants may not know the precise value1 . Since each bid is assumed to be a realization from the value distribution, the average bid is assumed to be the estimate of the common value. For this reason, if bidders bid their estimate then the highest bid is an over-estimate of the common value, leading to the “Winner’s Curse” phenomenon. To avoid this curse, bidders in such auctions must “shade” their bid, i.e. shrink their estimate. According to auction theory, the amount of shrinkage depends on the number of bidders in the auction: the more bidders, the more shading is required. Using this reasoning, if we see bid shading taking place, this can be indicative of a common-value component. If there is no bid shading, this can indicate either a private-value item or else a common-value auction that experiences the winner’s curse. But more importantly, from the degree of bid shading we can learn about how bidders are behaving regardless of our knowledge of whether the item is common or private value. 2. The winner’s curse test and its limitations The winner’s curse phenomenon has been investigated in the context of real online auction data by several authors. They look for the presence of bid shading which would be indicative of a common-value auction. This is done by determining whether the change in bid magnitude is affected by the number of bidders. Had the series of bids during an auction consisted of independent events, then we could have treated the pairs {bids,# Bidders} as a random sample and examine the change in bid magnitude as a function of the number of bidders. A linear relationship would indicate that the bid increase is equally affected by an additional bidder irrespective of the number of bidders (i.e., whether the increase is from 2 to 3 bidders or from 10 to 11 bidders, the bid increase is similar). This is what we would expect to see in a common value auction with no bid shading. Alternatively, in the presence of bid shading we would expect to see a nonlinear relationship, where the rate of bid increase per additional bidder decreases as a function of the number of bidders. This is essentially the setting that Bajari & Hortacsu (2003)2 use, relying on Milgrom & Weber’s (1982)3 “winner’s curse test” which compares the average bid in auctions with different numbers of bidders. Like the above setting, they treat the set of all bids as an iid sample. They consider the bid magnitude as their response variable and then make the restricting assumption that the relationship of bid magnitude and number of bidders is linear, whether or not there is bid shading. They then fit a linear regression of bids to the number of bidders, and find a negative correlation between the two. They take this as suggestive evidence for the presence of bid shading, and thus a common-value auction. However, a negative correlation between bid magnitudes and the number of bidders can be explained by other factors: First, there are usually less auctions with many bidders, which means fewer data points with many bidders. The regression estimates are therefore susceptible to influential points at the high number of bidders range. Even a single auction with many bidders and a relatively low last bid can lead to a negative coefficient. Second, if the underlying relationship between the rate of bid change and the number of bidders is linear (y 0 = β0 + β1 x) when there is no bid shading, then the relationship between the bid magnitude and number of bidders is expected to be quadratic: y ∝ β0 x + β1 x2 . In fact, a scatterplot of bids vs. number of bidders for a sample of 97 Cartier watch auctions does show some curvature (Figure 1, top left). Finding a negative slope coefficient in a linear regression of bids vs. number of bidders is therefore not necessarily 1 from “Common value auction” in Wikipedia. P and Hortacsu, A (2003) “The Winner’s Curse, Reserve Prices and Endogenous Entry: Empirical Insights from eBay Auctions”, The Rand Journal of Economics 3(2), 329-355. 3 Milgrom, P R and Weber, R J (1982), “A Theory of Auctions and Competitive Bidding”, Econometrica, 50, 10891122. 2 Bajari, 1 3500 3000 3000 2500 2500 Proxy bid ($) Proxy bid ($) 3500 2000 1500 2000 1500 1000 1000 500 500 0 0 5 10 Number of bidders 15 0 20 2000 0 5 10 Number of bidders 15 20 2000 1500 1500 Residual velocity Bid velocity 1000 500 0 1000 500 0 −500 −500 −1000 −1500 0 5 10 Number of bidders 15 −1000 20 2 4 6 8 10 Number of bidders 12 14 16 Figure 1: Top: Bids vs number of bidders for 97 Cartier wristwatch auctions (left) and auction curves (right). Bottom: bid velocity per number of bidders before (left) and after (right) factoring out bid timing. indicative of no bid shading. Instead, we should look at the rate of bid change, or bid velocity.And third, the assumption that the price at time t is not related to the price at time t0 , (t0 < t) is not reasonable. In fact, the underlying assumption in common-value item auctions is that bidders derive information from the other signals, and in our case this information is manifested by the observable previous bids. 3. Quantifying the curse: a functional approach We introduce a new method that directly measures the degree of bid shading and overcomes the shortcomings above. The method is based on a functional data analytic approach, which is capable of capturing the dynamics that take place during an auction. The functional approach uses the entire set of bids, not just those in the last minutes of the auction; it maintains the relationship between the bid amount and bid timing; and it captures the temporal dependencies between bids in the same auction. Using a functional approach, we treat the set of {bid,# bidders} pairs from an auction as a time series and use a curve to represent it (Figure 1, top right). We then treat the set of auction curves as our sample. The first derivative of these curves tells us about the price increase per additional bidder (Figure 1, bottom left). This is almost a reasonable measure of bid shading, except that it is biased by the bid timing: in closed-ended online auctions there tends to be a surge of bidding towards the auction end, and sometimes some early bidding as well. This means that the bid-increase-per-bidder is likely to vary during different times of the auction. Bajari & Hortacsu (2003) addressed this by using data only from the last few minutes of the auction. However, with our method we can use the entire bid data. Our solution is to first factor out the bid time and then fit curves and estimate derivatives. The factoring out of bid timing is done by fitting a (linear or nonlinear) regression between the bid amounts and bid times. The residuals from this regression are then free of bid timing effects, and are used to obtain curves and curve derivatives (Figure 1, bottom right). Our Cartier watch auction sample shows that before removing the bid timing the average bid velocity is slightly positive and constant in the number of bidders, indicative of no shading (Figure 1, bottom left). However, the average velocity changes, once bid timing is factored out, to a constant zero bid velocity (Figure 1, bottom right) again indicative of no shading (and therefore presence of winner’s curse if these are common value items). Thus, the initial positive velocity was reflecting the heat of the auction rather than the effect of the number of bidders, which in other instances might conceal the presence of bid shading. 2