1: \documentclass[aps,pre,twocolumn,floatfix,nofootinbib,showpacs]{revtex4}

2: %\documentclass[pre,preprint,showcase,floatfix]{revtex4}

3: \usepackage[dvips]{graphicx}

4: \usepackage{epsf,amssymb,amsmath,mathptmx}

5: \begin{document}

6: \title{Bidding process in online auctions and winning strategy: 

7: rate equation approach}

8: \author{I.~Yang and B. Kahng}

9: \affiliation{School of Physics and Center for Theoretical Physics,

10: Seoul National University, Seoul 151-747, Korea}

11: \date{\today}

12: \begin{abstract}

13: Online auctions have expanded rapidly over the last decade and have 

14: become a fascinating new type of business or commercial transaction 

15: in this digital era. Here we introduce a master equation for the

16: bidding process that takes place in online auctions. 

17: We find that the number of distinct bidders who bid $k$ times, 

18: called the $k$-frequent bidder, up to the $t$-th bidding progresses 

19: as $n_k(t)\sim tk^{-2.4}$. 

20: The successfully transmitted bidding rate by the $k$-frequent

21: bidder is obtained as $q_k(t) \sim k^{-1.4}$, independent of

22: $t$ for large $t$. This theoretical prediction is in agreement with

23: empirical data. These results imply that bidding at the

24: last moment is a rational and effective strategy to win in an eBay

25: auction.

26: \end{abstract}

27: \pacs{89.75.-k, 89.75.Da, 89.65.-s, 89.65.Gh}

28: \maketitle

29: Electronic commerce (e-commerce) refers to any type of

30: business or commercial transaction that involves information

31: transfer across the Internet. As a formation of e-commerce, 

32: the online auction, i.e., the auction via the Internet~\cite{heck}, 

33: has expanded rapidly over the last decade and has become a 

34: fascinating new type of business or commercial transaction in 

35: this digital era. Online auction technology has several benefits 

36: compared with traditional auctions. Traditional auctions require 

37: the simultaneous participation of all bidders or agents at the 

38: same location; these limitations do not exist in online auction 

39: systems. Owing to this convenience, ``eBay.com," the largest 

40: online auction site, boasts over 40 million registered consumers 

41: and has experienced rapid revenue growth in recent years.

42: 

43: Interestingly, the activities arising in online auctions generated

44: by individual agents proceed in a self-organized 

45: manner~\cite{mantegna,bouchard, stanley,challet,pennock,hulst}. For

46: example, the total number of bids placed in a single item or

47: category and the bid frequency submitted by each agent follow power-law

48: distributions~\cite{yang}. These power-law behaviors~

49: \cite{simon,zhang,albert} are rooted in the fact that an agent who

50: makes frequent bids up to a certain time is more likely to bid

51: in the next time interval. This pattern is theoretically analogous 

52: to the process that is often referred to as preferential attachment, 

53: which is responsible for the emergence of scaling in complex

54: networks~\cite{ba}. This is reminiscent of the mechanism of 

55: generating the Zipf law~\cite{zhang,pareto}. The accumulated data of

56: a detailed bidding process enable us to quantitatively characterize 

57: the dynamic process. In this paper, we describe a master equation 

58: for the bidding process. 

59: The master-equation approach is useful to capture the dynamics 

60: of the online bidding process because it takes into account of the effect 

61: of openness and the non-equilibrium nature of the auction. 

62: This model is in contrast to the existing equilibrium approach~\cite{lb,kauf} 

63: in which there is a fixed number of bidders. The equilibrium approach 

64: is relevant to traditional auctions; however, it is unrealistic 

65: to apply this approach to Internet auctions. The power-law 

66: behavior of the bidding frequency submitted by individual agents 

67: can be reproduced from the master equation. 

68: Moreover, we consider the probability of an agent 

69: who has bidden $k$ times, called the $k$-frequent 

70: bidder, becoming the final winner. We conclude that the winner is 

71: likely to be the one who bids at the last moment but who placed 

72: infrequent bids in the past.

73: 

74: Our study is based on empirical data collected from two

75: different sources~\cite{yang}. The first dataset was downloaded

76: from the web, http://www.eBay.com, and is composed of all the auctions

77: that closed in a single day. The data include 264,073 auctioned items,

78: grouped into 194 subcategories. The dataset allows us

79: to identify 384,058 distinct agents via their unique user IDs. To

80: verify the validity of our findings in different markets and time

81: spans, the second dataset was accumulated over a period of one year from

82: eBay's Korean partner, auction.co.kr. The dataset comprised 215,852 agents

83: that bid on 287,018 articles in 355 lowest categories.

84: 

85: An auction is a public sale in which property or items of merchandise

86: are sold to the bidder who proposes the highest price. Typically,

87: most online auction companies adopt the approach of English auction, in

88: which an article or item is initially offered at a low price that is

89: progressively raised until a transaction is made. Both ``eBay.com"

90: and ``auction.co.kr" adopt this rule and many bidders submit multiple 

91: bids in the course of the auction. An agent is not allowed to place 

92: two or more bids in direct succession. It is important to notice 

93: that the eBay auction has a fixed end time: It typically ends a 

94: week after the auction begins, at the same time of day to 

95: the second. The winner is the latest agent to bid within this period. 

96: In such an auction that has a fixed deadline, bidding that takes 

97: place very close to the deadline does not give other bidders sufficient 

98: time to respond.

99: In this case, a sniper--the last moment bidder--might win the

100: auction, while the bid that follows has a substantial probability 

101: of not being transmitted successfully. While such a bidding pattern is

102: well known empirically, no quantitative analysis has been

103: performed on it as yet. In this study we analyze this issue through 

104: the rate equation approach.

105: 

106: To characterize the dynamic process, we first introduce several

107: quantities for each item or article as follows:

108: \begin{itemize}

109: \item[(i)] When a bid is successfully transmitted, time $t$ 

110: increases by one. 

111: \item[(ii)] Terminal time $T$ is the time at which an 

112: auction ends. Thus, the index of bids runs from $i=1$ to $T$.

113: \item[(iii)] $N(t)$ is the number of distinct bidders who 

114: successfully bid at least once up to time $t$. 

115: Thus, the index of bidders (or agent) runs from $i=1$ to $N(t)$. 

116: \item[(iv)] $k_i(t)$ is the number of successful

117: bids transmitted by an agent $i$ up to time $t$.

118: \item[(v)] $n_k(t)$ is the number of bidders with 

119: frequency $k$ up to time $t$.

120: \end{itemize}

121: From the above, we obtain the relations

122: \begin{equation}

123: N(t)=\sum_k n_k(t)

124: \end{equation} and

125: \begin{equation} t=\sum_{k} k n_{k}(t)

126: \end{equation}

127: for any time $t$ including the terminal time $T$.

128: 

129: It is numerically found that $T$ is linearly proportional to $N(T)$,

130: that is, $T \propto N(T)$. The average value of the

131: proportional coefficient $a$ for different items or articles listed 

132: in eBay is estimated to be $a \approx 1$ when the total number

133: of bidders $N(T)$ exceeds $20$. However, when the number of

134: bidders is lower, the proportional coefficient is very large, as

135: shown in Fig.~\ref{fig:K_N_each}. For the Korean auction, $a

136: \approx 4.5$, regardless of the number of bidders. On the other

137: hand, the bidding frequencies and the number of bidders for each

138: article are not uniform. Their distributions, denoted as $P_f (T)$

139: and $P_n (N)$, respectively, follow the exponential functions $P_f (T)\sim

140: \exp(-T/T_c)$ and $P_n(N)\sim \exp(-N/N_c)$, respectively, where

141: $T_c\approx 7.4$ and 10.8 for the eBay and Korean auctions,

142: respectively, and $N_c\approx 2.5$ and 5.6 for the eBay and 

143: Korean auction, respectively (Fig.\ref{fig2}).

144: 

145: %%%%%%%%%FIGURE

146: \begin{figure}[t]

147: \centerline{\epsfxsize=8cm \epsfbox{fig1.eps}}

148: \caption{Plot of $T$ versus $N(T)$ for the eBay.com (a) and the

149: Korean auction (b). The dotted line has a slope of 1 both in (a) 

150: and 4.5 in (b).} \label{fig:K_N_each}

151: \end{figure}

152: %%%%%%%%%%%%%%%%%%%

153: 

154: %%%%%%%%%FIGURE

155: \begin{figure}

156: \centerline{\epsfxsize=8cm \epsfbox{fig2.eps}} 

157: \caption{Plot of $P_f(T)$ versus $T$ in (a) and (c), and $P_n(N)$ 

158: versus $N$ in (b) and (d) for the eBay  (a) and (b) and Korean 

159: auction (c) and (d)

160: in semi-logarithmic scale. The dotted lines have slopes of $2.5$ in

161: (a), $5.6$ in (b), $7.4$ in (c), and $10.8$ in (d).} \label{fig2}

162: \end{figure}

163: %%%%%%%%%%%%%%%%%%%

164: %%%%%%%%%FIGURE

165: \begin{figure}

166: \centerline{\epsfxsize=8cm \epsfbox{fig3.eps}}

167: \caption{Plot of $\langle dk/dt \rangle$ versus $k/t$ for the eBay (a)

168: and for the Korean auction (b). The dotted lines obtained 

169: by the least square fit in the range [0.1:1] (a) and 

170: [0.01:1] (b), respectively, fit to the formula, $\approx 0.7k/t$.}

171: \label{fig:w_k_t}

172: \end{figure}

173: %%%%%%%%%%%%%%

174: 

175: %%%%%%\section{Stochastic process}

176: We introduce the master equation for the bidding process as

177: \begin{equation}

178: n_k(t+1)-n_k(t)=w_{k-1}(t)n_{k-1}(t)-w_k(t)n_k(t)+\delta_{k,1}u_t,

179: \label{discrete}

180: \end{equation}

181: where $w_k(t)$ is the transition probability that a bidder, who

182: has bid $k-1$ times up to time $t-2$, bids at time $t$. 

183: In this case, the total bid frequency of that agent up to time $t$ 

184: becomes $k$.

185: Note that a bidder is not allowed to bid successively. In the

186: master equation, we presume that the bidding pattern is similar

187: over different items when $N(T)$ is sufficiently large. Then,

188: ${w_k}(t)$ may be written as ${w_k}(t) \approx \langle dk/dt

189: \rangle$ on average over different items. Empirically, we find

190: that

191: \begin{equation}

192: {w_k}(t)\approx \langle dk/dt \rangle \approx bk/t, \label{kernel}

193: \end{equation}

194: where $b$ is estimated to be $b\approx 0.7$ for both the eBay and

195: Korean auctions (Fig.~\ref{fig:w_k_t}). The fact that $w_k

196: \propto k$ is reminiscent of the preferential attachment rule in

197: the growing model of the complex network~\cite{ba}. $u_t$ is the

198: probability that a new bidder makes a bid at time $t$. Using the

199: property that $\sum_k n_k (t)=N(t)$, we obtain 

200: \begin{equation}

201: u_t=N(t+1)-N(t).

202: \end{equation}

203: 

204: Next we then change the discrete equation, Eq.~(\ref{discrete}), to

205: a continuous equation as follows:

206: \begin{equation}

207: \frac {\partial n_k(t)}{\partial t}=-\frac {\partial}{\partial

208: k}\big({w_k}(t)n_k(t)\big)+\delta_{k,1} u_t, \label{continuous}

209: \end{equation}

210: which can be rewritten as

211: \begin{equation}

212: \frac {\partial n_k(t)}{\partial t}=-\frac{b}{t}\frac

213: {\partial}{\partial k}\big(k n_k(t)\big)+\delta_{k,1} u_t.

214: \label{continuous2}

215: \end{equation}

216: When $k > 1 $, we use the method of separation of variables,

217: $n_k(t)=I(k)T(t)$, thus obtaining 

218: \begin{equation}

219: \frac{\partial}{\partial k}(k I(k))+\ell I(k)=0, \label{eq_K}

220: \end{equation}

221: where $\ell$ is a constant of separation, and

222: \begin{equation}

223: \frac{\partial T(t)}{\partial t}= \frac {b\ell}{t}T(t).

224: \end{equation}

225: Thus, we obtain 

226: \begin{equation}

227: n_k(t)\sim t^{b\ell} k^{-(1+\ell)}.

228: \end{equation}

229: When $k=1$,

230: \begin{eqnarray}

231: \frac{\partial n_1(t)}{\partial t}=-\frac{b}{t}n_1(t)

232: +u_t.

233: \label{eq:n_1}

234: \end{eqnarray}

235: 

236: Next from the fact that $N=\sum_k n_k$, we obtain

237: \begin{eqnarray*}

238: \frac{\partial N}{\partial t} &=&

239: \sum_{k > 1}\frac{\partial n_k}{\partial t}+\frac{\partial n_1}{\partial t}\\

240: &=&\sum_{k>1}-\frac{b}{t}\frac{\partial}{\partial k}\Big(kn_k\Big)

241: -\frac{b}{t}n_1+\frac{\partial N}{\partial t}\\

242: &=& \frac{b\ell}{t}(N-n_1)-\frac{b}{t}n_1+\frac{\partial N}{\partial t}.

243: \end{eqnarray*}

244: Therefore, we obtain $N(t)=(1+1/\ell)n_1(t)$ and $n_1(t)\sim

245: t^{b\ell}$ by using Eq.~(\ref{eq:n_1}). Note that $N(t) < t$, and

246: the linear relationship holds asymptotically. The linear relationship 

247: breaks down for small $t$. From the empirical data,

248: Fig.~\ref{fig:N_t_fig}, we find that $\ell b\approx 1$. Since

249: $b\approx 0.7$ in Fig.~\ref{fig:w_k_t}, we obtain $\ell

250: \approx 1/b\approx 1.4$. Therefore,

251: \begin{equation}

252: n_k(t)\sim t k^{-2.4} \label{n_k} \end{equation} 

253: for large $t$, which fits reasonably with the numerical data 

254: shown in Fig.\ref{fig:N_DNST}.

255: 

256: %%%%%%%%%FIGURE

257: \begin{figure}

258: \centerline{\epsfxsize=8cm \epsfbox{fig4.eps}}

259: \caption{Plot of $\langle N(t) \rangle$ versus $t$, on average

260: over different items for the eBay data. The straight line has a 

261: slope of $0.7$ obtained from the least square fit.

262:  } \label{fig:N_t_fig}

263: \end{figure}

264: %%%%%%%%%%%%%%%%%%%

265: 

266: %%%%%%%%%FIGURE

267: \begin{figure}[b]

268: \centerline{\epsfxsize=8cm \epsfbox{fig5.eps}}

269: \caption{Plot of $n_k$ versus $k$ for the eBay auction 

270: at ebay.com (a) and for the Korean auction at

271: auction.co.kr (b) for various terminal times $T$. The solid lines

272: have a slope of -2.4 drawn for guidance.}

273: \label{fig:N_DNST}

274: \end{figure}

275: %%%%%%%%%%%%%%%%%%%

276: 

277: In eBay auctions, the winner is the last bidder in the bidding

278: sequence. Now, we trace the bidding activity of the winner in the

279: bidding sequence in order to find the winning strategy. To

280: proceed, let me define $q_{k}(t+1)$ as the probability that a

281: bidder, who has bid $k-1$ times up to time $t-1$, bids at time

282: $t+1$ successfully. Note that a bidder is not allowed to bid 

283: successively. In this case, $q_k(T)$ is nothing but the probability that a

284: $k$-frequent bidder becomes the final winner. The probability

285: $q_k(t+1)$ satisfies the relation,

286: \begin{eqnarray}

287: q_{k}(t+1)&=&

288: (1-u_{t+1})\sum_{j=1}^{N(t)}q_{j}(t)\frac{(k-1)(n_{k-1}(t)-

289: \delta_{j,k-1})}{t-j}\nonumber \\

290: &+&\delta_{k,1}u_{t+1}\label{eq:win}

291: \end{eqnarray}

292: with the boundary conditions $q_{1}(1)=1$ and $q_{1}(2)=1$. The

293: first term on the right hand side of Eq.~(\ref{eq:win}) is

294: composed of three factors: (i) $1-u_{t+1}$ is the

295: probability that one of the existing bidders bids successfully at

296: time $t+1$, (ii) $q_j(t)$ means that bidding at

297: time $t$ is carried out by the $j$-frequent bidder, and (iii) the last

298: factor is derived from the bidding rate, Eq.~(\ref{kernel}), where the

299: contribution by the bidder at time $t$ is excluded because he/she

300: is not allowed to bid at time $t+1$. The second term represents

301: the addition of a new bidder at time $t$.

302: 

303: The rate equation, Eq.~(\ref{eq:win}), can be solved recursively.

304: To proceed, we simplify Eq.~(\ref{eq:win}) by assuming that

305: $n_{k-1}(t)$ is significantly larger than $\delta_{j,k-1}$, which is

306: relevant when the number of bidders is large. Then,

307: \begin{widetext}

308: \begin{eqnarray}

309: q_k(t+1)& \approx & (1-u_{t+1})

310: \sum_{i=1}^{N(t)}q_i(t)\frac{(k-1)n_{k-1}(t)}

311: {t-i}+\delta_{k,1}u_{t+1}\nonumber \\

312: &=& (k-1)n_{k-1}(t)\prod_{\tau=2}^{t} (1-u_{\tau+1})

313: \Big[ \sum_{i=1}^{\tau-1} \frac {(i-1)n_{i-1}(\tau)}{(\tau-i)} \Big] q_{1}(2) \\

314: &+&(1-u_{t+1})(k-1) n_{k-1}(t)\sum_{\tau=3}^{t}  \frac{

315: u_{\tau}}{\tau-1}\prod_{\tau'=\tau+1}^{t}(1-u_{\tau'}) \Big[

316: \sum_{i=1}^{\tau^{\prime}-1} \frac

317: {(i-1)n_{i-1}(\tau^{\prime})}{(\tau^{\prime}-i)}\Big]+

318: u_{t+1}\delta_{k,1}\nonumber.

319: \end{eqnarray}

320: \end{widetext}

321: Since $1-u_t\approx 0.3 < 1$, $q_k(t)$ is obtained to be

322: \begin{equation}

323: q_k(t)\approx

324: (1-u_{t-1})\frac{(k-1)n_{k-1}(t-1)}{t-2}+\delta_{k,1}u_t

325: \end{equation}

326: within the leading order. Considering that $n_k(t)\sim

327: tk^{-2.4}$ in Eq.~(\ref{n_k}) and $u_t$ is constant, we obtain

328: $q_k(t)\sim (t-1)k^{-1.4}/(t-2)$ for large $k$ and $t$,

329: with a weak dependence on $t$. Thus, the winning probability

330: by the $k$-frequent bidder is simply given as 

331: \begin{equation}

332: q_k(T)\sim k^{-1.4} \end{equation}

333: in the limit $t\to \infty$. This result is confirmed by the

334: empirical data in Fig.~\ref{fig:winning}.

335: 

336: %%%%%%%%%FIGURE

337: \begin{figure}

338: \centerline{\epsfxsize=8cm \epsfbox{fig6.eps}}

339: \caption{Plot of the relative winning probability $q_k(T)/q_1(T)$

340: of the $k$-frequent bidder to that of the one-frequent bidder at

341: the last moment versus frequency $k$. 

342: The dotted line has a slope of -1.4 drawn for guidance.} 

343: \label{fig:winning}

344: \end{figure}

345: %%%%%%%%%%%%%%%%%%%

346: 

347: Our analysis explicitly shows that the winning strategy is to bid 

348: at the last moment as the first attempt rather than

349: incremental bidding from the start. This result is consistent

350: with the empirical finding by Roth and Ockenfels~\cite{roth} in

351: eBay. According to them, the bidders who have won the most items tend 

352: to wait till the last minute to submit bids, albeit there is some 

353: probability of bids not being successfully transmitted. 

354: As evidence, they studied 240 eBay auctions and found that 89 

355: bids were submitted in the last minute and

356: 29 in the last ten seconds. Our result supports these empirical

357: results.

358: 

359: In conclusion, we have analyzed the statistical properties of

360: emerging patterns created by a large number of agents based on the

361: empirical data collected from eBay.com and auction.co.kr. 

362: The number of bidders and the winning probability

363: decay in power laws as $n_k(t)\sim tk^{-2.4}$ and $q_k(t)\sim

364: k^{-1.4}$, respectively, with bid frequency $k$, which

365: has been confirmed by empirical data.\\

366: 

367: This work is supported by the KRF Grant No. R14-2002-059-01000-0

368: in the ABRL program funded by the Korean government MOEHRD and 

369: the CNS research fellowship in SNU (BK).

370: 

371: \begin{thebibliography}{99}

372: \bibitem{heck} E. van Heck and P. Vervest, Communication of the ACM

373: {\bf 41}, 99 (1998).

374: \bibitem{mantegna} R.N. Mantegna and H.E. Stanley, 

375: {\textit An introduction to econophysics:

376: Correlations and complexity in finance} (Cambridge University Press,

377: Cambridge, 2000).

378: \bibitem{bouchard} J.P. Bouchard and M. Potters, {\textit 

379: Theory of financial risks:

380: From statistical physics to risk management} (Cambridge University Press,

381: Cambridge, 2000).

382: \bibitem{stanley} M.H.R. Stanley, L.A.N. Amaral, S.V. Buldyrev, S. Havlin,

383: H. Leschhorn, P. Maass, M.A. Salinger and H.E. Stanley,

384: {Nature} {\bf 379}, 804 (1996).

385: \bibitem{challet} D. Challet and Y.-C. Zhang, {Physica A}

386: {\bf 246}, 407 (1997).

387: \bibitem{pennock} D.M. Pennock, S. Lawrence, C.L. Giles and F.A. Nielsen,

388: {Science} {\bf 291}, 987 (2001).

389: \bibitem{hulst} R. D'Hulst and G.J. Rodgers, {Physica A} {\bf 294}, 447 (2001).

390: \bibitem{yang} I. Yang, H. Jeong, B. Kahng, and A.-L. Barab\'asi,

391: Phys. Rev. E {\bf 68}, 016102 (2003).

392: \bibitem{simon} H.A. Simon, {Biometrika} {\bf 42}, 425 (1955).

393: \bibitem{zhang} M. Marsili and Y.-C. Zhang, {Phys. Rev. Lett.} {\bf 80},

394: 2741 (1998).

395: \bibitem{albert} R. Albert and A.-L. Barabasi, {Rev. Mod. Phys.} {\bf 74}, 47

396: (2002).

397: \bibitem{ba} A.-L. Barabasi and R. Albert, {Science} {\bf 286}, 509 (1999).

398: \bibitem{pareto} V. Pareto, Cours d'economie politique (Rouge,

399: Lausanne et Paris, 1897).

400: \bibitem{lb} Y. Shoham and M. Tennenholtz,

401: Games and Economic Behavior, {\bf 35}, 197 (2001).

402: \bibitem{kauf} R.J. Kauffman and C.A. Wood, {Proc. of ICIS} (2000).

403: \bibitem{roth} A.E. Roth and A. Ockenfels, {American Economic Review}, 

404: {\bf 92}, 1093 (2002).

405: \end{thebibliography}

406: \end{document}

407: 

408: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

409: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

410: %%%%%%%%END OF DOCUMENT%%%%%%%%%

411: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

412: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

413: