Information Technology Reference
In-Depth Information
Ta b l e 5 .
Customer sales statistics for each agent
Agent
PCs sold PCs bid on Percent Won ASP
FreeAgent
54660
166293
0.33 0.90
Mr.UMBC
56748
172517
0.33 0.83
UMTac-04
56341
167746
0.34 0.74
Botticelli
17290
28581
0.60 0.81
Deep Maize
33166
124485
0.27 0.92
SouthamptonSCM
41798
93936
0.44 0.85
they did. That the three agents bid on almost identical fractions of the market and win
similar fractions of bids suggests that this difference in ASPs is due to targeting different
types of markets.
Deep Maize
had the highest ASP of any agent and the lowest winning
percentage, suggesting that this agent made systematically higher offers to customers.
Conversely,
Botticelli
won a very high fraction of its bids with a relatively low ASP,
suggesting systematically lower bids.
SouthamptonSCM
won a high fraction of bids
with mid-range ASPs, but bid on a much smaller fraction of the market than the other
three agents with similar component purchases.
4.2
Market Behavior
To better understand why the agents had different ASPs, we consider features that dif-
ferentiate markets along the dimension of ASP. For this analysis we consider a “market”
to be the set of customer requests for a PC type on a simulation day. We identify four
factors that are strongly correlated with overall market ASPs. Figure 5 shows these rela-
tionships as scatter plots, with superimposed lines representing binned averages. Except
for simulation day, all of these factors are measures of supply and demand motivated by
basic economic principles. The simulation day is important primarily due to start- and
end-game effects.
Plot 5(a) shows the relationship between prices and simulation day. Prices start very
high early in the game as agents build inventory and decrease over time. At the end
of the game prices can fall very low as agents try to recover some value for excess
inventory. The second factor, shown in 5(b), is market demand (i.e., total quantity re-
quested). Prices increase as demand increases, with any demand level less than 100
occasionally subject to very low ASPs. Plot 5(c) is a measure of bid density calculated
this by summing the number of bids for each individual PC and dividing by the total
number of PCs requested. ASPs fall approximately linearly as bid density increases.
ASPs also fall approximately linearly as manufacturer PC inventory increases, as seen
in 5(d).
We ran linear regressions to test the strength of these relationships with ASPs (all
of the plots suggest a linear relationship, so this is a reasonably approximation). The
individual
R
2
values were 0.29 for simulation day, 0.13 for market demand, 0.42 for
bid density, and 0.44 for PC inventory. A multiple regression using all four factors
yields an
R
2
value of 0.65, with all coefficients significantly different from 0. We could
certainly improve this model by considering additional factors (e.g. reserve prices, lead
times, smoothed market demand). However, a simple linear fit to these four variables is
