ABC Process-Based Capital Budgeting.

Managers were very concerned about the fact that net cash flows were negative for the first six years as shown in Table 1. Even though the NPV was positive when considering low discount rates and including the residual value, managers felt that the choice of a discount rate did not adequately capture the project risk because of the distant payoffs.

ABC -- BASED SIMULATION OF THE CYBERMALL PROJECT

Thus, due to the limitations noted above, a simulation approach was used to analyze the major uncertainties in the base case. The cash flow model in Table 1 is a deterministic approach with a single set of estimates. As such, the managers cannot study the risks they saw in the project. A probabilistic model is more powerful than a deterministic one because uncertain cash flow items are assigned probability distributions. Using Monte Carlo simulation, distributions of uncertain cash flows are generated and an output variable such as NPV is sampled for each iteration of the simulation. This process is continued for each iteration of the simulation.

In a typical simulation study (Meimban et. al., 1992) original estimates of cash flow items are varied by some amount and variation in NPV is examined. The items being varied are the revenue and expense items from a typical income statement. The drawback to using these items is that one is studying the result of many underlying decisions and processes and not the processes themselves. Although the ABC approach was developed to analyze historical cost relationships, it also provides a very powerful framework for conducting a risk analysis of proposed capital investments. The ABC approach allows the manager to vary the underlying activity in order to study the impact of different levels of the activity itself. Managers have the potential to learn much more about the inherent risks of their decision when they study the uncertainty in the business processes, rather than highly aggregated items such as total revenues, labor costs, and material costs.

Simulation Model

Accordingly, a model was constructed to study the underlying risks of the cybermall decision. First, the various Excel spreadsheets in the model were linked together in order to run effective simulations on the major uncertainties previously cited. In a few instances, the base case was found not to be fully integrated. Thus, the original model was extended into a general simulation model with appropriate linkages to all of the underlying activity drivers as shown in Figures I and II. Interviews with managers determined which activity drivers to simulate and the bounds on the variability of the drivers. Table 2 summarizes these major sources of uncertainty, the linkages in Figures I and II, and the assumptions about the probability distributions assigned to them.

There were three general problems to deal with in developing the inputs to the simulation model. First, probability distributions for the various input variables were needed. Going from point estimates to probability distributions is difficult for managers, and managers at this company were no exception. Specifying a range of possible values is one thing, but choosing between a normal distribution and (say) a geometric probability distribution is another. For example, the normal distribution requires the analyst to specify both the mean and standard deviation of the distribution, as well as assume symmetry about the mean. Since managers generally were willing to estimate pessimistic, most likely, and optimistic levels for the inputs, triangular probability distributions were specified. A triangular distribution is defined by three values: the minimum, the most likely and the maximum. This distribution is often used in simulation studies when the analyst does not know anything more about the probability distr ibution of a variable than these three values. In an ideal world, each distribution would be fit to real data using the method of maximum likelihood (Law et. al., 1994), but there was no data to collect for the cybermall project. It was an entirely new line of business. There was little choice but to use the triangular distribution, even though using it when the true distribution is of another type can lead to large errors in the simulation results (Law et. al., 1994).

The second problem concerned the modeling of the interactive adoption percentage, or rate, over time. In the base case model, analysts in the company estimated percentage amounts each year over the life of the project. In general, the pattern of adoption of new products tends to have the same general shape across a wide range of new products. Early adoption occurs slowly but then increases over time. In the middle part of the process, adoption occurs at a rapid rate, but then slows down during the final phase of the adoption cycle. When plotted in a graph, these estimates form an S-shape that is characteristic of the cumulative adoption rates associated with other new product introductions (Moore and Pessemier, 1993). The adoption rate was modeled over time using the Fisher-Pry equation which gives a very close fit for a wide range of new product introductions (Moore and Pessemier, 1993, and Meade and Towhidul, 1995). Because the adoption rate is so important to the revenue projections, parameter estimates f or the Fisher-Pry model developed by company planners were also varied in the simulation to investigate the sensitivity of the results to model specification.

The third issue in setting up the simulation model was to develop a model for forecasting the number of video servers. In the base case model, these estimates were simply inputted as individual numbers in the interactive deployment schedule that was not linked to the project. A regression model was developed for forecasting the number of video servers needed each year. In the regression, the number of video servers was regressed against the total number of subscribers. The coefficients from the regressions were used in the simulation model to link the number of servers projected in the interactive deployment schedule to the number of avenue shoppers or subscribers.

What-If Analysis

After the simulation model was developed, the next step was to decide how many of the input variables were important. The TopRank software from the Palisade Corporation (1996) was used to perform a what-if analysis on the variables of interest (see Table 3) to determine which ones should be studied further. Using built-in functions available in the TopRank software, each input in Table 2 was varied from its minimum to maximum value. The software computed a value for NPV for each different value and saved it for later analysis. After all of the inputs had been varied across the specified ranges, TopRank analyzed the impact of each variable on NPV using regression analysis. A graph in Figure III (called a Tornado diagram) of the standardized regression coefficients displayed (in descending order of magnitude) the relative importance of the input variables to NPV. The most important variables were those with the largest coefficients (in absolute value). The standardized regression coefficient showed the number of standard deviations by which NPV increases if the independent variable increases by one standard deviation. For example, increasing the number of cable subscribers by one standard deviation increased NPV by 1.25 standard deviations.

The Tornado diagram for NPV with the residual value is in Figure III. Two key inputs for the NPV were identified in order of importance: the number of cable subscribers and the rate at which the interactive adoption percentage grew over time. The next step was to conduct a full simulation study of the model with these two variables as the key revenue inputs.

Simulation Analysis

Simulations were performed using the @Risk software, also from Palisade Corporation. Because simulation analysis incorporates risk in the variability of the cash flows, a risk-free interest rate was used as the discount rate to avoid double-counting risk (Brealey and Myers, 1996). The long-term Treasury Bond rate was used as an estimate of the riskless interest rate for the simulation analysis. At the time of the study, rates were about 8 percent. Since NPV is the net of several offsetting items, present values (PVs) were included to help interpret the simulation results for NPV, as shown in Panel A of Table 3. Inputs were modeled per Panel B of Table 3. NPV with a residual value, PV of Revenues, PV of ABC costs, PV of Variable costs, PV of Fixed costs, and PV of capital costs were selected as the output variables of the simulation.

It is clear that the residual value is a significant factor to the project. Without a residual value included, the mean NPV is statistically near zero at $28 million and a standard deviation of $70 million. On the other hand, including the residual value increases the mean NPV to $518 million and the standard deviation to about $348 million. One could conclude that the project has a positive net present value only if the business could be sold to generate a positive residual value. Also, including a residual value generated a 97.6 percent probability of a positive NPV. However, there was only a 41 percent probability of a positive NPV without a residual. Thus, there appears to be a large amount of risk related to the existence of a residual value.

Scenario Analysis

Output from the @Risk software explored the uncertainty in cash flows in several different ways. First, the annual cash flows over time were plotted using what is known as a Summary Graph (Figure IV). The line through the middle of the graph shows the mean level of annual cash flow, the inner band shows an interval equal to plus/minus one standard deviations from the mean, and the outer band shows an interval equal to plus/minus two standard deviations from the mean. With a 95 percent probability, annual cash flows were negative for the first six years. Thus, the initial negative cash flows observed in Table 1 appeared to be a general pattern for this project and positive NPVs were generated by cash flows occurring over the last few years of the project. This in turn suggests that the number of cable shoppers and the growth in the percentage of avenue subscribers were the key revenue variables crucial to the success of this project, as analyzed in the Tornado diagram in Figure III. This type of scenario anal ysis has also been advocated as a key component of a business plan (Sahlman, 1997).

Print Get the Mag Weekly Updates