BUSI 3010 Week 8 Forecast Method Comparison Example

Forecasting Monthly Kitchen Jobs at a Stone Fabrication Shop: A Three-Month Moving Average Against Exponential Smoothing at Alpha 0.3, Compared on Fifteen Months of Error, and Why Both Run Low in Spring

Student Name

Walden University

BUSI 3010: Operations Management

Instructor Name

Month Day, Year

Forecasting Monthly Kitchen Jobs at a Stone Fabrication Shop: A Three-Month Moving Average Against Exponential Smoothing at Alpha 0.3, Compared on Fifteen Months of Error, and Why Both Run Low in Spring

The Series and the Purpose

The countertop shop in this course needs a monthly forecast of kitchen jobs to plan CNC hours, installation crews and slab orders. The series below is the number of jobs sold each month from January 2027 through June 2028. A job is one kitchen order of about 55 square feet. The question is which of two simple methods fits this history better, judged by forecast error over the same fifteen months, April 2027 through June 2028.

The Two Methods

Under the three-month moving average, a month's forecast is simply the mean of the preceding three months, so April 2027 is the first month it can cover. Exponential smoothing forecasts each month as the previous forecast plus a fraction, alpha, of the previous month's error. Alpha was set at 0.3, a common middle value that responds to recent months without chasing every swing, and the first forecast was set equal to January's actual demand. Gardner (1985) reviewed the method's many variants and its record of performing well for short-range forecasting, which is why it is a standard comparison for a moving average.

Forecasts and Errors

Table 1

Monthly Jobs Sold and Forecasts, January 2027 to June 2028

Red River Stone & Surface and its sales records are composite. Forecasts are rounded to one decimal place.

Error was measured for April 2027 through June 2028, the fifteen months both methods cover. Three measures were used. Mean absolute deviation (MAD) is the average size of the error in jobs. Mean absolute percentage error (MAPE) expresses that error as a share of actual demand, which makes it easier to judge. The running sum of errors, actual minus forecast, reveals any steady lean toward overforecasting or underforecasting. Hyndman and Koehler (2006) caution that percentage errors can mislead when actual values are near zero; here monthly demand never falls below 48, so MAPE is safe to use.

Table 2

Forecast Error, April 2027 to June 2028

A positive sum means actual demand ran above the forecast.

Which Fits Better

The three-month moving average fits this series better on both accuracy measures: it misses by about 7.8 jobs a month on average, or 12 percent, against 8.7 jobs, or 13.3 percent, for exponential smoothing. With alpha at 0.3, exponential smoothing gives weight to months further back than three, so it reacts more slowly when demand climbs in spring and falls in late fall.

Neither method is good, though, and the sum of errors explains why. Both sums are positive and large, meaning both methods forecast too low most of the time. The error is concentrated in two periods: the spring climb from March to June, when each forecast lags rising demand, and the fall decline, when each forecast stays too high for a month or two. This series has a seasonal pattern, higher in late spring and summer, and an upward trend from 2027 to 2028: the first half of 2028 averaged 65.7 jobs a month against 60.8 in the first half of 2027. Methods that only average the past cannot anticipate either one.

What the Shop Should Use

For next month, July 2028, the moving average forecasts 74.0 jobs and exponential smoothing 69.6. Given the methods' record of running low in early summer, the moving average is the better of the two for short-term planning, and the shop should plan CNC hours for at least 74 jobs.

For anything beyond a month, neither method is adequate. A method that models trend and seasonality, such as Holt-Winters exponential smoothing, would fit this pattern better once two full years of history are available in January 2029. Makridakis and Hibon (2000), comparing forecasting methods across thousands of series in the M3 competition, reported that sophisticated statistical techniques were often no more accurate than simple ones. That finding argues for adding seasonality only because this series clearly shows it, not for complexity in general.

Limits

Eighteen months is a short history, and one unusually busy month would shift both forecasts. A single large builder contract, for example a twelve-unit townhouse project, would show up as a spike that neither method should be allowed to treat as normal demand, so such orders should be recorded separately. The comparison also measures fit to past data, not accuracy on future months; a fairer test would hold back the last six months, fit each method on the first twelve and compare forecasts against what actually happened.

References

Gardner, E. S., Jr. (1985). Exponential smoothing: The state of the art. Journal of Forecasting, 4(1), 1-28. https://doi.org/10.1002/for.3980040103

Hyndman, R. J., & Koehler, A. B. (2006). Another look at measures of forecast accuracy. International Journal of Forecasting, 22(4), 679-688. https://doi.org/10.1016/j.ijforecast.2006.03.001

Makridakis, S., & Hibon, M. (2000). The M3-Competition: Results, conclusions and implications. International Journal of Forecasting, 16(4), 451-476. https://doi.org/10.1016/S0169-2070(00)00057-1