BFO 7010 Module 1 Discussion: What Can and Cannot Be Forecast Example

Module 1 Discussion: Gallons, Not Guesses

Gallons, Not Guesses: Deciding What a Propane Distributor Can Forecast

I work in dispatch at a propane distributor in South Jersey, and Module 1 gave me words for something our drivers already know: some things we can forecast well and some we cannot. Hyndman and Athanasopoulos (2018) say predictability depends on how well we understand the factors involved, how much data we have, whether the future resembles the past and whether the forecast itself changes what it forecasts.

Delivery volume passes those tests. We know what drives it: cold weather, the number of customers and, more slowly, efficiency upgrades. We have six years of monthly gallons. Winter follows the same pattern every year, and our forecast of gallons does not change how cold January turns out. Wholesale propane prices fail at least two tests. They depend on crude oil, exports and storage across the country, factors we cannot observe well, and traders' expectations move prices directly, so any widely shared forecast feeds back into the market. We should forecast volume and hedge prices, not try to predict them.

The module also separates three things that our managers blur together. A forecast is our best estimate of what will happen; a goal is what we want to happen; a plan is how we respond to both. Last year the sales manager's goal of 5 percent growth became the volume forecast for the delivery schedule, and we hired too few winter drivers when the cold came anyway. Fildes and Goodwin (2007) describe how organizational pressures let targets seep into forecasts, and our schedule was a case of exactly that.

Setting up a proper forecasting task would follow the book's steps: define the problem (monthly gallons, a year ahead, for driver and truck planning), gather data and the judgment of experienced dispatchers, explore the series with graphs, choose and fit a few models, then use and evaluate the forecasts as actual gallons arrive. The M4 competition found that simple and combined methods often match complex ones (Makridakis et al., 2018), so we can start simple.

Reply to Deshawn

Deshawn, your example of forecasting emergency room arrivals at the hospital where you work fits the predictable side well: arrivals follow daily and weekly cycles, and there are years of hourly records. I would add that the forecast changes over a shorter horizon than ours. You need to know tonight's arrivals by hour, while we need next winter's gallons by month, so the same principles lead to very different data and models. Have you found that holidays break the weekly pattern, the way a warm Christmas week breaks ours?

Reply to Karen

Karen, you wrote that a forecast that misses is a bad forecast. I see it differently after this module. A forecast is a distribution, not a single number, and Hyndman and Athanasopoulos stress reporting prediction intervals. If our January forecast is 620,000 gallons with an interval of 560,000 to 680,000, and January brings 670,000, the forecast did its job: it told us to plan for the high end. A forecast is bad when its intervals are too narrow or when it is biased in the same direction month after month, not simply when one month lands away from the center.

References

Fildes, R., & Goodwin, P. (2007). Against your better judgment? How organizations can improve their use of management judgment in forecasting. Interfaces, 37(6), 570-576. https://doi.org/10.1287/inte.1070.0309

Hyndman, R. J., & Athanasopoulos, G. (2018). Forecasting: Principles and practice (2nd ed.). OTexts.

Makridakis, S., Spiliotis, E., & Assimakopoulos, V. (2018). The M4 Competition: Results, findings, conclusion and way forward. International Journal of Forecasting, 34(4), 802-808. https://doi.org/10.1016/j.ijforecast.2018.06.001