“The only thing I cannot predict is the future.”
― Amit Trivedi, Riding The Roller Coaster: Lessons from financial market cycles we repeatedly forget
It goes without saying that every business is keenly interested in knowing what the future will bring.
Will sales grow next year? By how much? Will suppliers increase their prices? How fast will be the adoption of a new IoT product? How much warehouse capacity is needed for the next holiday period? Will some international event plunge the global economy into a recession?
Predicting the future is an exercise in probability rather than certainty. Businesses engage in various levels of sophistication in trying to bound the likelihood of future states to support their business plans.
Some have teams of economists and data scientists tasked with building complex forecasting models.
Many businesses, however, likely rely on less sophisticated means centered on spreadsheet models, trends and moving averages (or even educated guesses).
Time series methodology is a moderately sophisticated yet cost effective way to generate forecasts. It is a statistical approach which bases forecasts on the past behavior of the data series in question (e.g. monthly sales).
And it accounts for other characteristics of a time series which can yield a more accurate forecast than, say, a simple straight-line trend model.
More time, more data
We have all heard the forecasts about data growth, the proverbial “hockey stick.”
By one account, human and machine-generated data is growing at 10x the rate of traditional business data. And machine-generated data is growing at 50X the rate.
A washing machine can monitor, collect and post performance data to the cloud. Using these data to forecast product failure can lead to a pro-active maintenance visit by your friendly but lonely Maytag repairman.
Similarly, a household’s electricity usage can be monitored, modeled and forecasted leading to cost-savings suggestions by an energy service provider under a time-of-day pricing scheme.
The electric utility company itself can use the data from all the IoT appliances in households to generate better residential load forecasts and help better manage the electricity grid.
Even if a more traditional businesses source like sales, inventories, deliveries, workforce utilization, IT usage and the like, advances in data collection, storage and proliferation are making time series data more readily accessible.
Thus, there will be an increased demand for product managers, economists, statisticians and data scientists to make use of these data and tell us what will happen next.
Time series methods
The premise of time series methods (and of most quantitatively-based forecasting methods) is that the future will be much like the past.
If sales have been growing at a consistently healthy rate with strong seasonal variation (e.g. holiday periods) for the last year, then it is likely the next year will be similar, all else constant. If done correctly, the methodology can yield a defensible forecast of likely sales each month during the “forecast horizon.”
But, as with all forecasting methodologies, there are pitfalls of which one should be aware.
Practical time series methods
This is the first of a series of articles on practical time series methods for short-run business forecasting.
There are abundant, excellent resources covering the basics of business forecasting including time series methods, ranging from blog posts to online courses to open-source textbooks.
And time series methods are a mainstay of advanced courses in econometrics and business forecasting (resources we recommend are Elements of Forecasting by Diebold and Econometric Models and Economic Forecasts by Pindyck and Rubinfeld).
Rather than being a treatise on forecasting, this series of articles will present a practical methodology and some of the lessons we have learned performing time series forecasting for clients.