I’m often asked how I forecast workforce needs. How do I know how many people need to be working in our call centers or customer care offices at some given time in the future?
Forecasting seems like the “reading tea leaves” of business functions on the surface. There’s no established consensus on how it should exactly be done. It’s hard to get such a consensus because the truth is that forecasts are nearly always wrong. And there are many different kinds of forecast needs: long-term vs short-term, aggregate vs individual, etc.
However, most forecasters generally work with a pre-existing pool of techniques. Some businesses choose to be subjective. They use methods like sales force composites (essentially a collection of opinions), customer surveys or leadership opinions. Your business needs may warrant the use of such methods because of how fast and easy they are. The value for time + effort is great. However, if your competition uses more sophisticated and data-driven forecasting, you may not have the margins to play fast-and-loose with your forecasting. I’ll spend the rest of this post talking about the data-driven forecasting methods.
There are two kinds of models when we talk about data-driven forecasting: causal models and time series models. Real-life forecasting often mixes and matches the two models to create what uniquely works for the business need in question. Causal models use the cause-and-effect principle. When I managed the forecast at Sears, I would see increased workforce need in the days and weeks before school would start and Christmas (among other times). This was due to increased demand from shoppers for the back-to-school and Christmas seasons. Workforce needs were also affected if our top manufacturing partners increased or decreased their prices for their most popular products – because this too affected foot traffic and ultimately purchases in the stores. But how does this help me forecast for business needs during non-seasonal periods or when causality is not apparent?
This is where time series models come in. I call this the bread-and-butter of forecasting. Most base forecasts use a time series model to build out future data points. The most popular time series model is linear regression. This is great for assessing trends. Are workforce needs generally increasing or decreasing with time? Those are trends I can assess with linear regression. In the healthcare space though, demand for our services can be fairly stable. It is preferable in those cases to use the moving average time series model. This takes the average of the last few observations instead of drawing a line through them (as in linear regression). And if demand is constantly changing, then the more intricate exponential smoothing method can be used, which takes the weighted average of the last forecast with the current value of demand.
Some final thoughts: trend and seasonality must be accounted for in most time series models. Linear regression helps with assessing trend and seasonality requires building a normalized seasonal index. While not overly complex, this new rabbit hole is best explored in a different blog post. Finally, while forecasts are always wrong, you can measure their accuracy through the following metric: Mean Absolute Percentage Error (aka MAPE). I hope you’ve seen by now that forecasting is a little more involving and interesting than reading tea leaves!