Forecasting

Whenever planning future activities of your business, you will almost immediately be confronted with the problem of predicting the development of your major business parameters. Based on medium term demand figures you will have to plan your production and transportation capacities, based on the anticipated success of your online activities you will buy equipment or hire IT-staff. In the short term, you may want to be able to know how much of which products to put on your shelves in the following weeks, and in which areas you will need to increase the number of temporary workers.

In the big picture, you already have developed an intuition about how to treat these problems, usually based on your and your coworkers' experience from comparable situations. By experience you know the basic drivers of the developments, such as

• long time trend
• yearly or weekly patterns / seasonality
• patterns for specific clients or regions

• special, nonrecurring orders,
• holidays,
• changes in price
• changes in the economic climate of your industry

Thus, when making plans in the medium to long term, a more or less intuitive approach to predicting the future developments is to

• clean your data of non-recurrent effects (if you know about them),
• take into account seasonality,
• make a rough assumption about the future general market environment,
• infer the trend in your data,

For the short term, sales or bookings of the last weeks give a good indicator of things to come, and not few of your larger clients may even be in the good habit of announcing larger orders beforehand or warn you, if they may need to cancel.

For purposes of strategy building, in the purely long term picture, you may choose an entirely different, more hypothesis-based approach, sketching out a number of scenarios based on your assessment of the main forces in your industry.

While these long term problems also lend themselves to a quantitative approach (after all, even if with hypotheses only you have to keep track of inputs in a systematic way), it is in the short to medium term that statistical forecasting methods or time series models show their real potential. Even if the average planning figures are well-founded, a lot can go wrong on a daily or weekly scale: Flights are overbooked, trucks run at half their capacity, shelves remain empty, and staff has to be paid expensive extra hours due to erroneous short term volume predictions. In short: Though, as we have seen, your actual forecasting is probably not bad at all, even a small improvement will have a direct impact on your costs.

Again, most of the remaining part will be explained by known drivers in your business. The task is to make a list of everything that may contribute to explain the observed data, like

• known external influencing factors
• rules known from experience
• the importance of the recent history on the current figures
• types of influence
• etc.

The result is a forecasting model that is as precise as possible and as close as possible to your business experience. Be aware that a certain amount of randomness is always to be expected to be present in the data and no forecasting method in the world will be able to remove it. By testing the model on past data you will also have a diagnosis of the typical forecasting error. Also, by coding the forecasting rules that were used before, the performance of the old and the new systems can be compared, giving you an exact idea of the what application of the new forecasting method will mean money-wise.

The forecasting model can then be integrated (with or without error margins) into your planning tools. Usually that will be done in a way to allow users to overrule the automatically generated figures, if these seem counterintuitive or if additional information, like an order-announcement by a customer, that is not in the model has become available.