News + Publications

Demand forecasting, predicting the quantity of items desired at a future date, plays a key role in many businesses. Traditionally, the predicted items are goods and services for a consumer market, but the same techniques can be transferred to other areas such as predicting hospital admissions or the number of participants requiring access to a homeless shelter. Demand forecasting has a long history, using both classical statistic methods (e.g. ARIMA models) and more modern machine learning methods including time-series specific forecasting (e.g. LSTMs).

These methods can be adapted depending on the data and forecasting goal. For instance, goods data are often hierarchical where the sales are reported geographically by province, region, and individual store; and grouped by main product type, a sub-category, all the way down to an individual product number. In these cases, using both the aggregate product type and regional data, and the finer product number and store data together can produce more tailored models. These models can predict both national sales and the amount of a specific product to stock at a particular store.

In some scenarios, the demand will be greater than the available stock or capacity. A provider needs to be able to react to such situations by increasing inventory or adjusting the sales forecast. One example is providing beds at homeless shelters. If a particular shelter can predict that they will have a demand for more beds than they have available, they can plan ahead and liaise with other nearby shelters. A similar example is related to hospital intakes with the goal of predicting both the number of intakes, how many of those expected intakes will require a bed, and their likely length of stay.

Accurate forecasting is key to many industries and businesses. The nuances of a particular application – available data, specific prediction goal, extra conditions (e.g. storage space, capacity, etc.) – mean that there is no universal “best” model and that the best performance is likely obtained from a customized, application-specific model.