Daily Forecasting including Call Centre Forecasting
Daily forecasting, of course, has the complication that there is usually a profile of day within week that needs to be taken into account in the demand forecast in addition to normal seasonality. Examples where daily forecasting may be necessary include supply planning for short shelf-life products, such as bread and cakes, or for short-term workforce planning in call centres.
The presence of two different demand cycles i.e. week-year and day-week, makes it difficult to use a single forecasting method for daily forecasting. The usual approach is to create a set of forecasts for total week, incorporating any weekly seasonality, then split those forecasts into days based on the profile of day within week.
Forecast Solutions can help with daily forecasting and with call centre forecasting.
Achieving the Day-Week Profile
One easy way to proceed is to calculate a set of daily indices with a very simple analysis across a period of time, for example the latest 12 months. So the index for Monday is calculated as the average of the Mondays divided by the average of the complete weeks. A minor adjustment should be made to determine that the daily indices add to 1.00. Alternatively, simple exponential smoothing can be used to track the daily proportions.
Another approach sometimes used, if suitable software is available, is to firstly do a daily forecast using exponential smoothing with seasonal adjustment (for example the Holt Winters method), with a seasonal periodicity of seven days to represent day of week. Then a top-down force is carried out from a separate forecast that has been prepared at the total week level, thus forcing in the weekly trend and seasonality.
The final forecast is calculated as the weekly forecast multiplied by the daily indices.
With causal modelling, this could potentially be carried out on weekly data, then splitting to days using daily profiles, but in most instances the need will be to do the causal analysis at a daily level. A decision is needed as to whether to seasonally adjust the historical data that is being used, prior to the regression analysis, and how to deal with any unrelated events or trends e.g. a pandemic.
The causal analysis can then be conducted at the daily level with exploration of time-lagged responses as necessary. The daily profile can be removed prior to causal analysis or can be dealt with in the linear regression using dummy variables for each day of the working week. It may be possible to include certain events, such as bank holidays, as additional causal factors.
It is best to use statistical software for the regression analysis, rather than Excel, because the causal analysis needs the capability to deal with time-lagged responses and with correlations between causal factors. A combination of time series and causal modelling methods is quite often required and the method used for each such project should be carefully considered.
Forecasting System for Daily Forecasting
In a forecasting system built on time series forecasting methods there is an ongoing need to incorporate the latest daily actual data and to revisit at appropriate intervals the seasonal indices and daily profile.
If causal modelling is being used there needs to be a regular update of the latest actual data and a forecast for the causal factors. With weather sensitivity models a forecast of the weather measures is only good over a short horizon, but a longer term demand forecast can be made on a 'weather neutral' basis using average seasonality.
Forecasting in Excel can be a good option, particularly if a mix of time series and causal models is needed. If the number of entities to be forecast is large there may be a need to consider specialist forecasting software.