Part 2. MM4XL Tools > 1. Strategic Tools > Forecast Manager > 3. Technicalities > Forecasting Technique Selection

Forecast Manager

Forecasting Technique Selection

The selection of the appropriate forecasting technique depends on several critical factors, some of which are more common then others. In our experience, 3 factors are worth considering in this help manual: forecast horizon, data pattern, and level of accuracy.

Forecast horizon

This is the number of periods the forecast should go into the future after the last known value. Typically, decision-makers are interested in one of the following:

  1. Short-term forecast = one to six periods.
  2. Intermediate-term = seven to 12 periods.
  3. Long-term = beyond 12 periods.

The long-term forecast tends to relate to trend factors (e.g., product demand, market size, industry structure, etc.). The short-term is tied to seasonality and cyclical variations. In general, long-term horizon forecasts find causal methods (regression) more valuable and autoregressive methods become less valuable. In the short-term, however, when dealing with stable series (that exhibit few turning-points) autoregressive methods may become very powerful.

Level of accuracy

It is dependent on the project for which the forecast has to be made. In some cases, a rough approximation of the trend pattern may be enough to the end user. When a high level of accuracy is required, Forecast Manager makes available all indices and charts needed for judging whether the model is accurate enough or a better one can be devised among the 14 available models.

Tip:
After a visual inspection of all fitted curves you find that other models predicted the latest part of your input data better than the best curve chosen automatically. In this case, compute again the fit measures in the hidden sheet by reducing the number of rows in the formulae and find out if you are right and a different model than the selected one should be used.

The accuracy of forecast can be inspected visually or tested by means of statistical measures. In both cases the process is based on the analysis of the error associated with each forecasted items, and its ultimate goal is to test how well one model forecasts (fits the actual curve).

The forecast error is computed with the formula:

Sales Forecast Software with Sales History Data

The smaller the error term

Sales Forecast Software with Sales History Data
the better the forecast.
Sales Forecast Software with Sales History Data
represents the actual sales level and
Sales Forecast Software with Sales History Data
the forecasted one. However, one model can forecast better under certain circumstances and less well under others. Therefore, it is highly recommended to test the quality of forecast by means of several measures and to compare forecasts obtained with different models, as we shall see.

Graphical methods inspect the reliability of forecasts quickly and accurately, and they help to identify systematic error patterns produced by the model.

Sales Forecast Software with Sales History Data

The pictures above show how the same actual data were fitted with different models. Sometimes one cannot recognize by simple visual inspection, which model is best, so plotting error terms helps in discerning among the various methods. When the model forecasts accurately the error terms are randomly scattered around the zero error-line.

There are three most common ways for inspecting forecast reliability by means of graphical methods:

  • Plotting cumulative error terms (called CuSum chart in Forecast Manager).
  • Plotting error terms (called Special Events chart in Forecast Manager).
  • Using turning-point diagrams.

Data pattern

Trend and seasonality are the two core elements of any short-term forecasting exercise made by means of autoregressive models. Both elements can be found or not in a time series. When they are present, also separately, they take either an additive or a multiplicative form. Additive trend or seasonality is one that increases over time at a regular rate (panel A and B in the picture below). On the other side, a multiplicative trend or seasonality increases at a faster rate than in the past (panel C and D).

Sales Forecast Software with Sales History Data

In general, regression models can replicate any pattern, given the forecast manager can identify, measure, and gather the relevant independent variables that explain the model. When forecasting short-term, the autoregressive and regression models available in Forecast Manager offer a valuable solution to fit any curve form resembling the curves shown above. However, unstable series, series that show a high number of turning-points, may produce weak results.

Lifetime license:

Price: euro

Vote this tool
We proudly serve
Your vote
vote1 vote2 vote3 vote4 vote5
Email:
Gender:
M
F
Age:
Position:
Department:
Comment: