Sales Forecasts

Sales forecast is a requirement to any attempt on Supply Chain planning.

We often notice the following characteristics:

• The forecasts are not reliable

• The more we work on a forecast at an aggregated level, the more reliable it is.

• The more the horizon of forecast is long, the less it is reliable

A- Forecast methods

1. Forecast horizon

Forecast horizon varies depending on its utilisation:

• Short term forecasts (days) for:

  • Shipping planning

  • Short term sales …

• Medium term forecasts (weeks) for:

  • Resources dimensioning

  • Supply

• Long term forecasts (months) for:

  • Budget establishment

  • Resources dimensioning

  • Trend identification

2. Subjective forecast methods

In certain sectors, ones use subjective forecast (established on opinions and not on mathematical evaluations). Among these methods, we have:

• Opinion of the concerned individuals (ex: salesmen)

• Polls among potential customers

• Domain experts opinion or Delphi method

3. Objective forecast methods

Mostly used by companies, integrated into the demand forecast and planning tools.

3.1. Temporal series

They use sales historical data and models the future behaviour based on the past..

They use the analysis of:

• The trend : growth or decrease stability in the time (linear or non-linear)

• The seasonality: an particular event occurring on a regular basis (weekly, monthly, annual …)

• Erratic characteristic: is said from a very irregular and un-forecasted event.

3.2. Stationary series

A series is stationary when each event can be represented by a constant plus a randow fluctuation

Two methods are mainly utilized:

• Mobile average

A level N mobile average is the arithmetical average of the last N observations. In forecast methods, this average becomes the next forecast.

• Exponential smoothing

A popular and classical forecast method, the current forecast is a weighted average of the last forecast and the actual demand.

See Simple-Exponential-Smoothing

3.3. Trend Analysis

Linear regression

It allows data analysis and trend calculation based on a first degree equation (equation of type Y = a.X+b)

See Linear regression

Holt's Double Exponential method

Holt’s method is based on a double smoothing to evaluate temporal series with trend factors.

3.4. Seasonal series

A seasonal series is a series which profile repeats itself every N periods during a given number of periods.

Seasonal factors

The best known method is to evaluate multiplication factors of each season and weight the trend in function of these factors.

Winters method

Winters method is a triple exponential smoothing.

B- Forecast reliability control

=Minimisation of the following values

Be “et” the error on forecast at time t, we define “et” as the difference between estimated value of forecast and the actual demand value.

et = Ft - Dt

Be e1, e2, e3, ... , en the observed error on n periods,

Hence:

The mean absolute deviation (MAD) will be:

And the mean square error (MSE) will be:

Finally, the mean absolute in percentage will be: