Documente Academic
Documente Profesional
Documente Cultură
Demand Planning
Univariate Forecasting
– Moving Average
– Simple Linear Regression
– Exponential Smoothing
– Holt-Winters
– Croston’s Model (for sporadic demand)
Causal Analysis
– Multiple Linear Regression
Composite Forecasting
– Weighted Averaging of Multiple Models
Univariate Profile in APO
Entries
Historical input
Version,
Version and key figure
Key figure
Model parameters
Forecast strategy,
Strategy, seasonal length,
Periods per season,
smoothing parameters
Parameters: a, b, g, s
Control parameters
Without Leading Zeros,
Outlier correction, adjustment Outliers,
of corrected history,
workdays correction Days in Period
History
500
400
Units
300
200
100
0
Y-4 Y-3 Y-2 Y-1 Y+1
Time Now
Determine Base Level & Trend
History Trend
500
400
Units
200
100
0
Y-4 Y-3 Y-2 Y-1 Y+1
Now
Model Seasonality
400
Units
300
200
100
0
Y-4 Y-3 Y-2 Y-1 Y+1
Now
Forecast Future Sales / Demand
400
Units
300
200
100
0
Y-4 Y-3 Y-2 Y-1 Y+1
Time Now
Elements Of Univariate Forecast in APO
Past Future
Data smoothing – Significance of a
The most common values for alpha lie, between 0.1 and 0.5.
Example – Constant using alpha 01 Vs 0.5
Weighting of historical periods
Constant & Trend using alpha & beta 01. Vs 0.5
Seasonal Method – Weighting factor Gamma
Smaller the sigma factor, Less the tolerance, greater the control
values
Outlier Calculation logic
Moving Average –
– Assumption : Data varies 12
Series 3 point MA 5 point MA
Volume
Weighted Moving Average –
– Assumption : Data varies 10
around a constant value and
shows no seasonality and
trend
9
Jan Apr Jul Oct Jan Apr Jul Oct Jan Apr Jul Oct
– Every historical value is
History Future
weighted with a factor
Forecasting Models - Explanation
Volume
2
Manual Forecast –
– Assumption – Not interested in system to propose basic & trend values
– Define Basic, trend value yourself.
Croston Method
– Assumption : Demand Occurs every X period
Forecasting Models - Explanation
Auto Model I
– Assumption : Data has visible trend & seasonal pattern, 1st order
exponential smoothing required
– System checks intermittent pattern in historical data – If data is not there
for more than 66% of total period, Propose Croston Model.
– System checks seasonal effect by determining Auto co-relation
– System checks trend by trend significance test
– If any of these checks are not valid, results in no forecast
Auto Model II
– Assumption : Data has visible trend & seasonal pattern, 1st order / 2nd
order exponential smoothing required
– System checks intermittent pattern in historical data – If data is not there
for more than 66% of total period, Propose Croston Model.
– System checks white noise, if it finds, Propose Constant model.
– Checks for Constant / Trend / Seasonal / Seasonal trend / Seasonal
Linear Regression starting from alpha / beta / gamma equal to 0.1
Trend Method – Linear Regression
– Step 3 –
• Every historical value is divided
by the corresponding average
seasonal index
• This results in historical data
corrected by seasonality
– Step 4 –
• Linear regression is carried out
for the corrected historical data
• The linear regression forecast is
multiplied by the seasonal index
Error Total
– Used primarily as a measure of systematic error
– Bias: Measure of error that can create large cumulative error
From all the available methods for forecast errors most common is:
– Mean Absolute Percent Error (MAPE)
Reason -
– It’s easy to Understand
– It’s relatively easy to calculate and correlate with the business results
– All errors weighted equally
Forecast Bias
Definition-
– Variance of historical sales volume over the forecasted sales volume as a
positive or negative percentage.
– Positive is oversell or under-forecasted &
– Negative is undersell or over-forecasting
Develop Custom Reports for generating forecast bias, This is a good method
to keep track of forecast Vs. current sales.
– In APO there is no default calculation for forecast bias.
Questions and Answers
???