Class 05: Forecasting – Advanced Methods

 Chapter 9

Copyright © 2016 Pearson Education, Inc. 9-1

 Chapter Objectives
Be able to:
• Be able to calculate the following forecasts:
• Exponential Smoothing
• Basic Linear Forecast (simplified version)
• Understand the principles used within Causal Forecasting Models
• Linear Regression (Principles, not calculations)
• Multiple Linear Regression (Principles, not calculations)
• Understand the difference between Causality and Correlation
9-2
 Quantitative Forecasting Approaches
Time Series Models: demand follows a trend and/or pattern over time.
• Last Period or Naïve Forecast
• Moving Average
• Weighted Moving Average
• Exponential Smoothing
• Adjusted Exponential Smoothing*
• Linear Regression*
Causal Models: demand is predicted by observing environmental factors
such as economic indicators
• Linear Regression*
• Multiple Regression*
4-3
 Exponential Smoothing Model
A form of moving average model in which the forecast for the next period is calculated as
the weighted average of the current period’s actual value and forecast.

Ft+1 = αDt + (1- α)Ft

Allows one to adjust the balance between forecasted values and actual values.
The closer a is to 1, the greater the weight put on the most recent actual demand value.

9-4
 Exponential Smoothing
Dt = 110 Ft+1 = αDt + (1- α)Ft
Error = Dt – Ft = 10
If α = .1, Ft+1 = .1(110) + .9(100) = 101
Ft =100
If α = .3, Ft+1 = .3(110) + .7(100) = 103
If α = 1.0, Ft+1 = 1.(110) + 0(100) = 110

t t+1 t+2
time
• Because Exponential Smoothing calculations involve utilizing the previous forecast, an
initial forecast (using some other method) will always be given.
• Because Exponential Smoothing calculations involve utilizing the previous forecast,
which is based upon the previous forecast, which is based upon the forecast before that ….
A Forecast based upon Exponential Smoothing is somewhat influenced by ALL previous
forecasts.
 Exponential Smoothing Model
 Features
 Weighted smoothing model with greater weight on most recent data.
 Requires very little stored data (only the past period forecast and demand) and easy to automate.
 The general rules for determining the a value:
 Stability
• The greater the randomness in the time series data, the lower the a value (the more weight that is
placed upon the forecast values.)
• The less randomness in the time series data, the higher the a value (the more weight that is placed
upon the most recent demand value.)
 Responsiveness
• The greater the instability in the underlying data pattern, the higher the a value.
• The less instability in the underlying demand pattern, the lower the a value.
 a is selected to minimize forecast error.
 Smoothing constant is typically 0.05 ≤ α ≤ 0.3 9-6
 Exponential Smoothing Example
You are organizing a Cougar Alumni Meeting for 2019 and have the past
attendance since 2014. What would have been the forecast attendance for
2015-2019 using exponential smoothing ( = .10) if the 2014 forecast was
175.

2014
2015
2016
2017
2018
2019
 Exponential Smoothing when Data has a Trend or Seasonality
5000 7000

4500
6000
4000

3500 5000

3000
4000
2500
3000
2000

1500 2000
1000
1000
500

0 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29

Alpha = 0.1 Alpha = 0.5 Alpha = 0.8 Alpha = 0.1 Alpha = 0.5 Alpha = 0.8

Which α value is most responsive?

Exponential Smoothing is not well suited to forecast trend or seasonality

Exponential Smoothing forecasts will tend to lag trending and seasonal data.
 Alternative Methods to Address
Trending or Seasonality
 Adjusted Exponential Smoothing
 Expanded Exponential Smoothing Calculation considers Trending
 Contains an “Unadjusted Forecast” component (a)
 Contains a “Trend Adjustment” component (b)
 Linear Regression
 Attempts to model underlying trends using a straight line
 Form of the line is Forecast = constant + slope (time)
 Seasonally Adjusted Linear Regression
 Multiplies Linear Regression forecasts by a seasonal index

9-9
 Linear “Regression” Forecast
• Persistent, up or down, generally linear movement in a time series
• Trends will change over time. 5000

• Increase, decrease, remain constant 4500

4000 f(x) = 119.38094957077 x + 748.331245796288

Demand Forecast = a + bx, where 3500

• a = “intercept” at period 0
3000
• b = demand increase per period
2500
• x = desired forecast period
2000

What was the forecast at period 16? 1500

Forecast = 1000

500
What is the predicted forecast at period 37?
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29
Forecast = 4-10
 Causal Forecasting Models
Also known as correlation models
 Quantitative Forecasting method that predicts a dependent variable as a
mathematical function of something other than time (the independent
variable)
 Linear Regression: Establishes a mathematical relation between one dependent variable
and one independent variable
 Multiple Regression: Establishes a mathematical relation between one dependent
variable and multiple independent variables
 Dependent Variable: The variable that is assumed to be “caused”
 Independent Variable: The variable(s) that is(are) assumed to do the “causing”
 Examples:
 The money spent on drought relief is mathematically related to rainfall levels
 The amount of food consumed at a party is mathematically related to the number of
people at the party, when they last ate, how many food stations are available, the taste
9-11 of
 Causal Forecasting Models
WARNING: Just because a Linear Regression or Multiple Regression is
calculated does NOT prove that one variable CAUSES another one to change.

9-12
 Reminders - Assignments
 As of Class 05, you may now sign up for an Exam Time/Date Reservation
with CASA. www.casa.uh.edu

 Due to be completed on MyOMLab prior to MIDNIGHT on SUNDAY

 GLM Chapter 9

Copyright © 2016 Pearson Education, Inc. 9-13