Вы находитесь на странице: 1из 56

What Now?

How to Proceed When


Automatic Forecasting Doesn’t Work

Presented by Business Forecast Systems, Inc.


Eric Stellwagen 68 Leonard Street
Vice President & Cofounder Belmont, MA 02478 USA
estellwagen@forecastpro.com (617) 484-5050
www.forecastpro.com
On-demand Webinars and Handouts

Today’s webinar along with a pdf version of the slide set will
be posted on our Website next week.
Previously presented Webinars are available for viewing on
www.forecastpro.com.
Participants will receive an email confirming availability
after the Webinar and slide set is posted.

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Eric Stellwagen
Vice President & Cofounder of
Business Forecast Systems, Inc.
Coauthor of Forecast Pro product line.
Over 27 years in forecasting.
Currently serving on the board of directors of
the International Institute of Forecasters and
on the practitioner advisory board of
Foresight: The International Journal of Applied
Forecasting.
© 2013 Business Forecast Systems, Inc. www.forecastpro.com
What We’ll Cover
Overview

The Role of Judgment

Automatic Forecasting

Exponential Smoothing

New Product Forecasting


Top-down Approaches

Multivariate Approaches

Summary
© 2013 Business Forecast Systems, Inc. www.forecastpro.com
Common Forecasting Methods
Judgmental Approaches Multivariate Methods:
Event-index Models
Simple Time Series Methods:
Moving Averages Dynamic Regression

Same as Last Year New Product Methods:


Percentage Growth Forecast by Analogy (looks-like)

Statistical Time Series Methods: Assumption-based Models


Exponential Smoothing Diffusion Models (e.g., Bass)
Box-Jenkins (ARIMA)
Croston’s Intermittent Demand
Model
© 2013 Business Forecast Systems, Inc. www.forecastpro.com
Evolution of Forecasting Process

Phase 1 Phase 2 Phase 3

Judgment & Automatic Time Customized


Spreadsheets Series Approaches Approaches

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


The Role of Judgment

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Judgmental Forecasting

Pros:
Does not require statistical expertise.
Allows forecaster to incorporate domain knowledge.
This knowledge can come from many sources including
experience with similar products, feedback from sales
staff, customer surveys, focus groups, etc.
Does not require historical data.

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Judgmental Forecasting

Cons:
Is subjective.
Can be biased by company politics, sales goals, etc.
Is difficult to monitor performance and fine tune
future forecasts.
Is not automatic—can be very time consuming.

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Judgmental Forecasting
Judgment often plays an important role in forecasting, particularly with new
products, short product-life-cycle products, rapidly changing environments
and instances where the forecaster’s domain knowledge is not captured in the
statistical forecasting model.

A strong recommendation is to add judgment in the form of an override to a


statistically generated base-line forecast. This practice provides the ability to
track the effectiveness of the judgmental override and introduces more
accountability into the process.

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Automatic Time Series Approaches

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


3 Months of Data

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


12 Months of Data

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


24 Months of Data

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


36 Months of Data

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


6 Years of Data

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Automatic Time Series Approaches
Pros:
Simple to understand and explain
Widely accepted and used
Often quite accurate
Adaptive
Easy to apply

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Automatic Time Series Approaches
Cons:
Requires adequate demand history
Assumes continuity between past and future
Does not capture response to noncalendar-based events
(e.g., promotions)
Does not capture response to explanatory variables
Implementations vary and some are poor
© 2013 Business Forecast Systems, Inc. www.forecastpro.com
Rejecting Automatic Models

When you disagree with the forecasts generated using an


automatic time series approach you should reject them.
Generally there are three ways to do this:
Judgmentally override the forecasted values.
Dictate that a different forecasting model be used.
Reconfigure the input data.

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Why is it Wrong?
Domain knowledge
• Your knowledge of the future leads you to reject the
forecast.
• The solution is most often to judgmentally adjust the
forecast.
• Might be done informally or as part of a structured
process (e.g., S&OP)

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Adding Judgment

Best practices:
Retain statistical forecast and adjusted forecast for comparison
Document reason for changes

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Why is it Wrong?
Chose wrong time series model
• Often a case of misidentifying trends and/or seasonal
patterns.

• Dictating an appropriate exponential smoothing


model is often a good solution.

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Exponential Smoothing

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Main Characteristics of Exponential Smoothing

A family of models.
Models three data components—level, trend, and seasonality.
Assumes that each component is changing in time.
Assumes that there is random variation (noise).
Uses weights to reflect the relative emphasis given to the recent vs.
the distant past.
Estimates final values of the components and uses them to construct
forecasts.

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Holt-Winters Exponential Smoothing
Additive Multiplicative
Nonseasonal
Seasonal Seasonal
(SIMPLE)
Constant
Level

(HOLT) (WINTERS)
Linear
Trend

Damped
Trend
(0.95)

Exponential
Trend
(1.05)

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Why is it Wrong?
Inadequate data
• Data too short (new product)

• Data too low level (not enough structure)

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


New Product Forecasting
and Top-down Approaches

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Types of New Products
There are different types of new products including:
Replacement Products
Product Line Extensions
New-to-Company
New-to-World

The type of new product you are trying to forecast will often dictate what
approaches should be considered.

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Replacement Products and Product-Line Extensions

When forecasting replacement products or product line extensions


you will often want to leverage the data that exist for the
predecessor products.
Approaches can include:
Judgment and Market Research
Item Supersession (i.e., mapping histories)
Top-down Forecasting

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Item Supersession
You create a “forecast history” for the new product using the demand
histories of predecessor products and the new product.

For a replacement product this may be as simple as merging the old


product’s history with the new product’s history.

More complex mapping may be necessary depending on the


circumstances.

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Item Supersession
Old Product New Product

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Item Supersession
Forecast History

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Top-down Forecasting
Cough Syrup

SKU 1 SKU 2 Etc.

Group-level data are higher volume, will often exhibit more structure and will
have a longer demand history than the product line extensions and
replacement products.

To generate a top-down forecast, you first forecast at the group level using the
aggregated history. Then, you forecast at the lower levels. Finally, you apply
proportionality factors to lower-level forecasts so that the forecasts sum to the
top-level forecast.

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


New Product Forecasting
When forecasting new-to-company or new-to-world products you
obviously don’t have internal data—external data may or may not
be available.
Approaches can include:
Judgment and Market Research
Forecasting by Analogy (looks-like)
Assumption-based Models
Market Share Forecasting
Diffusion Models
© 2013 Business Forecast Systems, Inc. www.forecastpro.com
Why is it Wrong?
Inadequate data
• Data too low level (not enough structure)
Simplify the hierarchy

Forecast top-down

Aggregate the time buckets (e.g., switch from weeks to months)

Use bucket conversions

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Bucket Conversion
It is not uncommon for a company that needs a daily
or weekly forecast to discover that the data can’t
support statistical modeling at these periodicities.

One solution is to forecast at an aggregated level and


allocate down to the lower level. This process is
sometimes referred to as “bucket conversion”.

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Bucket Conversion
Bucket conversion can take different forms:
S
Flat allocation
Proportional allocation based on historical profiles
Proportional allocation based on forecasts (bucket
synchronization)

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Why is it Wrong?
Time series model is not appropriate
when demand is driven by:
• Events that move around the calendar

• Explanatory variables that can change abruptly

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Multivariate Methods

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Common Events
Product promotions
Moveable holidays (e.g., Easter, Rosh Hashanah,
Ramadan)
Catastrophes (e.g., earthquakes, hurricanes, 9/11)
Labor strikes
Acquisitions
New legislation or regulations

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Forecasting Event-Driven Data
Judgmentally adjust history to remove impact.
Separate base demand from event-driven demand.
Use a time series extension model (e.g., event-
index model, ARIMA intervention model, etc.).
Use a multivariate model (e.g., regression).

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


What is an Event-Index Model?
An extension of exponential smoothing.
An index-based approach.
The model introduces an additional smoothing
weight and updating equation.
The model requires an event schedule.

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Example: Sales of Mouthwash

Demand for mouthwash is not seasonal and for this brand not trended.

Price promotions by both the manufacturer and competitors introduce significant peaks and valleys.

The timing of promotions is similar from year to year (but not exactly the same), and thus without
domain knowledge the data appear to be seasonal.

The introduction of EDLP with Wal-Mart changes the response to promotions.

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Why is it Wrong?
Time series model is not appropriate
when demand is driven by:
• Events that move around the calendar

• Explanatory variables that can change abruptly

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Dynamic Regression
Pros:
s
Allows for the introduction of explanatory variables.
Lends insight into relationships between variables.
Allows for “what if” scenarios.
Can exploit leading indicators.

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Dynamic Regression
Cons:
Is not automatic, requires considerable expertise and time.
Will produce poor forecasts of the dependent variable if
there are difficulties in forecasting the explanatory
variables.
Requires large amounts of data.

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Independent Variables
Internal Variables External Variables
Prices Weather

Promotion Economy

Competition

Demographics

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Example: Sales of Electricity

Temperature is an important driver.


© 2013 Business Forecast Systems, Inc. www.forecastpro.com
Example: Sales of Electricity

Time series models cannot capture adequate response to temperature.

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Example: Sales of Electricity

Dynamic regression models capture response to temperature and work well.

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Summary
Judgment is always important and is best applied as an adjustment to a
statistically generated base-line forecast.
Automatic time series methods work well when you have adequate data
and are generally superior to spreadsheet models.
When you reject an automatic time series forecast, your options include:
• Judgmentally overriding the forecasted values.
• Dictating that a different forecasting model be used.
• Reconfiguring the input data.
Forecast Training and Workshops
S
BFS offers forecasting webinars and product training
workshops.
On-site, and remote-based (via WebEx) classes are
available.
Learn more at www.forecastpro.com

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Forecast Pro

Examples from today’s Webinar used Forecast Pro.


To learn more about Forecast Pro:
Request a live WebEx demo for your team (submit
your request as a question right now)
Visit www.forecastpro.com
Call us at 617-484-5050

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Our Next Webinar

Tracking Accuracy: An Essential Step to Improve Your


Forecasting Process, April 11, 2013 1:30 p.m. EDT
Eric Stellwagen, Vice President of Business Forecast Systems
Highlights include how to apply best practices, the pros and
cons of different error measurements and how to spot poor
performance.
Visit www.forecastpro.com to sign up!

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Questions?

© 2013 Business Forecast Systems, Inc. www.forecastpro.com


Thank you for attending!

© 2013 Business Forecast Systems, Inc. www.forecastpro.com

Вам также может понравиться