Week 333

Forecast Range
Very long range forecasts do not work well!!
• Forecasts are done under the assumption that the market and other conditions in future are very
much like the present
neerajshin@gmail.com
• Not that there
LZSQFTU6A9 will be no change in the market
• But the change is gradual, not a drastic change
• A financial crash like 2008 US market will send all forecasts into a tizzy
• Events like Demonetization would throw the forecasts into disarray
Based on the amount of data availability, one should not try to forecast more than a few periods
ahead
This file is meant for personal use by neerajshin@gmail.com only.

Proprietary content. ©Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited.
Sharing or publishing the contents in part or full is liable for legal action.
Scope of Module
• Only univariate TS is explored

• No attempt made to forecast
LZSQFTU6A9
more than one interdependent
TS
• The variable measured in TS is
assumed to be continuous and
have fairly decent volume
• Intermittent demand TS is not
consideredThis file is meant for personal use by neerajshin@gmail.com only.
Gathering Information
Historical data required for future prediction
If volume of data is limited, forecasts will not

be reliable enough
LZSQFTU6A9
If data is available for very long past, data may

not be useful at all

Example: Clay Brick Production
LZSQFTU6A9

Example: Clay Brick Production
Series not stable

LZSQFTU6A9
Use stable part for forecast

Forecast Range
Different industry needs forecast for different range for different purpose
with different accuracy
Example: Airlines industry: Interested in passenger volume forecast

• Long-term forecast: 5-10 years
neerajshin@gmail.com ̶ Strategic decision making
LZSQFTU6A9
• Mid-term forecast: 2-5 years

̶ Manpower hiring / Route alteration
• Short-term forecast: 2 weeks – 6 months

̶ Pilot / Cabin Crew rostering
̶ Dynamic pricing

LZSQFTU6A9
Forecast

Simple Forecast
 Naïve Forecast: Use the last observed value
 Average Forecast:
 Moving Average Forecast:
neerajshin@gmail.com • Take average over a window of certain width
LZSQFTU6A9
• Move the window
None of these work well even for the most regular

series

Moving Average
• One of the most popular and often-
used technical indicators to understand stock
movement
LZSQFTU6A9
• The moving average is easy to calculate and,

once plotted on a chart, is a powerful visual
trend-spotting tool.

Long Term Movement
LZSQFTU6A9

EXPONENTIAL SMOOTHING
LZSQFTU6A9
FORECAST

Model Validation
Predictive power of a model is estimated by comparing its
forecasting performance on a Test Data
A part of sample data is used to train (develop) the model:

LZSQFTU6A9 Training Data
A part of sample is withheld from estimation process: Test
Data
The model which gives smallest measure of error between

forecast and actual series is the ‘best’

Model Validation
• Training Data is used to identify a few working models
• The forecasts for training data are called fitted values
• Each of the models is tested against the observed values of the series
for hold-out period
• The model is selected to be the best where observed and forecasted
LZSQFTU6A9
values are the closest

Model Validation
• For other predicting models hold-out sample is randomly
chosen from the total sample
• Usually in 80:20 ratio
• For time series hold-out sample has to be the most recent
LZSQFTU6A9 period because of the ordered nature of the data
Training Data Test Data

Model Validation
• If there is seasonality in the data, the at least one
full season’s data need to be held out for validation
• Once a model is selected, held-out period must be
neerajshin@gmail.com incorporated into the sample for final forecasts for
LZSQFTU6A9
future

Measures of Forecast
Accuracy
• Performance of forecast method is tested by comparing the
forecast values with the test sample observations
෣ Forecast Values
𝑌(𝑡)
LZSQFTU6A9
Y(t)Observed Values
Compare the observed and forecast values through
various methods

Measures of Forecast Accuracy
 Residual sum of squares
෣
RSS = σ 𝑌 𝑡 − 𝑌(𝑡) 2
 Mean sum of squares

1
෣
MSS = σ 𝑌 𝑡 − 𝑌(𝑡) 2
𝑇
LZSQFTU6A9
 Mean absolute deviation
1 ෣
MAD =𝑇 σ |𝑌 𝑡 − 𝑌 𝑡 |
 Mean absolute percent error

1
MAPE = 𝑇 σ[ |𝑌 𝑡 − 𝑌෣
𝑡 | / 𝑌 𝑡 ] x 100

Forecast by Average
• Using mean of all past observations – easiest,

intuitive, naïve
LZSQFTU6A9
• Naturally does not work well!!
• Can this scheme be modified to get useful
forecasts?

Exponential Smoothing
• Weighted averages of past observations
• Weights decaying as observations get older
• Practically speaking, only the recent observations
LZSQFTU6A9 matter
• One or more parameters control how fast the
weights decay
• These parameters have values between 0 and 1

Decreasing weight given 0  1

to older observations

 (1   )
LZSQFTU6A9
 (1   ) 2
 (1   ) 3

Simple Exponential (SES)
If the time series neither has a pronounced trend nor
seasonality: Almost non-available!
LZSQFTU6A9

Petrol Consumption
LZSQFTU6A9

• Performance of the smoothing parameter α controls
performance of the method
• If α is closer to 1, forecasts follow the actual observations
more closely
LZSQFTU6A9 • If α is closer to 0, forecasts are farther from the actual
observations and the line is smooth

Double Exponential
• Applicable when data has Trend but no seasonality

• An extension of SES
neerajshin@gmail.com • Two separate components are considered: Level and Trend
LZSQFTU6A9
• Level is the local mean
• One smoothing parameter α corresponds to the level series
• A second smoothing parameter β corresponds to the trend
series
• Also known as Holt model

Caselet V: US GDP
LZSQFTU6A9
Trend Series

Caselet V: US GDP
α=0.99
α=0.99 β=0.05
LZSQFTU6A9
MAPE=5% MAPE=4.5%

Petrol Consumption
LZSQFTU6A9

Petrol Consumption
LZSQFTU6A9

with Seasonality: Holt-
Winters’ Model
• Because Seasonality can be additive or multiplicative, HW
model can be additive or multiplicative
LZSQFTU6A9
• Simultaneously smooths the level, trend and seasonality
• Three separate smoothing parameters
• α: Smooths level; 0<α<1

• β: Smooths trend; 0<β<1
• γ : Smooths seasonality; 0<γ<1

Champagne Sales
Model Estimates
Smoothing parameters:
alpha = 0.0842
beta = 1e-04
LZSQFTU6A9
gamma = 0.7608
• Smoothing parameter for trend (β) almost 0 corroborates well with insignificant
YOY movement
• Smoothing parameter for seasonality(γ) fairly high that almost all fluctuations
are due to seasonality

Champagne Sales
High sales volume forecasts

neerajshin@gmail.com Are close but not low sales MAPE=16.7%
LZSQFTU6A9
Volume months

Champagne Sales
High sales volume forecasts

neerajshin@gmail.com as well as low sales volume MAPE=10.6%
LZSQFTU6A9
months

Passenger Volume
Significant trend
and
seasonality both
LZSQFTU6A9

Passenger Volume
Model Estimates
Smoothing parameters:
LZSQFTU6A9
alpha = 0.3515
beta = 0.0147
gamma = 0.6481
• Smoothing parameter for trend (β) small compared to
other parameters
• Indicates almost a straight line trend
• Smoothing parameter for seasonality(γ) fairly high
Passenger Volume
Forecast values show

small bias MAPE=7%
LZSQFTU6A9

LZSQFTU6A9
HAPPY LEARNING
This file is meant for personal use by neerajshin@gmail.com only. 36 prohibited.

Proprietary content. ©Great Learning. All Rights Reserved. Unauthorized use or distribution

Week 333

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Week 333

Uploaded by

Copyright:

Available Formats

Forecast Range

Very long range forecasts do not work well!!

This file is meant for personal use by neerajshin@gmail.com only.

• Only univariate TS is explored

If volume of data is limited, forecasts will not

If data is available for very long past, data may

This file is meant for personal use by neerajshin@gmail.com only.

This file is meant for personal use by neerajshin@gmail.com only.

Series not stable

Use stable part for forecast

This file is meant for personal use by neerajshin@gmail.com only.

Example: Airlines industry: Interested in passenger volume forecast

• Mid-term forecast: 2-5 years

• Short-term forecast: 2 weeks – 6 months

This file is meant for personal use by neerajshin@gmail.com only.

This file is meant for personal use by neerajshin@gmail.com only.

None of these work well even for the most regular

This file is meant for personal use by neerajshin@gmail.com only.

• The moving average is easy to calculate and,

This file is meant for personal use by neerajshin@gmail.com only.

This file is meant for personal use by neerajshin@gmail.com only.

This file is meant for personal use by neerajshin@gmail.com only.

A part of sample data is used to train (develop) the model:

The model which gives smallest measure of error between

This file is meant for personal use by neerajshin@gmail.com only.

This file is meant for personal use by neerajshin@gmail.com only.

Training Data Test Data

This file is meant for personal use by neerajshin@gmail.com only.

This file is meant for personal use by neerajshin@gmail.com only.

This file is meant for personal use by neerajshin@gmail.com only.

 Mean sum of squares

 Mean absolute percent error

This file is meant for personal use by neerajshin@gmail.com only.

• Using mean of all past observations – easiest,

This file is meant for personal use by neerajshin@gmail.com only.

This file is meant for personal use by neerajshin@gmail.com only.

Decreasing weight given 0  1

This file is meant for personal use by neerajshin@gmail.com only.

This file is meant for personal use by neerajshin@gmail.com only.

This file is meant for personal use by neerajshin@gmail.com only.

This file is meant for personal use by neerajshin@gmail.com only.

• Applicable when data has Trend but no seasonality

This file is meant for personal use by neerajshin@gmail.com only.

This file is meant for personal use by neerajshin@gmail.com only.

This file is meant for personal use by neerajshin@gmail.com only.

This file is meant for personal use by neerajshin@gmail.com only.

This file is meant for personal use by neerajshin@gmail.com only.

• Simultaneously smooths the level, trend and seasonality

• Three separate smoothing parameters

• α: Smooths level; 0<α<1

This file is meant for personal use by neerajshin@gmail.com only.

This file is meant for personal use by neerajshin@gmail.com only.

High sales volume forecasts

This file is meant for personal use by neerajshin@gmail.com only.

High sales volume forecasts

This file is meant for personal use by neerajshin@gmail.com only.

This file is meant for personal use by neerajshin@gmail.com only.

Forecast values show

This file is meant for personal use by neerajshin@gmail.com only.

This file is meant for personal use by neerajshin@gmail.com only. 36 prohibited.

You might also like