Elements of Forecasting
Introduction
Dr. Sung Won Kim
Quotes about Forecasting
"Those who have knowledge, don't predict. Those
who predict, don't have knowledge.“
—Lao Tzu, 6th Century BC Chinese Poet
[Link]
Quotes about Forecasting (cont.)
“Forecasts may tell you a great deal about the
forecaster; they tell you nothing about the future.”
—Warren Buffett
[Link]
Quotes about Forecasting (cont.)
“We have two classes of forecasters: Those who
don’t know—and those who don’t know they don’t
know.”
—Economist John Kenneth Galbraith
[Link]
Subjective Forecasting Methods
Subjective and judgmental methods are often
used when there is a limited amount of data
available.
Subjective methods may even perform better than
quantitative methods.
Very long-term predictions (e.g., 50 years out)
Quantitative models do not include the necessary types
of data.
Subjective Forecasting Methods (cont.)
Guess
Sales-force composite – ask people closest to the
problem to make forecasts. Ask for best, normal, and
worst cases
Jury of executive opinion – experts deliberate as a
group
Subjective Forecasting Methods (cont.)
Delphi Method
1. Participants are selected.
2. Participants fill out questionnaire.
3. Questionnaire results are summarized.
4. Participants review and consider results to update
their forecasts.
5. Repeat until a consensus is achieved.
Disadvantages of Subjective
Forecasting
They are almost always biased.
They are not consistently accurate over time.
It takes years of experience for someone to
convert intuitive judgement into good forecasts.
Chase,Charles W.,,Jr. (1991). Forecasting consumer products. The Journal of Business Forecasting Methods & Systems, 10(1)
What Is Business Forecasting?
Forecasting is the process of predicting future
business events.
Forecasting relies on historical data and
forecasting techniques.
Forecasting requires managerial oversight to
ensure that the correct forecasting techniques are
implemented to prevent costly errors.
What Is Time Series Data?
Time series data is a collection of numerical data
collected at regular intervals.
Example SPY Closing Prices
Date Close
5/21/2019 286.51
5/22/2019 285.63
5/23/2019 282.14
5/24/2019 282.78
5/28/2019 280.15
Yahoo Finance
R: Create Time Series
See 10 EOF intro.R
Annual Coal Production: USA 1920–1968
(in tons)
References
Hanke, J. E., Reitsch, A. G., & Wichern, D. W.
(2001). Business forecasting (9th ed.). Upper
Saddle River, NJ: Prentice Hall.
Wilson, J. H., & Keating, B. (1994). Business
forecasting. Tata McGraw-Hill Education.
Financial Analytics
Introduction
Dr. Sung Won Kim
What Is Analytics?
Scientific process of uncovering and showcasing
patterns in data
Converts raw data into useful knowledge
Blend of statistics, computer programming, and
mathematical modeling
What Is Financial Analytics?
Example cases of financial analytics:
Use business operations data to assess a firm’s
performance and recommend improvements.
Evaluate an investment balancing risk and return for a
portfolio.
Assess future risk in property values, policy, and
claims. (Risk Management Information Systems)
Financial Analytics: Overview
[Link]
services-06195123
Financial Analytics: Automate
Algorithmic trading for fast-paced automated
trading
Customer credit risk evaluation for credit decisions
Customer complaint management for root-cause
analysis and rapid response
[Link]
financial-services-06195123
Financial Analytics: Informate
Warning predictions based in liabilities
Predicting loan delinquency risk of a customer
Forecasting churn risk for individual customers
Detecting financial crime like fraud, money
laundering, etc.
[Link]
financial-services-06195123
Financial Analytics: Transform
• Growth
• Profitability
• Liquidity
• Cash Flow
• Valuation
• Leverage
• Portfolio Management
[Link]
Financial Analytics Topics
Time Series Analysis – understanding
autocorrelated data
Portfolio Management involves evaluating a group
of investment’s risk and return trade-off.
Elements of Forecasting
Performance Measurements
Dr. Sung Won Kim
Forecast vs. Actual
Actual value (𝑦" ), also called observed value, for a
particular variable is obtained by observing the available
data.
Forecasted value (𝑓" ), also called predicted value, for a
particular variable.
Difference/Distance between actual and forecast values is
called residual or error
Distance Measure
0 𝑖𝑓 𝑦" = 𝑓"
Similarity/Positional distance = %
1 𝑖𝑓 𝑦" ≠ 𝑓"
Absolute distance = ∑-"+, 𝑦" − 𝑓"
Euclidean Distance = ∑-"+,(𝑦" − 𝑓" )1
Performance Measurement
We will discuss some commonly used performance measures.
In the following definitions:
𝑦" represents the actual value at time t.
𝑓" represents the forecasted value (estimated value) at time t.
𝑒" = 𝑦" − 𝑓" represents the forecast error (residual) at time t.
𝑛 is the size of the test set.
Performance Measurement
The mean forecast error (MFE)
, -
𝑀𝐹𝐸 = ∑ 𝑒, average of all forecast errors, which indicates
- "+, "
forecast bias by showcasing direction of error
If MFE = 0 (desirable goal), then forecasts are accurate with
minimum bias.
Affected by scale of measurement and data transformation.
It does not penalize extreme error values.
Performance Measurement
The mean absolute value (MAE)
, -
𝑀𝐴𝐸 = ∑"+, |𝑒" |,
average of all forecast errors (in absolute
-
terms), which indicates magnitude of overall error
Small values of MAE are desirable.
Affected by scale of measurement and data transformation.
It does not penalize extreme error values.
Performance Measurement
The mean percentage error (MPE)
, - :
𝑀𝑃𝐸 = ∑"+, ; ×100, average error percentage indicating
- <;
direction of error.
MPE close to zero is desirable.
Performance Measurement
The mean absolute percentage error (MAPE)
, :;
𝑀𝐴𝑃𝐸 = ∑-"+, ×100, is a percentage of average absolute error
- <;
without any information about error direction.
Affected by data transformation, but not scale of measurement
It does not penalize extreme error values.
Performance Measurement
The mean squared error (MSE)
, -
𝑀𝑆𝐸 = ∑"+, 𝑒"1,gives overall idea of the forecast errors even if there are
-
positive & negative error values cancelling each other.
No information about error direction
Affected by scale of measurement and data transformation
It does penalize extreme error values.
Performance Measurement
The sum of squared error (SSE)
𝑆𝑆𝐸 = ∑-"+, 𝑒"1, total of squared forecast errors.
Properties are same as MSE.
Performance Measurement
The signed mean squared error (SMSE)
, - :;
𝑆𝑀𝑆𝐸 = ∑
- "+, :;
𝑒"1, same as MSE incorporating error direction.
Affected by scale of measurement and data transformation
It does penalize extreme errors.
Performance Measurement
The Theil’s U-statistics
1 -
∑"+, 𝑒"1
𝑛
𝑈=
1 - 1 -
∑"+, 𝑓"1 ∑"+, 𝑦"1
𝑛 𝑛
Ranges 0 ≤ U ≤ 1
U = 0 signifies a perfect fit, which is desirable.
Affected by scale of measurement and data transformation
Performance Measurement
The root mean squared error (RMSE)
-
1
𝑅𝑀𝑆𝐸 = 𝑀𝑆𝐸 = A 𝑒"1
𝑛
"+,
Shares same MSE properties
Performance Measurement
The normalized mean squared error (NMSE)
CDE ,
𝑁𝑀𝑆𝐸 = FG
= FG - ∑-"+, 𝑒"1, it normalizes MSE dividing it by the test
variance
A balanced error measure, which signals forecast accuracy.
Smaller NMSE is desirable.