Transcript Diebold`s 6 Considerations

Econ 427 lecture 2 slides
Byron Gangnes
Lecture 2. Jan. 13, 2010
• Anyone need syllabus?
• See pdf EViews documentation on CDRom
• Problem set 1 will be available by Tues at
the latest.
Byron Gangnes
The forecasting problem
•
•
You’re given a forecasting assignment. What
things do you need to consider before deciding
how to develop your forecast?
Diebold’s 6 considerations for successful
forecasting
Byron Gangnes
The decision environment
•
How will the forecast be used? What will
constitute a “good” forecast?
– What are the implications of making
forecast errors?
•
•
•
How large are the costs of errors?
Are they symmetric?
An optimal forecast will be one that
minimizes expected losses.
Byron Gangnes
Loss functions
Error e  y  yφ
L(e)
Loss
What characteristics would you expect a loss
function to have?
Types of loss functions Lossfunction.xls
•
•
•
•
–
–
Absolute loss
Quadratic loss
•
–
•
Why is this one appealing/convenient?
Asymmetric loss functions
How do you decide which to use?
Byron Gangnes
Measures of Forecast Fit
• Making it concrete: some common
measures of forecast fit
– Notation: error of a forecast made at time t of
period t+h is:
et h,t  yt h  yt h,t
Byron Gangnes
Measures of Forecast Fit
– Mean absolute error MAE is
1 T
MAE   et  h,t
T t 1
– Mean squared error MSE is
1 T 2
MSE   et h,t
T t 1
• (see pp 260-262 in book)
– Look at my MAE/MSE forecast comparison
example MaeMseExample_Mine.xls
Byron Gangnes
Measures of Forecast Fit
– Do they give the same ranking? Need they
always?
– Would you want to use in-sample data for
this?
Byron Gangnes
The forecast object
• What kind of object are we trying to
forecast?
–
–
–
–
–
Event outcome
Event timing
*Time series
What are examples of each?
Other considerations: availability and quality
of data
Byron Gangnes
The forecast statement
• What sort of forecast of that object do we
want?
– Point forecast
– Interval forecast
– Density forecast
 
Byron Gangnes
The forecast horizon
• How far into the future do we need to
predict?
– The “h-step-ahead forecast”
• also, h-step-ahead extrapolative forecasts
– Likely dependence of optimal forecasting
model on fcst horizon
Byron Gangnes
The information set.
• What do we know that can inform the
forecast?


multivariate
T
univariate
T

 yT , yT 1 ,..., y1


 yT , xT , yT 1 , xT 1 ,..., y1 , x1
Byron Gangnes

Optimal model complexity
• The parsimony principle
– more accurate param ests, easier interp, easier
to commun intuition, avoids data mining
• The shrinkage principle
– imposing restriction—sometimes even if
wrong!—can improve forecast performance
• The KISS principle
– Keep it sophisticatedly simple
Byron Gangnes
Next time…
• Read Chapter 2 carefully before class.
Byron Gangnes