|
M3 Competition; Description of Data and AUTOBOX forecasts
Introduction
The Institute of International Forecasters (IIF) is conducting a third competition
which is designed to overcome the glaring deficiencies of the two prior competitions.
Prof. Makridakis , the M in the M3 title has collected Monthly (M) , Quarterly (Q)
, Annual (A) and Other (O) series ... 3003 in all. AFS includes here a brief description
and summary of these series and the call for participants to prepare forecasts.
AUTOBOX was used to analyze , automatically without any human intervention, all
How AUTOBOX was used in this competition
AUTOBOX was set up in a BATCH mode and conditions were set under which modelling
was to be performed. Each time series was analyzed under four sets of conditions
and the model selected was the one that minimized the error sums of squares. This
is equivalent to maximizing R SQUARED or the AIC. The four conditions were: 1. Perform
ARIMA modelling first then do INTERVENTION DETECTION, allowing local time trends
to be identified. 1. Perform ARIMA modelling first then do INTERVENTION DETECTION
, NOT allowing local time trends to be identified. 3. Perform ARIMA modelling second
after INTERVENTION DETECTION, allowing local time trends to be identified. 4. Perform
ARIMA modelling second after INTERVENTION DETECTION, NOT allowing local time trends
to be identified. The best of these four approaches was then declared the "winner"
and its forecasts saved for submission to IIF
How long did AUTOBOX take to perform the 3003 analysis ?
Since this analysis was done while multi-tasking, it was impossible to exactly
estimate the time required. We ran it on a Pentium 133 with 16mg ram. Print options
were set to a minimum as no details were required for this production run. The time
to develop a model varied as a result of the length of the series and the strength
of the model structure and of course the presence of anomalies or outliers, be
they Pulse, Level Shift, Seasonal Pulse or Time Trends. We would estimate that
if AUTOBOX had been run without any other tasks active the run time would have been
around 4 to 6 hours.
When will they get it right ?
The evaluation of a forecasting method , even with 3003 different time series still
fails to provide generality due to the design of the comptetion. The single largest
confusion in measuring and conducting forecasting competitions is the confusion
between forecast errors from a single origin and forecast errors for different lead
times. Single origin forecasts generate a correlated set of forecasts due to the
inherent bootstrapping procedures. That is to say the forecast error for one period
out is correlated with the forecast error for two periods out, etc. To correctly
measure forecast errors one has to compute k period projections from n origins.
In this way, one gets n independent measures of one period out erors, two period
out errors, etc. This requires an iterative process where the modeller is given
a set and asked for a k period forecasts and is then given 1 new value and is asked
to return another set of k period forecasts. In this way, the effect of the origin
or launch is designed out by virtue of the n replications. The developers of the
M3 competition could have done this by scaling and coding these series thus masking
the data and defeating any attempt to "cheat". A more important point
is the continued silliness inherent in auto-projective models, i.e univariate models.
The history of a series never causes or is responsible for the future. It is simply
a surrogate for the omitted "cause" series. Box and Jenkins not only codified
"rear-window driving" models (ARIMA) but developed a rigorous approach
to causal modelling known as Transfer Functions. Transfer Functions are simply distributed
lag models which are optimally tuned to the data. By extracting the impacts or elasticities
associated with causals or exogenous series one can project using the drivers rather
than the rear-view mirror.
Until both of these issues are spoken to the question of which approach
or model is optimal will remain unanswered.
M3-International Journal of Forecasting Competition
The M3-IJF Competition will compare the forecasts produced by a
variety of extraoplative forecasting methods. Participants from multiple disciplines
will use techniques in which they have expertise to forecast 3003 time series. Resulting
forecasts will be compared with actuals using multiple error measures.
The Series
The 3003 time series are distributed as follows
Types of Time Series Data
|
Time Interval
|
Micro
|
Industry
|
Macro
|
Finance
|
Demographic
|
Other
|
Total
|
Yearly
|
146
|
102
|
83
|
58
|
245
|
11
|
645
|
Quarterly
|
204
|
83
|
336
|
76
|
57
|
0
|
756
|
Monthly
|
474
|
334
|
312
|
145
|
111
|
52
|
1428
|
Other
|
4
|
0
|
0
|
29
|
0
|
141
|
174
|
Total
|
828
|
519
|
731
|
308
|
413
|
204
|
3003
|
For each time series the following will be available
- title
- starting date
- type of series (macro, micro, demographic, financial, others)
- time interval (yearly, quarterly, monthly, others)
- historical values
For those who want more details and to download the original distribution go
to M3-IJF Competition Data.
Click here to return to the home page
P.O. Box 563
Hatboro, PA 19040
Tel: (215) 675-0652
Fax: (215) 672-2534
sales@autobox.com
© Copyright 1996 AFS Inc.
All rights reserved.
|