(62
PONTIFICIAE ACADEMIAE SCIENTIARVM SCRIPTA VARIA - 4
mely the relative size of deviations between theoretical and
observed values. It goes without saying that this is a serious
limitation, and that the JANUS quotient therefore by no means
is a panacea in the testing of forecasting accuracy. Specific
reference is made to the importance of paying special attention
to turning points in the phenomena under analysis. To quote
a wellkown example from meteorology, Ref. 47, the use of
digital computers in short range forecasting was tried out on a
cyclone which on Thanksgiving Day 1950 swept the US con-
tinent in a wide and softly curved swing, and when coming to
‘he Atlantic made a sharp turn northwards through New Eng-
land. It was no difficulty to simulate and forecast the wide
swing on the computer, but for a considerable time all of the
trial models led the forecast path of the Thanksgiving Day
cyclone right out into the Atlantic, and it took a qualified com-
bination of meteorological thinking and data compilation to
construct a model that reproduced the sharp turn.
3-3. Overfitting.
In the nonsense department of statistical method everybody
has seen the pitfall of overfitting — the situation when a model
gives illusively close fit to the given data because the available
observations are outnumbered by the parameters. A case in
point that is actually on record is the time series analysis of
a sea level, in which study 122 annual data were graduated by
a sum of 40 sinusoids with different periods, phases and
amplitudes. The resulting fit in the observation range was
very very close, and the forecast for the next year was included
in the report. Just as the report was published the next obser-
vation emerged, showing an ample deviation from the forecast.
The comment of the author was that unfortunately he had
forgot to include the 41st component.
The parameter estimation of ID- and BEID-systems by
‘21
Wold - pag. 48