I hope you're still monitoring this long-expired question. If you
are, and need further information, post a comment here.
Regards,
AsiaTechnicals
Generally Accepted Practices
============================
When I was a utility-sector secutities analyst, I did occasionally
look at weather with respect to electricity sales. Most analysts I
have spoken to would simply make semi-qualitative comments, like "it
was an exceptionally hot summer". The more sophisticated amongst us
would regress demand vs temperature, but there are several tweaks you
may have to consider if you want to be more accurate:
Demand Trend
============
Electricity demand growth would interfere with your regression. First
you'll have to detrend the data. Then either remove seasonal
variation (and test time of year variations (e.g. unseasonally cold),
or preserve seasonality and explain non-weather seasonal factors.
Non-linear Relationship
=======================
The relationship is non-linear over the annual temperature variation,
so you will have to either develop your own model, or regress small
segments on the temperature scale.
For example, I covered South-East Asian utilities, so there wasn't
much heating-led demand. If you are looking at North America, there
may be some.
At higher temperatures, you will probably find a threshold above which
demand begins to increase.
At peak temperatures, you may see a plateau of demand, as close to
100% of cooling capacity is on-line. You may even see a small decline
above a certain level, as productivity drops.
Multiple Factors
================
Other factors play a role in determining demand. The primary one is
economic activity. Others are more subtle. For example, you may have
more luck testing a more subjective measure of temperature, involving
wind shear and precipitation, than outside air temperature.
Correlation
===========
You may find that the correlation is surprisingly weak, due to the
influence of the other explanatory factors. Also, the relationship
between temperature and demand is probably lagged somewhat.
Hence, a statistics-based normalisation may have a rather weak
explanatory power.
Conclusion
==========
Owing to the complex relationship between weather and demand, you will
have a hard time normalising the figures and an even harder time
convincing someone of your numbers.
Personally, I would approach a problem like this by through modelling.
I would program a relationship (as simple as possible, with reference
to historical observations over a long, but relevant timespan). I
would curve-fit my model to a sample of data and perform walk-forward
testing of my model on out-of-sample real-world data. I would smooth
the residuals, which would form the basis of my 'normalised' demand.
I would then hope to prove that my model is substantially more
accurate and robust than a standard regression-based normalisation.
Further Reading
===============
There's plenty written about traded weather derivatives and other
related topics available through a Google search under:
weather temperature "electricity demand" correlation |
