An unhelpful web interface for a forecasting model

Sometimes national companies get their PR wrong.  Tina and I have gas heating, which means that we have a gas meter by the back door of our house.  Once every quarter, we are sent a gas bill - and on that basis the company calculates how much we pay monthly.   Generally we are in credit, but that means that our household budget can be worked out more easily. 
When we were sent our latest bill, I was surprised to see that the meter had not been read for a year - and we had the option to send our own reading to ensure more accurate bills.  So I opened the meter cabinet and took a reading.  And it was very close to the estimate, only 2 units below the estimate; however, their estimate was datestamped one week earlier.  So, I thought, as it is about 52 weeks since the last reading, their forecast for our consumption over 52 weeks is our actual for 53 weeks ... an error of about 2% - pretty good.
However, their computer didn't agree.  When I entered the accurate reading on the website, there was an immediate response to the effect that this was outside the tolerance for their system.  It made me wonder what tolerance there was, as I consider 2% pretty good.  Tina suggested that the system rejected the reading because it was lower than their estimated reading and might correspond to the meter having gone "round the clock". 
I assume that somebody (O.R. or statistics) had a forecasting model working behind the website, and that it was programmed with rules for accepting readings by house-owners, but somebody seems to have made the model too strict, and the interface too severe.
Just another example of how an O.R. model can have faults which show up if there has been insufficient thought about the way that the model is used.


  1. Might they have programmed an absolute (rather than relative) tolerance, based on a (faulty) assumption of more frequent readings, and thus smaller increments?


Post a Comment

Popular Posts