Tuesday, December 04, 2007

Before it's too late!

Oldham currently entertain Luton Town at Boundary Park, the game will enter its second half very soon indeed. The current score is 1-1, with Oldham taking the lead, only to relinquish it within two minutes. The attendance has not yet been announced, but no doubt will be being determined right now within the deeper annals of Boundary Park. My attendance forecasting model suggests it will be 4,023. The attendance for the convincing performance, if unconvincing scoreline (only 1-0) on Saturday against Crewe was 3,900. More people would be expected to turn up for a league match, but it is Luton on a Tuesday. I'm not convinced it'll be over 4,000 but I could well be wrong.

Let's wait and see. And in the meantime, come on Oldham!

Sunday, December 02, 2007

Too high

The model, selected by Autometrics given the initial dataset that I specified, predicted an attendance as high as 4,587 for yesterday's Oldham Athletic v Crewe football match. The actual attendance was as low as 3,900, reflecting the somewhat odd appeal of the FA Cup. Only the League Cup visit of lower-league Mansfield Town in August attracted a lower attendance, of 3,155. League gates have edged below the 5,000 marker recently, as perhaps casual fans react to the fact that Oldham before yesterday hadn't won a home match since the previous (league) visit of Crewe, back at the end of September. Oldham won the match 1-0, and must be hoping for an enticing Cup draw this afternoon, when the big Premiership teams enter the fray, in order to get the casual fans back to Boundary Park.

But what is this Autometrics I speak of? It's not some management consultancy, but a model selection package, which builds upon many of the advances in PcGets, which was a development of David Hendry and Hans-Martin Krolzig. By selecting between variables and not models, it simplifies massively the unfeasibly difficult problem of selecting a model that besets all who do econometrics. Many research papers since 1999 have outlined the properties of this procedure. Theoretical size and power calculations are provided for the procedure, revealing that it retains irrelevant variables quite infrequently, and retains relevant variables almost always.

As yet, the procedure has been analysed and thought about as a model selection tool per se; using it for forecasting has not, as yet, been seriously analysed. Ideally this is what it could be used for, and I hope my research can delve further into this. For now though, it is my intention to make forecasts using Autometrics, and hopefully from time to time comment on possible strategies for using such a tool, based on simulation and theory. Naturally I also intend to report forecasts from using other methods, as another competing method for forecasting is to average over different models.