Hi all, I have figures for 2006 and 2007 from a data set (split into months), i also have the 2 first entries for 2008 (January and February), I need a projection for the yearly cumulative of 2008. What technique can/should i use to obtain an unbias figure?
It can't be done. You have WAY too few data points. Unfortunately, the corporate world forces many to make such projections and explain under duress any deviation. It's all a little silly.
The best predictor would be from your imagination, based on an examination of the data. Do they show any trend? Any seasonality? What sort of data are they? How were they collected? How were they split into months? Does anything account for months of different lengths? What level of confidence will be sufficient for your audience? What sort of business cycle realtes to these data? Are there any exogenous factors that relate to these data?
There are many questions to ask and answer before proceeding. You must build a decent model.
Thanks for your reply, yeah it does seem a little silly. It's for my job, i wish i could actually post the data, but i think i'd be risking my job if i did, oh well.
Really, in a practical setting, where you need to produce something, the idea that management cannot have what it wants normally is perceived as unsatisfactory.
Build a model and explain your assumptions. In general, lacking sufficient information, the default assumption is level or straight. You can display data with these assumptions, carefully disclosing these assumptions, and in this way force management to deal with the implications of the lack of information. If they have better assumptions, act all excited to get the additional information and make a new report showing those implications.
If it were me, I might have an inclination to look into some generalist modelling software so that assumptions can be coded quickly and results obtained very nicely. One such package is @RISK from Palisade Corporation. You can define any parameter with any distribution and let @RISK run simulations and tell you what the resultant distributions might look like. It's pretty cool stuff. If you've the budget for it, you can play with it for years and never get bored. "WinBugs" is another one that has some similar functionality, but considerably different in design and emphasis.
No need to give up just yet!