Life is about optimizing paths (own's and possibly other's) in an environment divided by tradeoffs (if you think `homo oeconomicus' then you may stop reading as well). I here review the well-known Bias-Variance Dilemma as well as the less-well-known Forecast AT-Dilemma and Signal-Extraction ATS-Trilemma. Whereas the first mentioned is a generic (unspecialized) statistical tradeoff, the former two address specifically forecasting and real-time signalextraction (though their essence is generic, too). A strongly counter-intuitive example is provided in order to flash-light into the darkness of the ATS-trilemma.
I just stumbled across a recent econbrowser entry posted by Menzie Chinn: therein I saw a Euro-area GDPquarterly (q-o-q) forecast by the Bundesbank, dated Nov.25 (2011) which attracted my attention. In this (SEFBlog-) entry I'd like to compare the previous GDP-forecast with another November release, namely a GDP Euro-area monthly (y-o-y) nowcastas published in 1. I'll attempt to derive some potentially interesting conclusions from this comparison by addressing tradeoffs of both designs.
Tucker Mc Elroy and myself up-dated our article on multi-step ahead forecasting. Here's the latest working paper: MultiFit_7-27-11.pdf. We are currently working on a new paper which picks-up this stuff in order to address a "generalized prediction problem": this will involve `traditional' as well as `less conventional' forecasting problems, real-time signal extraction and customized optimization criteria. I guess you see the topic: we shift progressively from `one-step ahead mean-square' to end-up in `MDFA'. I hope you'll enjoy this (next) paper, scheduled for 2012.
I just received the outlook for the Eurozone computed by the EFN (European Forecast Network): EFN2010_autumn.pdf which put me in the `forecasting mode'. I guess I'm not totally wrong by assuming that most readers/users of the report are interested in `forecasts'. For my part, I'm more interested in the way forecasts are produced: Am I able to identify the model from the forecasts (reverse engineering)? And should we really forecast something as capricious as a sentiment?
Please have a look at my previous Blog-entry concerning the choice of Data-Set F. Note that I release my results slightly before the deadline of the competition: I don't trust its design and I wouldn't be surprised if the whole competition would get a `cancel'. Therefore, posting my results prematurely is nomore of immediate concern to me. You might be tempted to ask: "Why did you participate"?
I provided prematurely results, see 2. Therefore I may propose forecasts too.
Sometimes, it's difficult to anticipate an experience that becomes an evidence afterwards. At least I didn't expect so many flaws in NNGC1 a priori.
Presently, my main motivation has been slightly downgraded and it's scope has been shifted. More precisely:
I took the opportunity to discuss the design of more `meaningful' future forecasting competitions, see 1.
Here, I'd like to discuss forecasting issues that might be of interest for practitioners. For that purpose I'll rely on data-set F of NNGC1.
Please allow me to show up a bit of ambition at this place: I'd like to reveal/confirm what forecasting is all about, namely the systematic and careful design of abest possible compromise.
In my previous Blog entries 1 and 2 on `forecasting competitions' I criticized the design of this year's tournament (series/results are available on the net: I provide some of the results in 1) and I proposed directions for the design of future competitions (more user-focused and more practically relevant dynamic and fair designs allowing for human interaction/intervention).
In 1, I mentioned that NNGC1 relies on six data sets, A to F, ranging from yearly to hourly data, and that data sets A to D are `possibly' subject to fraude i.e. results could be obtained on the net.
Data set E is daily transportation data, namely car counts in Swiss tunnels. It might be difficult to obtain results for `non-initiated' but our institute works, among others, on (Swiss) traffic research projects and therefore I conjecture that I would have access to the data (although I did not verify formally).
However, I was unable to find results/data for the F-series (hourly US-airport and Paris-metro data) on the net. Therefore, I conjecture that NNGC1 restricted to data set F is `clean'.
I decided to compute forecasts for the hourly data set F (even if the whole tournament is at risk of being cancelled). Here is my motivation for this decision.