For those of you interested in forecasting by “novel” methods (including particle swarm optimization), here is a competition to watch or indeed participate in .. I doubt I will have the time to do anything serious, but I have registered to get the data and I will have a look at it.

It is the 2008 Time Series Forecasting Competition for Computational Intelligence

The data is interesting .. daily cash withdrawals over two years from 111 ATM machines across England. It contains true zeros, and missing data (although I am unclear as to whether this is missing at random, or injected missingness, or how long the runs of missingness are), and by inspection it is very spiky. Quite why it should be so spiky I do not know .. some payday and local effects I suspect .. and I do not now what the units of observation are (pounds, thousands of pounds?). Perhaps one should treat this as discrete count data and fit Poisson hidden Markov models?

An idea of the thrust of the competition can be gained from a listing of the “acceptable” methods

* Feed forward Neural Networks (MLP etc.)
* Recurrent Neural Networks (TLRNN, ENN, ec.)
* Fuzzy Predictors
* Decision & Regression Trees
* Particle Swarm Optimisation
* Support Vector Regression (SVR)
* Evolutionary & Genetic Algorithms
* Composite & Hybrid approaches
* Others