The application of neural networks to politics has long been a personal interest – but until today that’s all it has been – an interest, not anything pursued practically.
My initial inspiration – more than twenty years ago – was the simple insight that, when canvassing for votes you often know the answer of the voter before they open the door and certainly in the majority of cases – at least, I thought – certainly before any words were exchanged: the brain of an experienced canvasser was able to compute a likely outcome from looks alone. The task, then, was to find some data that could substitute for that and speedily identify supporters using the non-linear calculations that neural nets, as analogues of the canvasser’s brain, made.
I still think that is a valid idea – though in the era of “big data” other approaches are now being used. But it wasn’t the one I experimented with today.
Working as a political consultant the thing I am often asked by clients (and other consultants) is: who do I think is going to win the next UK election? The truthful answer is “I don’t know” – there has never been a moment like this in my lifetime. But that is not a particular helpful answer – so is there something different we can do that goes beyond the uniform swing projections that anyone with an opinion poll and wikipedia could manage?
This is where a neural net might help – deep in the demographics perhaps there are non-linear relationships that will point us to the result in individual constituencies?
My initial experiments – with Christmas coming work is a little quieter and so I had a bit of time to mess about with some C code looking at this – suggest that if the relationships are there neural nets are not going to reveal them without effort.
I picked Scotland as my testing ground: the boundaries there have been unchanged since 2005 so that should give us two full sets of elections to test with and – because the insurgent populist party there, the SNP, is well established, it is a bit more data-rich than the English and Welsh experience with UKIP.
If I am able to make more progress with the model – perhaps as I read more of the really rather good Practical Neural Network Recipes in C++ – long since out of print but a great tour round the issues – I may write some more.