The Manchester Withington Predict-o-Matic 5000

The 2015 election has been notable for the many new models attempting to predict the result, mostly created by academics trying to avoid doing any proper work. In that spirit we at Tomsk79 have created our own model for predicting the result of the Manchester Withington constituency, which is detailed in full here.

The starting point for my model is the constituency poll carried out in June 2014 by Lord Ashcroft (blessed be his name). This poll suggests a decisive swing from the Lib Dems to Labour since the election:

The next stage is to correct the Ashcroft poll by movements in the national polls since June 2014. For this we use the BBC Poll of Polls, which reports changes as follows:

Con: 31 to 34 (+3)
Lab: 34 to 33 (-1)
LD: 10 to 9 (-1)
UKIP: 15 to 13 (-2)
Green: 5 to 5 (0)

We assume that the national changes apply uniformly to the Withington constituency.

Next, we apply a sub-seat correction using local election results from Withington wards. By analysing trends in council results from 2010-2015, we can divide the Ashcroft results up by ward and apply corrections that can be projected to polling day. This results in a single party state correction (SPSC) which we calculate to be +1.4% LD to Lab for next week's election.

Another important data input is the normalised voter contact ratio (NVCR), which we estimate based on leafleting frequency. Assuming the Tomsk79 HQ is typical of Withington households, each voter receives 11 Lib Dem communications and 5 from Labour, giving an NVCR of 1 / (2.2^2) = 0.21, which corresponds to a 5.2% Lab to LD swing.

Once the SPSC and NVCR have been applied to the whole seat prediction, we can then take into account historical trends in polling time series, including reversion to the mean or 'swingback' and marginal-indicated protest vote propensity (MIPVP). According to our tailored swingback model we expect a 2.3% Lab to LD switch over the period of the short campaign, and a time-static boost of 2.7% for the Greens and 0.8% for UKIP overall from MIPVP.

The last macrocorrection we make is to compensate for electorodynamic effects (rather than kinetic effects as in the previous corrections). As electorodynamic data is not readily available from polling or ground observation, we estimate it using a Frequentist-uprated Dynamic General Equilibrium model. This suggests a 0.06% LD to Lab swing and a "Green squeeze" of -0.4% which is shared 0.3% to Lab and 0.1% to LD.

We now turn to local environmental and geographical issues. First, we correct for the weather forecast next Thursday. It is well known that Labour voters cannot be bothered to turn up at the first sign of drizzle, and at the time of writing the BBC weather forecast for Withington is reporting light rain showers for polling day. Using the Fitzroy-Shannon model developed by the Met Office we calculate a weather-related voter intention depression (WRVID) of 0.3% to Labour's vote and renormalise the other parties accordingly. Naturally WRVID can be refined as we get to closer to the 7th, but we do not expect it to change significantly given that the constituency is in Manchester.

So far in our model the calculations have been computationally trivial, but with modern levels of parallel compute resource we can go beyond simple seat- and ward-level predictions. We can in fact model individual voters using our own purpose-built agent-based simulation, which combines the Mosaic voter profiling system with Hawk-Eye trajectory analysis technology. Through this we can predict individual voter trajectories right up and into the polling booth with unprecedented accuracy.

Naturally the resulting trajectories are classical and do not take into account any quantum fluctuations that may influence the result at a causational level. We therefore apply a quantum correction to the Mosaic/Hawk-Eye trajectories via a quasi-classical approach, using Ehrenfest's theorem under the constraint of Duverger's law.

Finally, we use a statistical technique called Tea-Leaf Analysis (TLA) to apply a low-pass filter over all the previous corrections, thereby deconvoluting most sources of noise.

So, after all that number crunching, what is the predicted result? For no extra charge I've converted the punchcard output into human-digestible numbergrams:

Con: 8%
Lab: 47%
LD: 35%
UKIP: 3%
Green: 7%
Mysterious independent: 1% 
(nb: result adds up to 101% due to the excellence of the model)

In conclusion, we believe the Predict-o-Matic 5000 to be the most accurate constituency prediction system ever devised. Indeed our analysis indicates an 80% chance of the model being more accurate than reality itself. Any error in the final result is therefore likely to be a fault in the universe rather than a problem with the calculations.