Monday, June 1, 2015

Quantitative Prediction for the Turkish General Election 2015


Final opinion polls are published. As of May 28, opinion polls are banned. At this stage of the campaign I wanted to model my quantitative prediction.

This is basically a meta-analysis. The idea is to combine different polls using weighted average. In a normal average, each data point contributes equally to the result. In a weighted average, some data points contribute more than others. I analyze how each polling organization performed in the past and consider (weigh) their latest 2015 predictions according their past performance.


Methodology:
 
I calculate the average error of each pollster's predictions for previous elections using RMS of Absolute Errors. A smaller average error indicates a more accurate pollster. I then transform these average errors into weights. For that, I simply set weight to equal the inverse of the pollster average error.

Finally I perform a weighted average of 2015 General Election predictions.


Some notes:

I used Absolute Error to calculate Delta, i.e. that's percentage points. There is no need to use Absolute Percent Error, because I am not interested in Mean Absolute Percent Error (MAPE) of all the pollsters, which is useful to get the prediction error of national opinion polls in general, i.e. forecasting accuracy of opinion polls. MAPE is good when assessing overall forecast accuracy. 

Assessing MAPE to each pollster is not a good strategy either, because MAPE is scale sensitive and should not be used when working with low-volume data.

Some people insist on using Standard Deviation on single errors in order to calculate the Mean Error, which is total nonsense. I will not explain why.

An auto-critique: "Past Performance is Not Necessarily Indicative of Future Results". If my prediction fails, this will be my strongest –and probably the only– excuse.

 
Results:

Past predictions:




 Pollsters sorted by descending weight:





Pollsters' 2015 predictions:




My quantitative prediction after calculating the weighted average:

AKP: 44.69%
CHP: 25.56%
MHP: 15.82%
HDP: 9.65%
OTHERS: 4.28%