Oct 10, 2020
Predicting Victorian State Elections: STRONG Transition Model
I’ve recently started a series on various approaches to forecasting the results of elections in the Victorian Legislative Assembly and testing how good of a job they would have done for the 2018 election.
The models tested so far have all been taken from Electoral Calculus (https://www.electoralcalculus.co.uk/)and are essentially deterministic models that attempt to transform statewide polls on voter intention into a concrete prediction of the seats held by different parties (on a TPP basis at least). So far we’ve looked at:
- The UNS model:
- The Transition model:
Which had pretty similar performance in terms of predicting the TPP vote share in each seat. Today’s article is about another model used by Electoral Calculus. The Strong Transition model.
This article won’t be very long because the results aren’t exactly novel. The Strong Transition model is designed to work by working like the UNS model for seats where the party’s support is high, and working proportionally for seats where the party’s support is low. Since we’re estimating TPP vote share where no parties support is sufficiently low, it behaves just like the UNS model.
Why did I bother writing this article then? Mainly to show the code used to implement this model. Since I’ve already implemented the model, I may as well share the code now since I may use it again in a more through model that goes beyond TPP vote shares.
I won’t go through the details because I am unsure if I’ll end up using this model and, like I said, its predictions for the TPP vote shares are the same as the UNS model. You can read the details and advantages of this model over the transition and UNS models here: https://www.electoralcalculus.co.uk/strongmodel.html.
Here’s the code:
import pandas as pd
pd.plotting.register_matplotlib_converters()
import matplotlib.pyplot as plt
from sklearn.metrics import mean_absolute_errordata2018 = pd.read_csv('~/Desktop/WORK/Python/Vic_Election/data/2pp2018.csv')
data2014 = pd.read_csv('~/Desktop/WORK/Python/Vic_Election/data/2pp2014.csv')def clean_voteshare(n):
n = list(n)
n.remove('%')
n = ''.join(n)
n = float(n)
return nfor x in range(0, len(data2014.index)):
data2014.loc[x, 'ALP_share'] = clean_voteshare(data2014.loc[x, 'ALP_share'])
data2018.loc[x, 'ALP_share'] = clean_voteshare(data2018.loc[x, 'ALP_share'])
data2014.loc[x, 'LNP_share'] = clean_voteshare(data2014.loc[x, 'LNP_share'])
data2018.loc[x, 'LNP_share'] = clean_voteshare(data2018.loc[x, 'LNP_share'])parties = ['ALP', 'LNP']# Set threshold and then get lists for statewide support and strong support of each partythresh = 20Vs = {}V = {}
for party in parties:
num = 0
for x in range(0, len(data2014.index)):
num += data2014.loc[x, 'Total votes']*data2014.loc[x, party+'_share']
V[party] = num/sum(list(data2014['Total votes']))for party in parties:
num = 0
for x in range(0, len(data2014.index)):
num += data2014.loc[x, 'Total votes']*max(data2014.loc[x, party+'_share'] - thresh, 0)
Vs[party] = num/sum(list(data2014['Total votes']))# Candidate P's:
# P = {'ALP': 57.3, 'LNP': 42.7}, actual election result, i.e. polls accurate
# P = {'ALP': 55, 'LNP': 45}, YouGov 24th Nov 2018P = {'ALP': 55, 'LNP': 45} # Can change this as needed
Pw = {}
Ps = {}# Here I set up the the predicted strong and weak support for each party.for party in parties:
Pw[party] = max(P[party] - Vs[party], 0)
Ps[party] = min(P[party], Vs[party])###################################################################V_splits = {}
for x in Vs.keys():
V_splits[x+'_strong'] = Vs[x]
V_splits[x+'_weak'] = V[x] - Vs[x]
P_splits = {}
for x in Pw.keys():
P_splits[x+'_weak'] = Pw[x]
for x in Ps.keys():
P_splits[x+'_strong'] = Ps[x]data2014['ALP_strong'] = 0
for x in range(0, len(data2014.index)):
data2014.loc[x, 'ALP_strong'] = max(data2014.loc[x, 'ALP_share'] - thresh,0)
data2014['ALP_weak'] = data2014['ALP_share'] - data2014['ALP_strong']data2014['LNP_strong'] = 0
for x in range(0, len(data2014.index)):
data2014.loc[x, 'LNP_strong'] = max(data2014.loc[x, 'LNP_share'] - thresh,0)
data2014['LNP_weak'] = data2014['LNP_share'] - data2014['LNP_strong']
# From here on out we are just implementing the transition model on our extended party list
# since I am lazy and this is mostly the same as the code in the transition model
# article, I have been very lax with commenting here (sorry, I know it's bad).
total_share = 0
for x in V_splits.keys():
total_share += max(P_splits[x] - V_splits[x], 0)
share = {}
for x in V_splits.keys():
share[x] = max(P_splits[x] - V_splits[x], 0)/total_share
class electorate(object):
def __init__(self, C, name):
self.C = C
self.name = name
def A(self, party):
if P_splits[party] - V_splits[party] < 0:
return self.C[party] * (P_splits[party]/V_splits[party])
else:
return self.C[party] + share[party]*SwingVoters[self.name]
districts = []for x in range(0, len(data2014.index)):
district = []
district.append(data2014.loc[x, 'District'])
district.append(data2014.loc[x, 'ALP_strong'])
district.append(data2014.loc[x, 'ALP_weak'])
district.append(data2014.loc[x, 'LNP_strong'])
district.append(data2014.loc[x, 'LNP_weak'])
districts.append(district)
SwingVoters = {}
district_objs = {}for x in districts:
C= {'ALP_strong': x[1], 'ALP_weak': x[2], 'LNP_strong': x[3], 'LNP_weak': x[4]}
swingvoters = 0
for party in V_splits.keys():
swingvoters += C[party] * max(1-P_splits[party]/V_splits[party] ,0)
SwingVoters[x[0]] = swingvoters
district_objs[x[0]] = electorate(C, x[0])preds_alp = {}
preds_lib = {}
for x in district_objs.keys():
preds_alp[x] = district_objs[x].A('ALP_strong') + district_objs[x].A('ALP_weak')
preds_lib[x] = district_objs[x].A('LNP_strong') + district_objs[x].A('LNP_weak')
real_alp = {}
real_lib = {}for x in range(0, len(data2018.index)):
district = data2014.loc[x, 'District']
real_alp[district] = data2018.loc[x, 'ALP_share']
real_lib[district] = data2018.loc[x, 'LNP_share']
d = {'electorate': list(district_objs.keys()), 'alp_real':list(real_alp.values()),
'alp_preds': list(preds_alp.values()), 'lib_real': list(real_lib.values()),
'lib_preds': list(preds_lib.values())}df = pd.DataFrame(data=d)lineStart = df.alp_real.min()
lineEnd = df.alp_real.max()fig = plt.figure(dpi=800)
plt.figure()
plt.scatter(x=df['alp_preds'], y=df['alp_real'], color='r', marker = '2')
plt.scatter(x=df['lib_preds'], y=df['lib_real'], color='b', marker = '2')
plt.plot([lineStart, lineEnd], [lineStart, lineEnd], 'k-', color = 'black')
plt.plot([50, 50], [lineStart, lineEnd], color='black', linestyle='dashed', linewidth=1)
plt.plot([lineStart, lineEnd], [50, 50], color='black', linestyle='dashed', linewidth=1)
plt.xlim(lineStart, lineEnd)
plt.ylim(lineStart, lineEnd)
plt.xlabel('Predicted Voteshare')
plt.ylabel('Actual Voteshare')
plt.title('STRONG Model Accuracy for 2PP Voteshare in 2018')
#plt.savefig('accuracy_Transition.pdf')
plt.show()# Calculating the MAE
# (since we are calculating this for 2PP votes where party[1] = 100 - party[0]
# the MAE will be the same regardless of whether we choose alp or lib data to compare)MAE = mean_absolute_error(df['alp_real'], df['alp_preds'])
print('The MAE of predictions is: ' + str(MAE))# Getting a list of incorrect predictions:
alp_fp = []
alp_fn = []for x in range(0, len(df.index)):
if df.loc[x, 'alp_real'] < 50 and df.loc[x, 'alp_preds'] >= 50:
alp_fp.append(df.loc[x,'electorate'])
if df.loc[x, 'alp_real'] >= 50 and df.loc[x, 'alp_preds'] < 50:
alp_fn.append(df.loc[x,'electorate'])print('False positives for the ALP/ False negatives for LNP were: ')
print(alp_fp)
print()
print('False negatives for the ALP/ False positives for the LNP were: ')
print(alp_fn)# Finally, let's look at the total predicted ALP seats v actual ALP seats:alp_wins = []
alp_pred_wins = []
lib_wins = []
lib_pred_wins = []for x in range(0, len(df.index)):
if df.loc[x, 'alp_real'] > 50:
alp_wins.append(df.loc[x,'electorate'])
else:
lib_wins.append(df.loc[x,'electorate'])
if df.loc[x, 'alp_preds'] > 50:
alp_pred_wins.append(df.loc[x,'electorate'])
else:
lib_pred_wins.append(df.loc[x,'electorate'])
print('The ALP was predicted to win: ' + str(len(alp_pred_wins)) + ' seats')
print('They actually won: ' + str(len(alp_wins)) + ' seats')
print()
print('The LNP was predicted to win: ' + str(len(lib_pred_wins)) + ' seats')
print('They actually won: ' + str(len(lib_wins)) + ' seats')
If I end up using this model again, I’ll be sure to go through the details myself.
Stay tuned for more.
