In [8]:
# special IPython command to prepare the notebook for matplotlib
%matplotlib inline
from fnmatch import fnmatch
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import requests
from pattern import web
# set some nicer defaults for matplotlib
from matplotlib import rcParams
#these colors come from colorbrewer2.org. Each is an RGB triplet
dark2_colors = [(0.10588235294117647, 0.6196078431372549, 0.4666666666666667),
(0.8509803921568627, 0.37254901960784315, 0.00784313725490196),
(0.4588235294117647, 0.4392156862745098, 0.7019607843137254),
(0.9058823529411765, 0.1607843137254902, 0.5411764705882353),
(0.4, 0.6509803921568628, 0.11764705882352941),
(0.9019607843137255, 0.6705882352941176, 0.00784313725490196),
(0.6509803921568628, 0.4627450980392157, 0.11372549019607843),
(0.4, 0.4, 0.4)]
rcParams['figure.figsize'] = (10, 6)
rcParams['figure.dpi'] = 150
rcParams['axes.color_cycle'] = dark2_colors
rcParams['lines.linewidth'] = 2
rcParams['axes.grid'] = True
rcParams['axes.facecolor'] = '#eeeeee'
rcParams['font.size'] = 14
rcParams['patch.edgecolor'] = 'none'
The comic refers to the fact that Nate Silver's statistical model (which is based mostly on combining information from pre-election polls) correctly predicted the outcome of the 2012 presidential race in all 50 states.
Polling data isn't a perfect predictor for the future, and some polls are more accurate than others. This means that election forecastors must consider prediction uncertainty when building models.
In this first assignment, you will perform a simple analysis of polling data about the upcoming Governor races. The assignment has three main parts:
First you will build some tools to download historical polling data from the web, and parse it into a more convenient format.
Next you will use these tools to aggregate and visualize several past Governor races
Finally you will run a bootstrap analysis to estimate the probable outcome of current Governor races, given the level of precision of historical polls.
The Real Clear Politics website archives many political polls. In addition, they combine related polls to form an "RCP average" estimate of public opinion over time. For example, the chart on this page shows historical polling data for the Obama-Romney presidential race. The chart is an average of the polling data table below the chart.
The data used to generate plots like this are stored as XML pages, with URLs like:
http://charts.realclearpolitics.com/charts/[id].xml
Here, [id] is a unique integer, found at the end of the URL of the page that displays the graph. The id for the Obama-Romney race is 1171:
http://charts.realclearpolitics.com/charts/1171.xml
Opening this page in Google Chrome or Firefox will show you the XML content in an easy-to-read format. Notice that XML tags are nested inside each other, hierarchically (the jargony term for this is the "Document Object Model", or "DOM"). The first step of webscraping is almost always exploring the HTML/XML source in a browser, and getting a sense of this hierarchy.
The above XML page includes 5 distinct tags (one, for example, is chart). List these tags, and depict how they nest inside each other using an indented list. For example:
We want to download and work with poll data like this. Like most programming tasks, we will break this into many smaller, easier pieces
Fill in the code for the get_poll_xml function, that finds and downloads an XML page discussed above
Hint
requests.get("http://www.google.com").text downloads the text from Google's homepage
In [9]:
"""
Function
--------
get_poll_xml
Given a poll_id, return the XML data as a text string
Inputs
------
poll_id : int
The ID of the poll to fetch
Returns
-------
xml : str
The text of the XML page for that poll_id
Example
-------
>>> get_poll_xml(1044)
u'<?xml version="1.0" encoding="UTF-8"?><chart><series><value xid=\'0\'>1/27/2009</value>
...etc...
"""
#your code here
def get_poll_xml(poll_id):
url = "http://charts.realclearpolitics.com/charts/%i.xml" % int(poll_id)
return requests.get(url).text
Here are some other functions we'll use later. plot_colors contains hints about parsing XML data.
In [10]:
# "r"egular "e"xpressions is kind of a mini-language to
# do pattern matching on text
import re
def _strip(s):
"""This function removes non-letter characters from a word
for example _strip('Hi there!') == 'Hi there'
"""
return re.sub(r'[\W_]+', '', s)
def plot_colors(xml):
"""
Given an XML document like the link above, returns a python dictionary
that maps a graph title to a graph color.
Both the title and color are parsed from attributes of the <graph> tag:
<graph title="the title", color="#ff0000"> -> {'the title': '#ff0000'}
These colors are in "hex string" format. This page explains them:
http://coding.smashingmagazine.com/2012/10/04/the-code-side-of-color/
Example
-------
>>> plot_colors(get_poll_xml(1044))
{u'Approve': u'#000000', u'Disapprove': u'#FF0000'}
"""
dom = web.Element(xml)
result = {}
for graph in dom.by_tag('graph'):
title = _strip(graph.attributes['title'])
result[title] = graph.attributes['color']
return result
Even though get_poll_xml pulls data from the web into Python, it does so as a block of text. This still isn't very useful. Use the web module in pattern to parse this text, and extract data into a pandas DataFrame.
Hints
pd.DataFrame({'column_label_1': list_1, 'column_label_2':list_2, ...})
pd.to_datetime to convert strings into dates
In [11]:
"""
Function
---------
rcp_poll_data
Extract poll information from an XML string, and convert to a DataFrame
Parameters
----------
xml : str
A string, containing the XML data from a page like
get_poll_xml(1044)
Returns
-------
A pandas DataFrame with the following columns:
date: The date for each entry
title_n: The data value for the gid=n graph (take the column name from the `title` tag)
This DataFrame should be sorted by date
Example
-------
Consider the following simple xml page:
<chart>
<series>
<value xid="0">1/27/2009</value>
<value xid="1">1/28/2009</value>
</series>
<graphs>
<graph gid="1" color="#000000" balloon_color="#000000" title="Approve">
<value xid="0">63.3</value>
<value xid="1">63.3</value>
</graph>
<graph gid="2" color="#FF0000" balloon_color="#FF0000" title="Disapprove">
<value xid="0">20.0</value>
<value xid="1">20.0</value>
</graph>
</graphs>
</chart>
Given this string, rcp_poll_data should return
result = pd.DataFrame({'date': pd.to_datetime(['1/27/2009', '1/28/2009']),
'Approve': [63.3, 63.3], 'Disapprove': [20.0, 20.0]})
"""
#your code here
def rcp_poll_data(xml):
dom = web.Element(xml)
result = {}
dates = dom.by_tag('series')[0]
dates = {n.attributes['xid']: str(n.content) for n in dates.by_tag('value')}
keys = dates.keys()
result['date'] = pd.to_datetime([dates[k] for k in keys])
for graph in dom.by_tag('graph'):
name = graph.attributes['title']
data = {n.attributes['xid']: float(n.content)
if n.content else np.nan for n in graph.by_tag('value')}
result[name] = [data[k] for k in keys]
result = pd.DataFrame(result)
result = result.sort(columns=['date'])
return result
The output from rcp_poll_data is much more useful for analysis. For example, we can plot with it:
In [12]:
def poll_plot(poll_id):
"""
Make a plot of an RCP Poll over time
Parameters
----------
poll_id : int
An RCP poll identifier
"""
# hey, you wrote two of these functions. Thanks for that!
xml = get_poll_xml(poll_id)
data = rcp_poll_data(xml)
colors = plot_colors(xml)
#remove characters like apostrophes
data = data.rename(columns = {c: _strip(c) for c in data.columns})
#normalize poll numbers so they add to 100%
norm = data[colors.keys()].sum(axis=1) / 100
for c in colors.keys():
data[c] /= norm
for label, color in colors.items():
plt.plot(data.date, data[label], color=color, label=label)
plt.xticks(rotation=70)
plt.legend(loc='best')
plt.xlabel("Date")
plt.ylabel("Normalized Poll Percentage")
If you've done everything right so far, the following code should reproduce the graph on this page
In [13]:
poll_plot(1044)
plt.title("Obama Job Approval")
Out[13]:
Unfortunately, these data don't have any error bars. If a candidate leads by 10% in the RCP average, is she a shoo-in to win? Or is this number too close to call? Does a 10% poll lead mean more 1 day before a race than it does 1 week before? Without error estimates, these questions are impossible to answer.
To get a sense of how accurate the RCP polls are, you will gather data from many previous Governor races, where the outcome is known.
This url has links to many governer races.
http://www.realclearpolitics.com/epolls/2010/governor/2010_elections_governor_map.html
Notice that each link to a governor race has the following URL pattern:
http://www.realclearpolitics.com/epolls/[YEAR]/governor/[STATE]/[TITLE]-[ID].html
Write a function that scans html for links to URLs like this
Hint The fnmatch function is useful for simple string matching tasks.
In [14]:
"""
Function
--------
find_governor_races
Find and return links to RCP races on a page like
http://www.realclearpolitics.com/epolls/2010/governor/2010_elections_governor_map.html
Parameters
----------
html : str
The HTML content of a page to scan
Returns
-------
A list of urls for Governer race pages
Example
-------
For a page like
<html>
<body>
<a href="http://www.realclearpolitics.com/epolls/2010/governor/ma/massachusetts_governor_baker_vs_patrick_vs_cahill-1154.html"></a>
<a href="http://www.realclearpolitics.com/epolls/2010/governor/ca/california_governor_whitman_vs_brown-1113.html"></a>
</body>
</html>
find_governor_races would return
['http://www.realclearpolitics.com/epolls/2010/governor/ma/massachusetts_governor_baker_vs_patrick_vs_cahill-1154.html',
'http://www.realclearpolitics.com/epolls/2010/governor/ca/california_governor_whitman_vs_brown-1113.html']
"""
#your code here
def is_gov_race(l):
"""return True if a URL refers to a Governor race"""
pattern = 'http://www.realclearpolitics.com/epolls/????/governor/??/*-*.html'
return fnmatch(l, pattern)
def find_governor_races(html):
dom = web.Element(html)
links = [a.attributes.get('href', '') for a in dom.by_tag('a')]
links = [l for l in links if is_gov_race(l)]
#eliminate duplicates!
links = list(set(links))
return links
At this point, you have functions to find a collection of governor races, download historical polling data from each one, parse them into a numerical DataFrame, and plot this data.
The main question we have about these data are how accurately they predict election outcomes. To answer this question, we need to grab the election outcome data.
Write a function that looks up and returns the election result on a page like this one.
Remember to look at the HTML source!
You can do this by selection view->developer->view source in Chrome, or Tools -> web developer -> page source in Firefox. Altenatively, you can right-click on a part of the page, and select "inspect element"
In [15]:
"""
Function
--------
race_result
Return the actual voting results on a race page
Parameters
----------
url : string
The website to search through
Returns
-------
A dictionary whose keys are candidate names,
and whose values is the percentage of votes they received.
If necessary, normalize these numbers so that they add up to 100%.
Example
--------
>>> url = 'http://www.realclearpolitics.com/epolls/2010/governor/ca/california_governor_whitman_vs_brown-1113.html'
>>> race_result(url)
{'Brown': 56.0126582278481, 'Whitman': 43.9873417721519}
"""
#your code here
def race_result(url):
dom = web.Element(requests.get(url).text)
table = dom.by_tag('div#polling-data-rcp')[0]
result_data = table.by_tag('tr.final')[0]
td = result_data.by_tag('td')
results = [float(t.content) for t in td[3:-1]]
tot = sum(results) / 100
#get table headers
headers = table.by_tag('th')
labels = [str(t.content).split('(')[0].strip() for t in headers[3:-1]]
return {l:r / tot for l, r in zip(labels, results)}
Here are some more utility functions that take advantage of what you've done so far.
In [16]:
def id_from_url(url):
"""Given a URL, look up the RCP identifier number"""
return url.split('-')[-1].split('.html')[0]
def plot_race(url):
"""Make a plot summarizing a senate race
Overplots the actual race results as dashed horizontal lines
"""
#hey, thanks again for these functions!
id = id_from_url(url)
xml = get_poll_xml(id)
colors = plot_colors(xml)
if len(colors) == 0:
return
#really, you shouldn't have
result = race_result(url)
poll_plot(id)
plt.xlabel("Date")
plt.ylabel("Polling Percentage")
for r in result:
plt.axhline(result[r], color=colors[_strip(r)], alpha=0.6, ls='--')
Now that this is done, we can easily visualize many historical Governer races. The solid line plots the poll history, the dotted line reports the actual result.
If this code block fails, you probably have a bug in one of your functions.
In [17]:
page = requests.get('http://www.realclearpolitics.com/epolls/2010/governor/2010_elections_governor_map.html').text.encode('ascii', 'ignore')
for race in find_governor_races(page):
plot_race(race)
plt.show()
Briefly summarize these graphs -- how accurate is the typical poll a day before the election? How often does a prediction one month before the election mispredict the actual winner?
Based on visual inspection, these polls generally seem to be within about 5% of the election outcome. The accuracy is higher shortly before the election (usually within 2% or so). In about 20% of cases, the eventual winner trails at some point during the last month of polling -- examples include the Brady/Quinn Illinois race, the Foley/Malloy Connecticut race, the Scott/Sink Florida Race, the Kitzhaber/Dudley Oregon Race, the Corzine/Christie New Jersey race, and the Caprio/Robitaille/Chafee Rhode Island Race.
You are (finally!) in a position to do some quantitative analysis.
We have provided an error_data function that builds upon the functions you have written. It computes a new DataFrame with information about polling errors.
Use error_data, find_governer_races, and pd.concat to construct a Data Frame summarizing the forecast errors
from all the Governor races
Hint
It's best to set ignore_index=True in pd.concat
In [18]:
def party_from_color(color):
if color in ['#0000CC', '#3B5998']:
return 'democrat'
if color in ['#FF0000', '#D30015']:
return 'republican'
return 'other'
def error_data(url):
"""
Given a Governor race URL, download the poll data and race result,
and construct a DataFrame with the following columns:
candidate: Name of the candidate
forecast_length: Number of days before the election
percentage: The percent of poll votes a candidate has.
Normalized to that the canddidate percentages add to 100%
error: Difference between percentage and actual race reulst
party: Political party of the candidate
The data are resampled as necessary, to provide one data point per day
"""
id = id_from_url(url)
xml = get_poll_xml(id)
colors = plot_colors(xml)
if len(colors) == 0:
return pd.DataFrame()
df = rcp_poll_data(xml)
result = race_result(url)
#remove non-letter characters from columns
df = df.rename(columns={c: _strip(c) for c in df.columns})
for k, v in result.items():
result[_strip(k)] = v
candidates = [c for c in df.columns if c is not 'date']
#turn into a timeseries...
df.index = df.date
#...so that we can resample at regular, daily intervals
df = df.resample('D')
df = df.dropna()
#compute forecast length in days
#(assuming that last forecast happens on the day of the election, for simplicity)
forecast_length = (df.date.max() - df.date).values
forecast_length = forecast_length / np.timedelta64(1, 'D') # convert to number of days
#compute forecast error
errors = {}
normalized = {}
poll_lead = {}
for c in candidates:
#turn raw percentage into percentage of poll votes
corr = df[c].values / df[candidates].sum(axis=1).values * 100.
err = corr - result[_strip(c)]
normalized[c] = corr
errors[c] = err
n = forecast_length.size
result = {}
result['percentage'] = np.hstack(normalized[c] for c in candidates)
result['error'] = np.hstack(errors[c] for c in candidates)
result['candidate'] = np.hstack(np.repeat(c, n) for c in candidates)
result['party'] = np.hstack(np.repeat(party_from_color(colors[_strip(c)]), n) for c in candidates)
result['forecast_length'] = np.hstack(forecast_length for _ in candidates)
result = pd.DataFrame(result)
return result
In [19]:
"""
function
---------
all_error_data
Calls error_data on all races from find_governer_races(page),
and concatenates into a single DataFrame
Parameters
-----------
None
Examples
--------
df = all_error_data()
"""
#your code here
def all_error_data():
data = [error_data(race_page) for race_page in find_governor_races(page)]
return pd.concat(data, ignore_index=True)
In [20]:
errors = all_error_data()
Here's a histogram of the error of every polling measurement in the data
In [21]:
errors.error.hist(bins=50)
plt.xlabel("Polling Error")
plt.ylabel('N')
Out[21]:
In [22]:
#your code here
errors.error.std()
Out[22]:
In [23]:
#your code here
print "< 7 days: %0.2f" % errors[errors.forecast_length < 7].error.std()
print ">30 days: %0.2f" % errors[errors.forecast_length > 30].error.std()
This is basically what we would expect, and what the plots suggest as well -- the accuracy of polls shortly before the election are about 40% more accurate than polls more than a month before the election.
Bootstrap resampling is a general purpose way to use empirical data like the errors DataFrame to estimate uncertainties. For example, consider the Viriginia Governor Race. If we wanted to estimate how likey it is that McAuliffe will win given the current RCP data, the approch would be:
errors.error. We are assuming that these numbers represent a reasonable error distribution for the current poll data.Bootstrapping isn't foolproof: it makes the assumption that the previous Governor race errors are representative of the Virginia race, and it does a bad job at estimating very rare events (with only ~30 races in the errors DataFrame, it would be hard to accurately predict probabilities for 1-in-a-million scenarios). Nevertheless, it's a versatile technique.
Use bootstrap resampling to estimate how likely it is that each candidate could win the following races.
Summarize your results in a paragraph. What conclusions do you draw from the bootstrap analysis, and what assumptions did you make in reaching this conclusion. What are some limitations of this analysis?
In [33]:
#your code here
def bootstrap_result(c1, c2, errors, nsample=1000):
"""
Given the current polling data for 2 candidates, return the
bootstrap-estimate for the win probability of each candidate
Parameters
----------
c1 : float
The current proportion of poll votes for candidate 1
c2 : float
The current proportio of poll votes for candidate 2
errors : DataFrame
The errors DataFrame
nsample : int
The number of bootstrap iteraionts. Default=1000
Returns
-------
p1, p2
The probability that each candidate will win, based on the bootstrap simulations
"""
#first, normalize votes to 100
tot = (c1 + c2)
c1 = 100. * c1 / tot
c2 = 100. * c2 / tot
indices = np.random.randint(0, errors.shape[0], nsample)
errors = errors.error.irow(indices).values
#errors are symmetrical -- an overestimate for candidate 1
#is an underestimate for candidate 2
c1_actual = c1 - errors
c2_actual = c2 + errors
p1 = (c1_actual > c2_actual).mean()
p2 = 1 - p1
return p1, p2
#Look up the data as of 9/24/2013
#virginia
nsample = 10000
mcauliffe, cuccinelli = 43.0, 39.0
pm, pc = bootstrap_result(mcauliffe, cuccinelli, errors, nsample=nsample)
print "Virginia Race"
print "-------------------------"
print "P(McAuliffe wins) = %0.2f" % pm
print "P(Cuccinelli wins) = %0.2f" % pc
#new jersey
print "\n\n"
print "New Jersey Race"
print "-----------------------"
christie, buono = 55.4, 31.8
pc, pb = bootstrap_result(christie, buono, errors, nsample=nsample)
print "P(Christie wins) = %0.2f" % pc
print "P(Buono wins) = %0.2f" % pb
The Virginia race is currently fairly close. Our bootstrap simulations assume that historical RCP poll accuracies are representative of the uncertainty in the current polls. In 10,000 of these simulations, McCauliffe won 70% of the time.
The New Jersey race, on the other hand, is less close. In 10,000 simulations, a candidate enjoying a lead as big as Chris Christie wins about 98% of the time
For comparison, most of the predictions in Nate Silver's presidental forecast had confidences of >95%. This is more precise than what we can estimate from the RCP poll alone. His approach, however, is the same basic idea (albeit he used many more polls, and carefully calibrated each based on demographic and other information). Homework 2 will dive into some of his techniques further.
To submit your homework, create a folder named lastname_firstinitial_hw0 and place this notebook file in the folder. If your notebook requires any additional data files to run (it shouldn't), add them to this directory as well. Compress the folder (please use .zip compression) and submit to the CS109 dropbox in the appropriate folder. If we cannot access your work because these directions are not followed correctly, we will not grade your work.
css tweaks in this cell