Horizon Graph

Originally called "two-tone pseudo-coloring", a horizon graph increases the density of time series graphs by dividing and layering filled line charts.

Intro:

We review Sizing the Horizon: The Effects of Chart Size and Layering on the Grahpical Perception of Time Series Visualizations.
We will use this example to illustrate these definitions:
1. Mirror
2. Offset
3. Band

Outline:

Problem: creating an effective presentation of multiple time series.
Goal: display more charts in a fixed area.
Reason: viewers need to quickly and reliably spot trends.

Initial thoughts?

Stock example.



In [1]:

    
%matplotlib inline  
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
from bokeh.charts import Horizon, output_file, show
from bokeh.io import output_notebook
%matplotlib inline  
output_notebook()









    





    
        
        Loading BokehJS ...

Visual attributes

Line Chart

Cleveland's Banking to the 45 - aspect ratio & perception.
Related paper on Banking to the 45

Stacked Graph

good: aggregate patterns
Cleveland: perception of position vs length

Horizon Graph Code

Using Bokeh, we can quickly build a stacked graph example.



In [2]:

    
# read in some stock data from the Yahoo Finance API
start_date = '2014'
end_date = '2017'
AAPL = pd.read_csv(
    "http://ichart.yahoo.com/table.csv?s=AAPL&a=0&b=1&c={sd}&d=0&e=1&f={ed}".format(sd=start_date,ed=end_date),
    parse_dates=['Date'])

MSFT = pd.read_csv(
    "http://ichart.yahoo.com/table.csv?s=MSFT&a=0&b=1&c={sd}&d=0&e=1&f={ed}".format(sd=start_date,ed=end_date),
    parse_dates=['Date'])

IBM = pd.read_csv(
    "http://ichart.yahoo.com/table.csv?s=IBM&a=0&b=1&c={sd}&d=0&e=1&f={ed}".format(sd=start_date,ed=end_date),
    parse_dates=['Date'])

TWTR = pd.read_csv(
    "http://ichart.yahoo.com/table.csv?s=TWTR&a=0&b=1&c={sd}&d=0&e=1&f={ed}".format(sd=start_date,ed=end_date),
    parse_dates=['Date'])
FB = pd.read_csv(
    "http://ichart.yahoo.com/table.csv?s=FB&a=0&b=1&c={sd}&d=0&e=1&f={ed}".format(sd=start_date,ed=end_date),
    parse_dates=['Date'])



In [3]:

    
data = dict([
    ('AAPL', AAPL['Adj Close']),
    ('Date', AAPL['Date']),
    ('FB', FB['Adj Close']),
    ('MSFT', MSFT['Adj Close']),
    #('IBM', IBM['Adj Close']), 
    ('TWTR', TWTR['Adj Close'])]
)

hp = Horizon(data, x='Date'
             , plot_width=800
             , plot_height=300,
             title="horizon plot using stock inputs")

#output_file("horizon.html")

show(hp)

Graphical Perception

Experiments

Estimation speed
Estimation accuracy

E1: mirror vs offset (vary bands)

bands
offset
mirrior

E2: line vs horizon

line chart
horizon chart

Measurement

Which point is larger? [pretty much everyone got this one correct]
What is the absolute difference between the values at these two points?

Stat note

80% trimmed mean per subject for each experiemental condition.

Issues

test size
no control for screen resolution

Results

18 subjects
More bands, longer time
2 or 3 bands, similar accuracy
4 bands descreases accuracy
less accuracy in line chart than horizon
accuracy most declined for line charts and 1 band horizon charts small in size
accuracy seems to increase with larger charts
layering increases estimation time, but no difference b/t line and 1 band horizon charts



In [ ]: