Numpy + Pandas

By: Matthew Qu & Asher Noel

Getting Started

Before we begin, we must first install the numpy and pandas libraries as they are not included in the standard Python library (check out our guide if you have any questions about installation). When we import these libraries, we typically abbreviate them as follows:

import numpy as np
import pandas as pd

This isn't necessary, but since we'll be calling functions from these libraries so often, it saves us quite a bit of typing!


Numpy (“Numerical Python”) is a general-purpose package that endows Python with efficient multi-dimensional array and matrix objects and operations.


Python lists are important, but slow. Numpy array objects aim to be 50x faster than traditional Python lists. Unlike Python lists, Numpy array objects are stored at one continuous place in memory.

On its own, Numpy’s matrix operations are commonly used for linear algebra and as the biases of all scientific work. Often, other libraries take advantage of numpy’s speed by building atop it, such as PyTorch for machine learning and Pandas for data analysis.

Why is NumPy so fast?

Python’s dynamic typing makes vanilla Python slow. Every time Python uses a variable, it has to check that variable’s data type. Lists are arrays of pointers. Even when all of the referenced objects are of the same type, memory is dynamically allocated.

In contrast, Numpy arrays are densely packed arrays of a homogeneous numerical data type. This allows memory allocation to be continuous, and operations can parallelize.



import numpy as np
x = np.array([1, 0, 0, 1])
y = np.array([-1, 5, 10, -1])
print(x + y)
# Output
array([0, 5, 10, 0])

Component-Wise Multiplicaiton

This is commonly referred to as a Hadamond product.

import numpy as np
A = np.array([[5, 10], [3, 4]])
y = np.array([[6, 20], [-4, -5]])
print(x * y)
# Output
array([[30, 200],
[-12, -20]])

With vectors, it is acceptable to use to calculate dot products. With matrices, numpy documentation says that it is preferable to use np.matmul():


In math, dot product and matrix multiplication require the same dimensions. In numpy, this is not the case. Dimensions are compatible if they are 1) the same value or 2) one of them is 1. The size of the result array is the maximum along that particular dimension.

As a simple example, we will add a scalar to an array:

import numpy as np
x = np.array([1, 10, 15, 100])
print(x + 10)
# Output
array([11, 20, 25, 110])

To read more about harder examples, check out the numpy documentation! For example, the dimension of adding a (3, 4, 5, 6) and (4, 6, 7) dimensional object will have dimension (3, 4, 5, 7).


Numpy comes with a trove of built-in functions, including np.matmul, np.zeros, np.arange, np.identity, and more. If you ever need to do something with a matrix, just check the documentation or google a query, and numpy can probably handle it!



Pandas is a powerful Python library that is specifically designed for data manipulation and analysis. Its name comes from the term panel data, which is data that contains information of individuals over a period of time.

Why use Pandas?

In general, Pandas makes it easy and intuitive to work with data; this includes cleaning, transforming, and analyzing data. Data from Pandas is also commonly used alongside other Python libraries such as SciPy, Matplotlib, and Scikit-learn for use in statistical analysis, data visualization, and machine learning, respectively.

NumPy and Pandas are almost always used in conjunction. In fact, Pandas is built on top of NumPy, and the two libraries work together internally as well. Because NumPy objects and operations are highly efficient, Pandas also executes very quickly.

Data Structures in Pandas

The two main data structures in pandas are Series and Dataframes. Series can be thought of as columns; they are a one-dimensional array of values. We can create a Series by passing in an iterable to the argument data:

import pandas as pd
s = pd.Series(data=[5, 10, 15, 20])
# Output
0 5
1 10
2 15
3 20
dtype: int64

The first column is the index; by default, it is numerically indexed starting from 0. We can create custom indices by setting the optional argument index equal to another iterable. This iterable must be of the same length as that passed into the data argument. If data is a dictionary, the keys become indices and the values make up the data.

If Series are like columns, then Dataframes are like tables with both rows and columns. As with a Series, we can create a Dataframe from scratch:

df = pd.DataFrame(data={'col1': [1, 2, 3, 4],
'col2': [5, 6, 7, 8]},
index=["row1", "row2", "row3", "row4"])
# Output
col1 col2
row1 1 5
row2 2 6
row3 3 7
row4 4 8

For a DataFrame, the keys of the dictionary become column names, not indices. We'll be working with DataFrames most of the time, but Series can arise when we extract data from a DataFrame. We can use the loc[] function to extract data based on the label of the index:

# Output
col1 1
col2 5
Name: row1, dtype: int64

Similarly, we can use the iloc[] function to extract data using numerical indices, not labels. This is useful for when the indices are relabelled.

# Output
col1 1
col2 5
Name: row1, dtype: int64

loc[] and iloc[] accept up to two arguments, the first being an expression that determines which rows to extract, and the same being the same but for the columns (by default, all columns are selected). For example, we can use slice notation to extract the entire first column:

print(df.iloc[:, 0])
# Output
row1 1
row2 2
row3 3
row4 4
Name: col1, dtype: int64

Importing and exporting data

We don't usually create DataFrames from scratch using dictionaries or lists - most of the time, we'll want to read external data stored in another file. Let's work with a real example. The data we'll be using comes from the U.S. Geological Survey of all earthquakes with magnitude 2.5 or greater that occurred on a randomly chosen day in 2020 (June 14). You can download the data here.

Pandas has a read_csv() function that automatically converts CSVs to DataFrames, using the first line as column names:

import pandas as pd
earthquakes = pd.read_csv("earthquakes.csv")

Pandas also has the functions read_json() and read_sql_query() to read data from JSON files and SQL databases.

We can similarly export data from DataFrames to other files. This can be done using the to_csv() function (or likewise the to_json() and to_sql() functions). For example, we can export the earthquakes DataFrame (presumably after some changes) and save it under the name new_earthquakes:


Working with data in Pandas

We often work with data that contains an abundance of information. For large DataFrames, we can use the head() function to examine the first five rows. We can also specify a different number of rows; for example, earthquakes.head(3) would print only the first three rows. Similarly, the tail() function can be used for viewing the end of the DataFrame.

# Output
time latitude ... locationSource magSource
0 2020-06-14T23:52:42.082Z 24.7959 ... us us
1 2020-06-14T22:52:49.059Z 38.1985 ... nn nn
2 2020-06-14T22:42:57.810Z 17.9665 ... pr pr
3 2020-06-14T22:40:36.458Z 39.4057 ... us us
4 2020-06-14T22:29:57.959Z 7.6562 ... us us
[5 rows x 22 columns]

The output is still condensed, however, and we can see that there are actually 22 columns in our DataFrame. We can inspect the column names as follows:

# Output
Index(['time', 'latitude', 'longitude', 'depth', 'mag', 'magType', 'nst',
'gap', 'dmin', 'rms', 'net', 'id', 'updated', 'place', 'type',
'horizontalError', 'depthError', 'magError', 'magNst', 'status',
'locationSource', 'magSource'],

We can also use the info() function to get column names as well as some other useful data, such as how many non-empty entries are in each column:
# Output
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 75 entries, 0 to 74
Data columns (total 22 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 time 75 non-null object
1 latitude 75 non-null float64
2 longitude 75 non-null float64
3 depth 75 non-null float64
4 mag 75 non-null float64
5 magType 75 non-null object
6 nst 23 non-null float64
7 gap 73 non-null float64
8 dmin 73 non-null float64
9 rms 75 non-null float64
10 net 75 non-null object
11 id 75 non-null object
12 updated 75 non-null object
13 place 75 non-null object
14 type 75 non-null object
15 horizontalError 72 non-null float64
16 depthError 75 non-null float64
17 magError 73 non-null float64
18 magNst 73 non-null float64
19 status 75 non-null object
20 locationSource 75 non-null object
21 magSource 75 non-null object
dtypes: float64(12), object(10)
memory usage: 13.0+ KB

Let's index by id, but to avoid creating a new DataFrame, we set an argument inplace to be True. In addition, let's extract just the first five columns because they contain the important data:

earthquakes.set_index("id", inplace=True) # modifies the existing DataFrame
earthquake_data = earthquakes.iloc[:, :5] # slice notation for every row and the first five columns
# Output
time latitude longitude depth mag
us6000aew6 2020-06-14T23:52:42.082Z 24.7959 123.2018 10.0 4.10
nn00748598 2020-06-14T22:52:49.059Z 38.1985 -117.9325 4.2 2.50
pr2020166021 2020-06-14T22:42:57.810Z 17.9665 -66.9666 7.0 2.52
us6000aew8 2020-06-14T22:40:36.458Z 39.4057 40.7696 10.0 4.50
us6000abue 2020-06-14T22:29:57.959Z 7.6562 121.7359 10.0 4.90

We can also sort by a specific column. Let's look at the most severe earthquakes for this day, so we want to sort by magnitude from highest to lowest:

earthquakes_data.sort_values(by="mag", ascending=False, inplace=True)
# Output
time latitude longitude depth mag
us6000abnv 2020-06-14T14:24:29.479Z 39.4317 40.7076 10.00 5.9
us6000abr2 2020-06-14T18:06:01.738Z 27.7759 53.3785 10.00 5.2
us6000abte 2020-06-14T21:23:19.165Z 49.1277 158.1032 10.00 5.2
us6000abp7 2020-06-14T14:43:08.390Z 23.3821 70.3548 10.00 5.1
us6000abjt 2020-06-14T03:35:19.148Z -8.9817 124.1386 68.88 5.1

We can also calculate the mean of rows and columns in the same way we would in numpy, specifying axis=0 to average over all rows (leaving columns) and axis=1 to average over all columns (leaving rows). Specifically, let's calculate the mean of the depth and mag columns, because taking the average of the other three wouldn't give us much insight:

means = earthquakes_data.iloc[:, 3:].mean(axis=0)
# Output
depth 46.2060
mag 3.8924
dtype: float64

If we want to make changes to a row or column, we can use the apply() function. This function is used on a Series and takes in a function to be applied on every element in the Series. For example, every entry in the time column has a format similar to 2020-06-14T14:24:29.479Z. We can splice the string and get rid of the date and the extra precision on the time:

new_time = earthquakes_data['time'].apply(lambda x: x[11:19]) # This slice returns the hour, minute, and second
# Output
us6000abnv 14:24:29
us6000abr2 18:06:01
us6000abte 21:23:19
us6000abp7 14:43:08
us6000abjt 03:35:19
Name: time, dtype: object

Don't forget to update the DataFrame with the new time!

earthquakes_data['time'] = new_time
# Output
time latitude longitude depth mag
us6000abnv 14:24:29 39.4317 40.7076 10.00 5.9
us6000abr2 18:06:01 27.7759 53.3785 10.00 5.2
us6000abte 21:23:19 49.1277 158.1032 10.00 5.2
us6000abp7 14:43:08 23.3821 70.3548 10.00 5.1
us6000abjt 03:35:19 -8.9817 124.1386 68.88 5.1

Pandas is also commonly used with other data visualization libraries. For example, we can use plotly to plot the locations of the epicenters on a world map:

import plotly.graph_objs as go
from plotly.offline import iplot
data = go.Scattergeo(
text='Magnitude: ' + earthquakes_data['mag'].apply(str),
marker={'symbol': 'circle', 'size': earthquakes_data['mag'] * 2}
layout = dict(
title='Worldwide Earthquakes on June 14, 2020',
projection={"type": "robinson"}
cmap = go.Figure(data=[data], layout=layout)

World map depicting earthquakes on June 14, 2020

Hopefully this guide has served as an introduction to numpy and pandas as well as their widespread usefulness in data science.