Part one of the intermediate series. Get started using Python, pandas, numpy, seaborn and matplotlib to analyze Fantasy Football. In this part, we look at the relationship between usage and fantasy points per game.
If you have any questions about the code here, feel free to reach out to me on Twitter or on Reddit.
If you like Fantasy Football and have an interest in learning how to code, check out our Ultimate Guide on Learning Python with Fantasy Football Online Course. Here is a link to purchase for 15% off. The course includes 15 chapters of material, 14 hours of video, hundreds of data sets, lifetime updates, and a Slack channel invite to join the Fantasy Football with Python community.
In this post we are going to be begin a series on using the programming language Python for fantasy football data analysis. Over the past season, I started using Python to make decisions on who to start every week, and maybe it’s just a coincidence – but I finally won the ‘ship this year (insert humble brag).
I believe Python can be an incredible tool for making fantasy football decisions, but there is little information on the internet on how to actually get started. This post is going to be about setting up a Python environment with anaconda, jupyter, pandas, seaborn, and matplotlib, and then at the end we will do some basic analysis of efficiency versus usage for running backs.
To start, you are going to need Anaconda, which is essentially a distribution of Python made for data science. You can find a link to the download section here.
Easily enough, anaconda already comes with almost everything we need to get started (Including jupyter, numpy, pandas, seaborn, and matplotlib). Anaconda also comes with it’s own terminal, which we’ll be using for the remainder of the series. Fire up the Anaconda Prompt after you’re done installing, and make a new directory for your jupyter notebooks (More on this in a bit).
We’re going to be doing almost all our Python in Jupyter notebooks. For those unfamiliar with Jupyter notebooks, they are essentially a web based interface where you can write Python code in “cells”. They’re really useful for exploratory hacking, and you’ll see why in a minute. In your directory, type the following command to open a new instance of Jupyter.
You should see that Jupyter is now running on your localhost. If it hasn’t already automatically popped up for you, type into your browser localhost:[whatever port your terminal says]/tree. Once you’re there, click on the “new” dropdown on the right hand side, and from the dropdown select to create a new Python3 notebook.
You should see this in your browser. Change the name from “Untitled” to whatever you want. I named mine the name of this post but it doesn’t really matter.
We are almost done with the setup before we start coding. The last thing we need to do is get a dataset. I was able to easily grab data sets from profootballreference.com for free, however, you’re free to use whatever source you want, as long as the file is in CSV/Excel format. For this part, we are going to be using a 2019 dataset consisting of every WR, RB, TE, and QB that played this year. Later, we will filter and separate this dataset into four different data sets based on position using the Python library pandas. I’ve included the CSV file taken from profootball reference. Download and place the file within your project directory (This is important. If it’s not in the same directory as your jupyter notebooks we won’t be able to read it). Save the file as 2019.csv
Finally, we can begin coding. Again, for those unfamiliar with Jupyter notebooks, we are writing python code within these “cells” and then running these cells using Shift+Enter. You can write multiple lines of code within each cell, and it’s probably better that you do. Just separate each cell based off function or purpose much like you would any other .py file. For more information on jupyter, check out the docs here https://jupyter.org/documentation.
We’ll begin by importing some of the libraries we are going to be using.
Briefly, pandas is a python library included in anaconda which allows us to work with “DataFrames”. DataFrames are like Excel spreadsheets that can be manipulated using pandas. Pandas will be the most important and powerful library we will use and probably provide the steepest learning curve. For the purposes of this post, we won’t be going over all the basics of pandas as that would necessitate an entire post. I advise you check out the pandas documentation here https://pandas.pydata.org/pandas-docs/stable/ . Matplotlib and seaborn are both used for data visualization and plotting, and the basics can be learned fairly quickly. Run this cell using Shift+Enter and ensure there are no import errors (There shouldn’t be).
Next, we are going to import our CSV file and create a pandas DataFrame, clean up the data, and then separate the DataFrame into four different DataFrames for each position being analyzed – RB, WR, QB, and TE.
This is a lot to unpack, so let’s take it step by step.
On the first line, we use the function from pandas called read_csv to read our CSV file and convert it to a DataFrame. If you type in df on the next line and run the cell, you can actually see our DataFrame.
You’ll probably notice some issues with our DataFrame right away. First of all, the formatting for the player names is all messed up. Second of all, there’s a lot of unnecessary columns that probably won’t be necessary for the purposes of our analysis. Third of all, we have repeating values (eg. TD, TD.1, TD.2, TD.3).
To drop columns from a pandas DataFrame, you simple use the .drop built-in method for the DataFrame, pass in a list of column names, set axis = 1, and set inplace = True. Setting axis = 1 tells pandas that we are committing a change on the column axis. In pandas, 0 is the row axis and 1 is the column axis. Setting inplace=True tells pandas to make a permanent change to our DataFrame.
By the way, if you want to see your DataFrame but don’t want to produce the entire table, use the built-in methods .head() and .tail(). By default, .head() produces a DataFrame with the top 5 rows (try to guess what .tail() does). You can pass in an argument for the number of rows but the default is 5. For example, df.head(10) would produce the top 10 rows.
Another built-in method for pandas DataFrames is rename, which allows you to rename column names. Again, we use axis=1 to indicate that we are making a change on the column axis, and we set inplace=True to tell pandas that we want to make a permanent change to our DataFrame.
Here, we are creating four different DataFrames based off the column row ‘FantPos’. For rb_df, for example, we are saying to pandas create a new DataFrame where ‘FantPos’ = ‘RB’. The syntax for how this is done is df[df[“Column Name”] == “Some Value”]]. We can also use our other operators besides ==, such as, >, <, !=, etc. We will see this is extremely powerful and useful when we want some minimum criteria. Later on, we will be evaluating the correlation between TD/Number of Carries (Efficiency) and Fantasy Points/GM. Obviously, we want to eliminate those players with a small number of carries. For example, fantasy vultures who only get utilized in the red zone (Looking at you, Taysom Hill). Although these types of players can be efficient, they are never really start-able. We want to eliminate these types of players through a minimum criteria so we do not have too many outliers.
This first cell looks kind of complicated but it’s really not. What we are doing here is essentially assigning columns to the different DataFrames. Although trick plays happen where running backs and wide receivers sometimes throw the ball, this is not something we can ever predict, so we can exclude those columns from our wr_df and rb_df for the sake of simplicity. The syntax for filtering DataFrames and creating new ones based on desired column names is df = df[[‘ColumnName1’, ‘ColumnName2’,]]
I wrote a little function to make the code more legible and simplistic. But essentially we are concatenating lists and following the syntax shown above. All players should have a fumble column, but it’s kind of miscellaneous stat that’s hard to predict so we leave it at the end (Unless your name is Daniel Jones – then we know you’re going to fumble). Our function returns a new DataFrame, and we assign it to a new variable. You’re welcome to skip this whole step if you want to include every column (as tight-ends do run the ball sometimes and wide receivers do sometimes pass the ball).
At the end, we run rb_df.head() to see our new DataFrame. We will be working with the rb_df for the remainder of the post for the sake of brevity, but you are welcome to run these data visualizations on any of the created DataFrames.
So now that we have our DataFrame ready, we can start to use pandas to answer some fantasy football questions. One maxim I followed this year when deciding on draft picks, waiver-wire pickups, and sit-start decisions was usage is king. Unless his name was Aaron Jones, I avoided running back committees at all cost. This maxim becomes more and more obvious the more you play fantasy, but the temptation to start the Will Fuller’s and Sammy Watkin’s of the fantasy world are forever present (Against my better judgement, I still started Will Fuller Week 16 against Tampa Bay, oops).
But let’s not just rely on intuition, as our intuition is sometimes wrong. Let’s try to answer the question –
“What is more correlated to fantasy performance for running backs, usage or efficiency?”
Below is the code for plotting a graph of usage (the x-axis) versus fantasy points per game (the y-axis)
The resulting graph is show below. No surprises here, usage and fantasy points per game are heavily correlated.
Let’s break down the code step by step again.
We can create a new column with pandas based off the value of other columns by doing df[“New column name” = df[“other column”] + …
Here, we want to create a column that calculates total fantasy points. We used 0.1 points per rushing yard + 6 points per rushing TD + 1 point per reception + 0.1 points per receiving yard + 6 points per receiving TD – 2 points per fumble lost.
You’re welcome to change the weighting of points per reception to 0.5 if you play in a 0.5PPR league or remove it altogether if you use standard.
We essentially did the same thing here. We created a new column based off a previous column, and then used the .apply() built-in method and a lambda function to round each answer to two decimal places.
Same thing here. We created a new column for “Usage/GM”. We defined usage as number of targets + number of carries. I like to use targets and not catches for analyzing players. I believe targets are a better measure of usage, and you should prioritize players who are seeing an increasing number of targets and target share. But that’s a whole different discussion.
There’s a bit of magic happening here and I don’t want to delve to deep into this for the sake of the brevity of this post. Check out the documentation for seaborn and matplotlib/pyplot here. But essentially, we are just setting styling with sns.setstyle(). A list of possible arguments for setstyle() can be found in the seaborn docs. What we are doing with plt.subplots() is essentially creating a canvas to plot our graphs. We then use fig.set_figure_inches(15, 10) to set the height and width of our “canvas”. That’s all you really need to know for now. And finally, we plot using seaborn and the regplot function, using the “Usage/GM” column as our x-axis and [“FantasyPoints/GM”] as our y-axis. Finally we set scatter = True to tell seaborn we want a scatter plot – and voila, our graph renders.
Below is the code evaluating the correlation between efficiency and fantasy points per game.
As you can see below, efficiency is not nearly as strongly correlated as usage. We expected this, but the comparison of the two graphs confirms how much more important prioritizing usage is if you want to win your fantasy football league.
Let’s break this additional code down briefly.
We are creating a new column again based off other columns, so the syntax is very similar to our previous graph. We define efficiency as TD’s per usage. I admit there are probably better measures of efficiency, but for simplicity’s sake, let’s just roll with this one. I actually decided on this measure because I drafted Aaron Jones and struggled with making the decision of whether to hold on to him or trade him for someone more consistent (I decided to hold on to him). Aaron Jones is the textbook definition of efficient. He’s part of a running back committee, yet somehow was top 5 in touchdowns almost all year.
Anyways, the setup for matplotlib is the same, we use plt.subplots() to create our “canvas”, and seaborn is essentially the same but we are using a different x and y axis. The one main difference here is that we are filtering for minimum rushing attempts. Remember when we filtered based off position to create our separate DataFrames? We are essentially doing the same thing here, but we are using a > operator and setting a minimum of 20 carries based off the RushingAtt column.
The main takeaway from all this should be to draft, pick up, start your running back workhorses. When you see players on the waivers seeing an increasing trend of targets and carries, pick them up. Don’t start running backs until they have proven that they are going to be heavily utilized in the offense (Mike Boone hype train anyone?).
This was a pretty basic analysis and in future posts we’ll dive deeper into using Python to help you win your league next year and crush the 2020 draft. I hope you all enjoyed this and this sends you down a rabbit hole of sports analytics! If you have any questions comment below, I’ll be sure to help out.
Check out part two of the series here.