How to Maximize Ad Revenue on Twitch?

A question anyone with at least one Affiliate on Twitch asks themselves... or at least some of them.

Hello Dreamers! I bet some of you are Twitch streamers as well. If you can code or are feeling brave, I have the beginnings of some code for you. This phase of the analysis is finding correlations. So, here is some code that is the start of this.

Twitch Analysis

How to run it:

First, you need Python 3 installed on your machine. I recommend 3.10. You will also need the following dependencies:

• Pandas

• Matplotlib

• Seaborn

These dependencies can be downloaded via python-pip3. Conda might also work.

Once the machine is ready, you now need data. The data you need to run this code is readily available from Twitch itself. All you need to do is go to the Creator Dashboard, then Analytics on the left-hand menu. The ones I used were:

• Overview

• Engagement

• Earnings

All you need to do is clone the git and download the Twitch Data, placing the resulting CSV files into the same folder as the scripts. Below is a correlation matrix in the pretty form of a heatmap of all factors in the data. A correlation matrix is simply a chart of the relationship between two quantitive measures in a number from -1 to 1: where a negative number is a negative correlation (one measure goes down as one goes up), a positive number is a positive correlation (if one measure goes up so does the other), and 0 is no correlation whatsoever.

Perhaps it's a little messy for you to see. Even after cleaning up the csvs downloaded straight from Twitch, there were still a lot of elements to look at. The purpose of a correlation matrix is to narrow down the elements to observe and experiment with. After all, imagine optimizing one element of the channel only to find that it doesn't make a difference at all in another measure we wanted to alter. In this case, we want to increase ad revenue, so we want to find measures we can control that will alter this number.

So, I loaded up a bar graph of all the correlations applicable to Ad Revenue. The result:

So, here, it's hard to see, but the highest element correlated appears to be Chat Messages.

The next thing I did was I took the ad revenues part of the correlation matrix and asked only for the top 10 and bottom 10 correlations. The results are rather shocking... at least for me.

Straight from the Python terminal:

                    Positive                     Negative

0              Chat Messages        Unfeatured Clip Views

1                 Live Views                         Date

2                   Chatters                   Clip Views

3             Unique Viewers                 Bits Revenue

4            Minutes Watched                Clips Created

5            Engaged Viewers          Featured Clip Views

6  Returning Engaged Viewers                  Sub Revenue

7            Average Viewers                  Tier 1 subs

8           Minutes Streamed              Total Paid Subs

9        Ad Breaks (Minutes)  Hosts and Raids Viewers (%)

Some of these make a lot of sense. More ad breaks more ads. The higher the average viewer rate the more people watching, and the more ads being watched. If someone is subbed they are paying for the privilege of not having to see a single ad.

However, what I found shocking is the positive, and rather high positive correlation between ad revenue and chat messages. Perhaps if someone is participating in chat they are more likely to watch an ad.

Now, there is the saying: "Correlation does not prove causality." Finding a correlation can narrow variables, but to truly want to hyper-tune, you need to find the mechanism for one causing the other and ensure the numbers are not due to a strange artifact in the data, or simple serendipity. (Like, the date of the year negatively affects ad revenue?)

So, the second looked for correlations for the correlations. For the positive ones:

Stay Dreaming Everyone!