Chris Low

Machine Learning Case Study: Telco Churn Prediction

One of the biggest challenges Telco companies face is their customers terminating their contract for various reasons (or known as "churn"), such as getting a better package or experience from the other companies.

Therefore, it is essential for a Telco company to investigate the underlying reasons of their customers churning. By doing so, the company will then be able to focus on the real problem, improve it, and ultimately reducing their churn rate.

Defining problem statements

To tackle the challenge aforementioned, I have defined the following problem statements that a Telco company may be interested in finding out:

1. What are the reasons that would make customers churn?
2. What kind of customers are more likely to churn in the future?

I utilized the framework by LEAD throughout the entire case study, called the OSEMN Framework, which comrpises of: Obtain, Scrub, Explore, Model and iNterpret

Let's get our hand dirty

In this case study, this dataset from Kaggle has been used as the source of data.


Upon inspecting the dateset, this company has a churn rate of 26%. According to Statista, the average churn rate for the Telecom industry is roughly 21%. This means that this company has a higher-than-average churn rate as compared to its competitors.

Let's dive deeper into the possible reasons that contribute to this high churn rate. There are indeed some interesting reasons that one group of customers are more likely than others to churn.

Exploratory data analysis

From the bar plots above, we can clearly see that customers having a monthly contract are more likely to churn as compared to their counterparts with a yearly or longer contract.

Not only that, more than half of the customers who use electronic check as their main payment method for the service are more likely to churn as compared to those who pay via mailed check, bank transfer or credit card.

This finding could possibly indicate the following events:

• When it comes to monthly contract, this telco's service may not be as good as its competitors, giving the customers no choice but to terminate their contract after a period of time.
• There could be a system failure to receive electronic check and this seems to require immediate attention, as more than half of the customers who choose this payment method actually churned.

Another interesting finding is that the subscription for anything related to the customers' phone security seem to play an essential role on the Telco's churn rate.

Apparently, customers who subscribe to the Telco's online security, online backup, device protection, and tech support are less likely to churn as compared to their counterparts. This brings forward another possible implication:

• Perhaps the Telco can consider including the phone security services as part of their packages, or even making it compulsory? They could really use some extra protection. 😀

Only part of the heatmap is extracted here to provide a clearer explanation. Full heatmap can be found in the source code.

Last but not least, let's interpret a little on the correlation heatmap above. Other than the few factors that we have discussed above, it's also interesting to note the correlation between customers who select Fiber optic as their internet service and their churn rate. It seems like people who subscribed to this Telco company's Fiber optic are far more likely to churn than those who subscribed to DSL service, or completely without internet service.

Therefore, it brings forth the final implication from the EDA:

• There could be some technical issues with the Fiber optic internet service that cause the customers to churn ultimately after a period of usage. It is worth discovering it to minimize the damage in the future.

Machine learning modeling: Who is more likely to churn?

Now that we have got our EDA done, it's time to build a model to predict who is more likely to churn.


From the correlation heatmap, I have chosen the following factors as my model features:

1. Tenure
2. Internet Service - Fiber Optic
3. Month-to-month Contract
4. Monthly Charges
5. Total Charges
6. Paperless Billing

The dataset is split into 2 groups, where 70% of them is the training data, and 30% of them is the test data.

Upon getting our data ready, it is then used to fit into our model, decision tree, with a depth of 3 levels for basic prediction. This is the result:


Finally, let's test the accuracy score of this model.


The accuracy score of this decision tree model is close to 80%. It seems to be a rather good accuracy.

According to this decision tree, customers who are on a monthly contract and are currently subscribing to fiber optic as their internet service, are very likely to churn. What increases the odd of them churning even more is that their tenure with the Telco is less than 15.5 months.

Conclusion

Coming back to our 2 problem statements defined in the beginning:

1. What are the reasons that would make customers churn?

There are a couple plausible reasons that could affect the Telco's churn rate that are worth investigating:

• Dissatisfied service experience provided to monthly-contract subscribers
• Potential technical issues on electronic check payment
• Incompetent fiber optic package

2. What kind of customers are more likely to churn in the future?

This answer to this is somewhat similar to the reasons mentioned above. Customers who are currently on a monthly contract and are currently subscribing to fiber optic are most likely to churn. If their tenure is less than 15.5 months, the odds of them churning increases by another 50%.


Tools / libraries used: Python [pandas, seaborn, matplotlib, numpy, sklearn(tree)]

source code

Linear Regression: Social Media Ads Forecasting

Marketing is an essential part of a business, especially those who rely on advertisements done digitally, via social media such as Facebook and Google. In this project, I attempt to answer the following business question:

• How would you allocate a marketing budget of $100,000 among the three social media platforms that we are using right now, based on our past performance?

Tools / libraries used: Python [pandas, seaborn, matplotlib, sklearn(linear regression)]

source code

The Elon Musk Effect

If you have been on the news for the past new months, you would have known a little about the rise and fall (and rise again recently) of the cryptocurrency Bitcoin. If you are in a little bit deeper, you would have heard how Elon Musk single-handedly "influenced" the cryptocurrency market, especially Bitcoin and Dogecoin.

The climax happened when Elon Musk announced that Telsa will be accepting Bitcoin as one of their payment methods back in March 24 and then later reversed his position in May 13.

You know how big the news is when your wife who didn't care about the market, suddenly asks: "How come Elon Musk so powerful one ah?" 😂

There are several articles written about the Elon Musk Effect and also how his Twitter activity has moved the market in an unexpected way.

Measuring the Elon Musk Effect

Musk's legendary brand impact on his audience often starts with a tweet. The Tesla CEO has recently shown his support toward different companies and projects on Twitter, instantly piquing the interest of his 47 million followers. Those endorsements range from Etsy to Gamestop (paywall) to Signal and, perhaps more surprisingly, Bitcoin.

Elon Musk being one of my personal idols before he gets his foot into the crypto market, this topic really interests me. Does he really have that big of an impact with just a few words on a social media platform?

Inspired by this awesome Kaggler who measured the return on Dogecoin after Elon Musk's tweets, I have decided to find out the same thing about Bitcoin.

So I have this one question:

Is the Elon Musk Effect real and will I be able to replicate the result?

Exploring the datasets

I have used two separate datasets in this project:

1. Elon Musk's Tweets from 2010 to March 24, 2021
2. Bitcoin historical daily price chart from Sep 17, 2014 to July 9, 2021

First, let's have a glance at Bitcoin's historical price chart.


It seems like Bitcoin received most of the attention starting from this year that its price shot up from $20,000 to its peak at $63,000. Throughout these years, it also experienced a heavy crash of more than 70% on 2018, as well as a 50% crash this year.

Moving on to the dataset that contains tweets from Elon Musk.

Because I'm interested in only his tweets that are related to Bitcoin, I have filtered the dataset with some keywords like "Bitcoin" and "BTC".


Do take note this filter only works on tweets that mention Bitcoin specifically on text, and this includes tweets from himself, and also tweets that he replied to some other users. Tweets that contain images that are relevant to Bitcoin are not included in our analysis.


There are a total of 21 instances where he mentioned Bitcoin throughout the years.

Apparently, Elon Musk started talking about Bitcoin on Twitter in November 2017. He then mentioned Bitcoin twice in 2018 and 2019, and subsequently talking about it more and more from 2020 onwards.

And out of those 21 tweets, 5 of them have negative sentiments, and the rest of them are either positive or neutral sentiments.

Exploratory Data Analysis

Now, it's time to put the datasets together and have a closer look.

I started off by merging the two datasets with a common column called "Date", which is the date of Elon Musk's tweet posting (I'll call it the Tweet Day), and the closing price of Bitcoin on that day.

After putting them together, I then calculated the difference between the price of Bitcoin in the next 5 days and the price of Bitcoin on the Tweet Day. This will be the day-to-day return after the Tweet Day. The tweets are be grouped together if Elon Musk tweeted them on the same day.

Last but not least, a line chart of the return in percentage after the Tweet Day is plotted.


The darker the line, the more recent the tweet is posted.

Based on line chart, we can see that most of Elon Musk's recent tweets have garnered a positive return right after his Tweet Day. Some of the returns even go as far up as 15% in just one day after he tweeted about Bitcoin.

Another box plot also has also been employed to further visualize the effect of Elon Musk's tweets.


As we can see in this plot, the median return up to 5 days after the Tweet day, has been a positive one, hanging around 2-4%. Even though there are intances of the return being negative, but the majority of the data are indicating an above average of return after Elon Musk's tweets.

What this means is that, if you own Bitcoin before Elon Musk tweets about it, you have a rather fair chance of seeing your Bitcoin value goes up by a few percent.

Disclaimer: The above references an opinion and is for information purposes only. It is not intended to be any financial advice.

Yes, the Elon Musk Effect is real

This case study certainly echoes the rest, indicating that Elon Musk has the potential of "moving" a market, simply by tweeting a few words about it.

It's actually a pretty scary thought that a single person who you may or may not like, could affect the overall value of your portfolio.

Moving forward

Having said that, there are many improvements that can be done onto this project:

• Instead of using a daily price chart, we can zoom into the minute price chart to track the immediate effect of his tweets right after he posted it.
• Use other techniques to include relevant images (including profile picture) as well, other than text-only tweets.
• Utilize a more recent dataset.
• Relate the tweet sentiments to the return.

Closing thoughts

In reality, Elon Musk is not the only one who has such power. In fact, there are tons of people, like celebrities and influencers, who can easily influence others to perform a trade, thereby moving the market price.

What's so scary about it is that this is a common phenomenon nowadays. We have victims everyday being manipulated by others and are losing money because of that.

Like it or not, it is important for us to get equipped with sufficient financial knowledge to avoid falling prey to those who have malicious intent. Do your own research before getting into any kind of investment vehicle.

Tools / libraries used: Python [pandas, seaborn, matplotlib, nltk(sentiment intensity analyzer)]

source code

Sentiment Analysis: Hotel's Review on Tripadvisor

Being in the hotel / homestay industry myself in the past, I know how important it is to always keep an eye on the reviews, as it could literally make or break the brand. This goal of this project is to analyze a hotel's reviews listed on TripAdvisor, and measure their sentiments.

Tools / libraries used: Python [pandas, seaborn, matplotlib, wordcloud, nltk(sentiment intensity analyzer)]

source code

HR Attrition Dashboard on Google Data Studio

A dashboard is a great information management tool that is used to track and display a series of data that interests a particular group of stakeholders. Depending on the the stakeholders' objectives, different dashboard may show different indicators, so that the party of interest can take action based on the data.

In this case study, I have built a dashboard for a stakeholder - the HR manager who has some particular busines questions in mind, regarding to the company's attrition rate.

Tackling the business quetions

So the imaginary HR manager has these hypothetical questions that she would like to find out the answers from the dashboard:

1. How satisfied are the employees in the company right now?
2. Does experience and also number of projects assigned affect an employee's attrition?
3. Which department has the highest attrition rate?

Bearing in mind these questions that she had, I have then proceeded to build a dashboard catering to her needs, so that she can actively track the relevant indicator to get her questions answered.

What's next?

Now that we have a happy HR manager, we can have more happy decision makers / stakeholders too, from different departments.

Ultimately, all departments should look at their own dashboard that answers their own business questions, instead of sharing one single dashboard with one another. This would then improve the business as a whole!

Simple tool, but with great effect!

Tools used: Excel spreadsheet, Google Data Studio

View Dashboard

Decision Tree on Risk of Conorary Heart Disease

There are many factors one could have contracted a Conorary Heart Disease (CHD). In this project, by using the Framingham Health Dataset, I attempt to predict who will most likely contract CHD by using a decision tree.

Tools / libraries used: Python [pandas, seaborn, sklearn(tree)]

source code