Master Data Analysis with ChatGPT (in just 12 minutes)

0:00Let's get straight to the point. All of

0:01us work with data regardless of our

0:03role. But very few of us were taught how

0:06to analyze data in a structured way. So

0:09in this video, I'll bridge that gap by

0:11sharing a simple three-step framework

0:13that basically turns chatbt into our

0:15personal data analyst with zero

0:17technical skills required. Let's get

0:19started. First, a bit of context. In all

0:21my past roles, management consultant,

0:22account manager, product marketing

0:24manager, I've had to work with data and

0:26present findings. Like most of you

0:28though, I've never received any formal

0:29training in data analysis. But I also

0:31knew AI could help. So after taking the

0:33top AI for data analysis course on

0:35Corsera, I learned that the key was

0:37giving Chachi PT a proven framework to

0:39follow and have the AI perform all the

0:42hard analytical tasks on our behalf.

0:45Diving right in the framework I learned

0:46is called DIG description,

0:48introspection, and goal setting. And in

0:50a nutshell, by using chatbt to apply the

0:52dig framework to any data set, we're

0:54able to number one, understand data

0:55we've never seen before in a matter of

0:57minutes instead of hours. And number

0:59two, extract insights that we as

1:01non-data analysts would have missed.

1:03Here's a simple visualization. When you

1:05get handed a spreadsheet with no

1:07context, you're at 0% understanding. But

1:10with every dig prompt you input into

1:12chatbt, your understanding of the data

1:15increases. And by the end of the dig

1:17framework, you've uncovered insights

1:19that would have taken hours to find

1:21manually, if you found them at all. Two

1:23quick things before diving to a real

1:24case study. First, I'm using a free

1:26Apple TV Plus data set that you can

1:28download and follow along. And it's

1:30actually pretty cool to analyze real

1:31data from popular TV shows and movies

1:33like Avatar the Last Airbender, The

1:35Godfather, and Sherlock. Second, the

1:37industry standard framework is actually

1:39called EDA, exploratory data analysis.

1:42But I'm using dig in this video because

1:43one, the principles are the same. And

1:45second, the professor on Corsera,

1:47probably used dig because it's easier to

1:49remember. Step one, description. Picture

1:52this hypothetical scenario. Your

1:53colleague, let's call him Tim Cookie,

1:55just rage quit for absolutely no reason

1:57and left you with a spreadsheet with

1:59zero context. At this point, we need

2:01ChachiBt to explain or describe what's

2:04in the file as quickly and effectively

2:06as possible. So let's open up ChachiBT,

2:08upload the data set, select the latest

2:11reasoning model, and start with a first

2:13description prompt. List all the columns

2:16in the attached spreadsheet and show me

2:18a sample of data from each column. The

2:20reason we start off with this prompt is

2:22because it forces chache to actually

2:24look at every single column in our data

2:26set and more importantly gives us a

2:28quick overview of the data we're working

2:30with. Looking at this output, I want to

2:32point out two things. First, having

2:33ChachiBT return just one sample output

2:36is much easier for us, the human to

2:38digest versus having to make sense of an

2:41entire spreadsheet we've never seen

2:43before. Second, the sample is selected

2:44as Forest Gump, a classic. Nice. And it

2:47returned all eight columns from the

2:48original spreadsheet. Great. But the

2:50release year is 994.0 and there are two

2:53genres separated by a comma. So, this

2:56might represent issues for Chacht down

2:58the road. So, we want to make a note of

2:59that. And I'm actually not really sure

3:01what IMDb ID means. Does every show in T

3:04does every show in movie have a unique

3:06ID? I'm not sure. So I might want to

3:08follow up and confirm. Next description

3:10prompt number two. Oh, I'll link to all

3:11these prompts down below. By the way, uh

3:13take five more random samples of the

3:15data for each column to make sure you

3:17understand the format and type of

3:18information in each column. So why are

3:21we asking for more samples? Because that

3:23one sample we received might be an

3:24outlier and therefore misleading.

3:27Multiple samples help us spot

3:28inconsistencies. Looking at this output,

3:30we see there are TV and movies under

3:33type, but we knew that already. This TV

3:36show has three genres. Okay, this these

3:38movies have one. All right, and uh some

3:42of this is available in one country,

3:44others are available in multiple

3:45countries. Okay, so our understanding of

3:47the data set is increasing. Moving on to

3:50description prompt number three. Run a

3:52data quality check on each column.

3:54specifically look for missing or empty

3:56values, unexpected formats or data

3:59types, outliers or suspicious values.

4:02This is pretty self-explanatory. We want

4:03Chhatabt to explicitly tell us if

4:06there's anything weird about the data we

4:07should know about before proceeding with

4:09our analysis. Okay, there's several

4:11tables here. This first one tells us how

4:13many values are missing from each

4:14column. So, for the title column, we're

4:16missing 589 values uh representing 3.1%.

4:20Going down, 10% is pretty high. Whoa.

4:22Okay. 99.7. This number tells us that

4:26for the available countries column,

4:28we're missing 99.7%

4:31of the values. I can double check really

4:33quickly by going into the raw data set

4:35and doing a quick filter or sort rather.

4:38Going to sort this and I scroll down.

4:40Yeah, most of these rows are completely

4:44empty for the available countries

4:45column. This means we should not perform

4:48any geographical analysis with this data

4:51set because we're missing that

4:52information. At this point, it should be

4:54pretty clear that first, although Chachi

4:56PT is not doing 100% of work for us,

4:58it's making our job as a human analyst

5:01much easier. Second, remember how the

5:03goal of the description step is for us

5:05to understand the data set as

5:08effectively as possible. I'm not going

5:10to waste your time here, but in real

5:12life, I would have at this point asked

5:14follow-up questions like, "Hey, what

5:16does this TT number mean here?" And

5:19chatbt would have confirmed it is the

5:22unique IMDb number. By the way, if you

5:24use Google Workspace tools at work, you

5:26might want to join my newsletter to

5:28receive an insanely actionable tip every

5:30week. Link down below. Next up, we have

5:32introspection. And the purpose of this

5:34step is to have Chat JBT brainstorm

5:36questions it could answer with our data.

5:38This shows whether ChachiPT truly gets

5:41our data and often services insights we

5:44hadn't considered. Prompt one, tell me

5:4610 interesting questions we could answer

5:48with this data set and explain why each

5:50would be valuable. And long story short,

5:53good questions mean CatchPT understands

5:55our data. Bad questions equal there's a

5:57misunderstanding that needs fixing

5:59before we proceed. These first three

6:02questions are solid. How has Apple TV's

6:04yearly output grown since launch? If

6:06we're putting out more TV shows and

6:08movies yearon year, it might mean we're

6:09capturing more market share. Uh, what

6:12share of releases are movies versus

6:13series each year? This might tell us

6:15about viewer behavior. Are we trending

6:17more towards TV shows or movies? I love

6:19this one. Which genres dominate the

6:21catalog and how have they shifted over

6:22time? Imagine you were on the Apple

6:24content team, right? You might want to

6:26invest more in the most popular genre

6:28next year. Or maybe additional analysis

6:31tells you the genre is oversaturated, so

6:33you want to pull back. Prompt two. For

6:35the first three questions, tell me

6:37exactly which columns you need to use

6:39and whether the current data is

6:40sufficient to answer it. This basically

6:43forces Chachbt to show its work and

6:45tells us whether or not we can perform

6:47these analyses. We see that for question

6:49one, yes, we just need to fix 0.3% of

6:53non-numeric entries. We can ignore that.

6:55For question two, yes, we need to do

6:57some light data cleanup. That's fine. We

6:59can tell chatbt to do it. For number

7:01three, yes, we have all the information

7:02we need to perform the analysis.

7:04Awesome. Prompt three is my personal

7:06favorite for the introspection step.

7:07What questions do you think someone

7:08would want to ask about this data but we

7:11can't answer due to missing information.

7:13This basically surfaces gaps in our data

7:16set and helps us manage our boss's

7:18expectations about what insights we can

7:21uncover. Here we see questions like

7:22what's the most watched genre? We can't

7:24answer that because we don't have the

7:25viewing metrics or from an ROI

7:28perspective which genres deliver the

7:30best cost per hour of content. We don't

7:32have the production budget, revenue, or

7:34cost fields. But here's where it gets

7:36interesting. What if I had access to

7:38some of that data? Let's say my friends

7:39at Apple invited me into Apple Part and

7:41I hacked their servers. That was

7:43obviously a joke. Very unrealistic

7:45scenario. I don't really have friends.

7:48So, I created a fake second data set.

7:51And to be very clear, this is made up.

7:52Don't report me to Apple. uh with the

7:54IMDb ID in column A, total viewership in

7:57column B, and the total cost of

7:59producing that show or movie in column

8:01C. I can now upload this onto the same

8:03ChachiPT thread and say, I just received

8:06this data set from a colleague. Your

8:07task is to explore and explain the

8:09relationships between this new data set

8:11with the original one and how they might

8:14be used to join the data together. After

8:16running for a bit, ChachiP confirms we

8:18can use the IMDb ID field to join the

8:21two data sets together and even gives us

8:24suggestions on how we can use this newly

8:26merged table. For example, we can

8:28calculate the cost per viewer ROI by

8:32genre. For example, after instructing

8:34Chacht to merge the data sets using the

8:36IMDb ID field, it gives us a sample

8:40output of this newly merged spreadsheet.

8:43Right? We see that for Forest Gump, if I

8:45scroll all the way to the right, yes, it

8:47now has the total viewership and total

8:49cost data in its row along with

8:53everything else. And I can even click

8:55here to download this merge CSV file.

8:59That's pretty awesome. Quick note, I'm

9:00focusing on the core dig prompts in this

9:02video to not waste your time. In real

9:05life, I would have branched out much

9:07sooner. For example, earlier when

9:09Chachib mentioned genre popularity, I'd

9:12immediately ask for that analysis and

9:14keep digging based on what it find. The

9:16third step, goal setting. This is

9:18extremely important to get right because

9:20imagine if our manager asked us to

9:22analyze the sales data and after working

9:24hard to create 20 beautiful slides, our

9:26manager says, "Wait, I just wanted to

9:28know if we should discontinue product

9:30X." This is what happens when our

9:32manager is an idiot. I mean uh when we

9:34analyze data without setting clear

9:35goals, we have something that's

9:37technically correct but ultimately

9:39useless. Obviously, the prompts we use

9:41in this step depend on the specific

9:42goal. So, I'll just share one example.

9:44Here's the prompt. My goal is to

9:46understand and you specify your goal

9:47here. I want to understand what content

9:49Apple TV should invest in next. Given

9:52this goal, which aspects of the data

9:54should we focus on? This is basically

9:56like giving CatchBT a mission briefing.

9:58It helps the AI prioritize what's

10:00important and ignore what's not. Okay,

10:03this is very useful. So, Traded D first

10:05breaks down our options for us. For

10:07example, if you're in the Apple content

10:09team, you might care about viewership,

10:11audience demand, and content supply,

10:12right? So, you would want to do this. If

10:14you're in the finance team, you might

10:16want to learn more about unit economics

10:18and do that. Scrolling down, we even see

10:20a step-by-step road map. So, first we

10:22might want to clean our data. Okay.

10:24Then, we build a genre scorecard. That

10:26could be very interesting. Then we rank

10:28our opportunities, layer in trend

10:30velocity. I would have never thought to

10:32do that. And finally, stress test with

10:34outliers. Makes sense. And here's the

10:36type of insight this process would

10:38surface. Uh, true crime series deliver

10:40three times the medium views of all

10:42series. They cost 18% less per finished

10:44hour and have climbed from 4% to 9%

10:47share of total watch time in the last 3

10:49years. Okay. Wow, that's that's really

10:51impressive. Pro tip. A final question I

10:53always like to ask Chad PT before any

10:55presentation is what are the key

10:57questions someone reading my analysis

10:59would ask and how should we proactively

11:02address them. This prompt

11:03single-handedly saved my ass multiple

11:05times by anticipating but Jeff what

11:08about this questions from managers and

11:11overly ambitious peers trying to put me

11:12down. Just kidding. Everyone loves me.

11:14How could they not? Two things I'd like

11:15to leave you with. First, the dig

11:17framework plus chajbt levels the playing

11:20field for regular untrained people like

11:22us. It's a simple repeatable process we

11:25can all use immediately. Second,

11:27although I cover the essentials today,

11:29the full Corsera course touches on other

11:31important concepts like how to mitigate

11:33hallucinations and debug weird data

11:35errors. So, if you want to level up your

11:37data skills, sign up for Corsera using

11:38the link in the description to take

11:40advantage of my special offer of 40% off

11:42for 3 months of Corsera Plus. If you

11:44enjoyed this, check out my comprehensive

11:46Chachi PT pro tips video next. See you

11:49all there. And in the meantime, have a

11:51great one.