Data Analysis By ChatGPT

by

For over a decade, I’ve been dreaming of an analytics platform that would ingest my Excel file, process it, and give me actionable insights – without any effort from me. My XLS file could have raw data on newsletter subscribers, sales pipeline, interviews of job applicants, and so on. If you think it sounds like magic, you wouldn’t be wrong. But that was my fantasy. And, until recently, it remained a fantasy.

In 2017, I was thrilled to learn that Gartner created a new subcategory of analytics called Smart Data Discovery, which “automatically finds, visualizes and narrates important findings such as correlations, exceptions, clusters, links and predictions in data that are relevant to users without requiring them to build models or write algorithms” (italics mine). It seemed like Smart Data Discovery would fulfill my dream.

But it didn’t. In the following months, I tried a dozen products including Excel Pivot Table, DataRobot, and a few SDD tools mentioned by Gartner.

None of them worked as advertised. After I uploaded my raw data, all of them asked me to define the fields in it. I thought that was dumb since this info was available in the header row of the XLS file I uploaded.

Over time, I got frustrated with my experience with these products, and added my dream platform to my laundry list of products that I should try and build myself when I had the bandwidth. I gave it the working name of MINER360 and spec’ced it as follows:

BACKGROUND: Virtually every business has some data, mostly in the form of Excel spreadsheets. There could be a considerable amount of insight locked in the data. To extract it, the company may use Excel Pivot Table or a more sophisticated BI software like COGNOS, DOMO, Tableau, etc. Some of them are too complex and the others, too expensive for the mainstream market. Besides, all of them require manual steps to describe the data and define the nature of insight expected. As Gartner notes in its latest BI and Analytics Platforms Magic Quadrant report, “current visual-based data discovery approach has accelerated data harmonization and the visual identification of patterns in data, as opposed to the previous IT-centric semantic-layer-based approach, (but) the tasks for creating insights are still largely manual and prone to bias.”
PAIN: Delays and difficulties in extracting insight from data lead to lost business opportunities.
SOLUTION: A platform where the user simply uploads their XLS file and presses a button. The platform automatically performs analytics and outputs actionable insights. Whether it should use data mining or predictive analytics or some other sophisticated analytics technique is decided by the platform user transparently based on the nature of data uploaded.

I didn’t find any such product for years.

That changed when ChatGPT launched 4o version in May 2024. You can find my first impressions of ChatGPT-4o here. Since then, 4o has come a long way. In the last quarter or two, I’ve had ChatGPT analyze many datasets and received tons of actionable insights.

Let me share a few examples of the data, analysis, and the insights.

1. Marketing Campaign Strategy

I uploaded an XLS file containing sales pipeline data, and entered the following prompt:

You’re a marketing analytics expert. We’re asking you to complete an analysis of the customers represented in this file. Please summarize the main insights we should take action on.

I got a pretty good response from ChatGPT, which is shown in the following exhibit.

In the above prompt, I followed the best practice of telling ChatGPT to assume the persona that’s fit-for-purpose for the given context, “marketing analytics expert” in this case.

2. ETF Composition 

I received a spreadsheet from HDFC AMC with the list of holdings in its NIFTY 50 ETF. I uploaded this XLS file and asked ChatGPT if it contained any actionable insights.

ChatGPT analyzed the scrips, sectors and holdings, and gave me a bunch of insights. Some of them were obvious since an ETF, by definition, must reflect the stocks making up the index, NIFTY 50 in this case. But there were others that were news to me. You can find Chat’s full response here.

3. Is G7 Obsolete?

When I observed that the US economy was vibrant whereas the economies of Germany, UK, and some other G7 countries were moribund, I wondered if the notion of “West” was no longer economically meaningful. I decided to do a deep dive into this. Towards that end, I prompted ChatGPT as follows:

Plot the contribution of the GDP of USA and “G7 Minus USA” to Global GDP for the last 10 years.

As you can see, I asked ChatGPT to also source the data by itself, not just do the analysis.

While you find Chat’s full reponse here, the chart I got from it is given in the following exhibit.

According to this chart, the dominance of USA is rising whereas the position of the other western nations is falling. This confirmed my suspicion that G7 is no longer the monolithic economic powerhouse that it has been projected as so far.

4. ChatGPT And The Art Of Motorcar Maintenance

Here’s an interesting case where tweeple @devaiahPB uploaded a car repair estimate to ChatGPT and sought its help to negotiate a better price with his auto mechanic:


tl;dr: ChatGPT helped @devaiahPB to reduce his repair cost by 14K.

As you can see in the above examples, ChatGPT carries out data analysis without asking any questions. This contrasts sharply with the various analytics tools that I’d tried before. While they touted to be business intelligence software, they betrayed their dumbness by asking fairly basic questions that a human with above room temperature IQ would not have. Another remarkable thing is that ChatGPT cleanses the data if required whereas the other tools expected me to carry out Extraction Transformation Loading (ETL) and feed them pristine data.

Once you start using ChatGPT for data analysis, you’ll find yourself reaching its rate limit quite quickly. Either you can upgrade to the paid version of ChatGPT or use it as an assistant to help you to do data analysis in ways that I described in my blog post titled Using ChatGPT as a Data Analysis Assistant.

Having gone through seven examples of data analysis by / with ChatGPT, you might have a few questions about hallucination, data sources, data quality, and accuracy of analysis. All valid questions.

I’ll try to answer them in a follow-on post. Watch this space!