For Economist (Private Sector / Consulting)s ·
What you'll accomplish
You rebuild growth, index and real-series columns often enough that a slipped reference is a matter of when, not if. A second set of eyes on the arithmetic is cheap: ChatGPT's data analysis can recompute the columns from your raw series, list the cells that disagree with its recomputation and draw a quick chart for a draft.
The workbook stays the record. ChatGPT points at cells to look at, and you open each one, read the formula and decide what is wrong. This guide uses an invented workbook called series-practice.xlsx with two planted errors, so you can see how the check works before you point it at a real series.
What you'll need
Data caution. Upload only public or synthetic files here. Client data, documents produced in litigation, anything under a protective order or NDA, and non-public company data stay in your firm's approved systems. If a check needs real client figures, run it in your own spreadsheet or in your firm's environment, because code that touches confidential data should never run inside a chatbot's code tool. Before you upload anything non-public, confirm that your firm and the engagement (and counsel, on litigation matters) approve ChatGPT for it. Then check ChatGPT's data controls: whether your chats and uploaded files are used to improve models, and how long they are kept.
Build the file in Excel or Google Sheets and save it as series-practice.xlsx. Every value is invented.
Year, in B1 type Series A (invented, nominal), and in C1 type Price index (invented, 2023 = 100).| Year | Series A (invented, nominal) | Price index (invented, 2023 = 100) |
|---|---|---|
| 2021 | 180 | 96 |
| 2022 | 190 | 98 |
| 2023 | 200 | 100 |
| 2024 | 210 | 102 |
| 2025 | 220.5 | 105 |
Growth, in E1 type Rebased (2021 = 100), and in F1 type Real.=B3/B2-1. In D4 type =B4/B2-1. In D5 type =B5/B4-1. In D6 type =B6/B5-1. Leave D2 empty.=B2/$B$2*100 and fill down to E6.=B2/(C2/100) and fill down to F5. In F6 type =B6*(C6/100).What you should see: Two cells carry planted errors. D4 divides 2023 by 2021 instead of by 2022. F6 multiplies by the index instead of dividing by it. Troubleshooting: Format column D as a percentage with two decimals, and columns E and F with two decimals. The tables below assume that.
Before ChatGPT sees the file, record what the sheet should show. This is the check on the check.
| Year | Growth (D) | Rebased 2021 = 100 (E) | Real (F) |
|---|---|---|---|
| 2021 | blank | 100.00 | 187.50 |
| 2022 | about 5.56 percent | 105.56 | 193.88 |
| 2023 | about 5.26 percent | 111.11 | 200.00 |
| 2024 | 5.00 percent | 116.67 | 205.88 |
| 2025 | 5.00 percent | 122.50 | 210.00 |
As built in Step 1, the sheet shows about 11.11 percent in D4 (200 divided by 180, less 1) and 231.525 in F6 (220.5 times 1.05). Those are the two cells ChatGPT should help you find.
What you should see: A correct-answer table in your notes and two wrong values in the workbook. Troubleshooting: If your sheet shows a different wrong value, one of the formulas in Step 1 was typed differently. Retype the formulas as written.
series-practice.xlsx.What you should see: The file name shown above or inside the message box before you send. Troubleshooting: OpenAI's help pages describe data analysis on spreadsheet files such as .xls, .xlsx and .csv, and recommend descriptive column headers, one record per row and no empty rows or columns splitting the data. If the upload fails, try the .csv version, and check OpenAI's file upload pages for current limits.
Send the prompt in "Your First Real Task" below. It asks ChatGPT to recompute from columns A to C only, so the answer does not depend on the formulas you want to test.
What you should see: ChatGPT works for a moment and returns tables of growth, rebased and real values, usually with a short note on the method. Troubleshooting: If ChatGPT does not run any analysis and answers from the file preview alone, send: "Run the calculation on the uploaded file with the data analysis tool and show the values for every year."
What you should see: A list that names D4 (about 11.11 percent in the sheet against about 5.26 percent recomputed) and F6 (231.525 in the sheet against 210.00 recomputed). Troubleshooting: ChatGPT may see only stored values, not formulas. Treat its list as a pointer. Open D4 and F6 yourself and read the formula in the formula bar.
series-practice.xlsx.=B4/B3-1.=B6/(C6/100).What you should see: D4 at about 5.26 percent and F6 at 210.00, with every other cell unchanged. Troubleshooting: If a value is off in the second decimal, check rounding before you look for an error. The table above rounds to two decimals.
What you should see: A chart with two lines. The nominal line ends at 220.5 and the real line ends at 210.00. Troubleshooting: If the chart uses a different year range or a different label, ask once for the change. Redraw it in your own tools for any deliverable.
What you should see: A workbook where every number traces to the raw columns and your formulas. Troubleshooting: If ChatGPT and the sheet disagree and you cannot see why, recompute that one cell by hand. The sheet and your reading of the formula decide, not the chat.
Recompute a derived column from raw columns
Using only columns [letters] of the attached file, recompute [growth rate, index, real value] for each row and show a table. List every row where the stored value in column [letter] differs from your result, with both values. Do not edit the file.
Describe the pattern in a series
Using only the attached file, describe in plain terms how [Series A] moves over the years, referring to specific rows. State no value that is not in the file, and do not say why it moves.
Chart for a draft
Make a line chart of [column] and [column] by year from the attached file. Title it "[title] (invented data)", label both axes, and list the plotted values underneath so I can check them.
Spot-check a formula in words
Here is a formula and the headers of the columns it refers to: [formula]. Explain in plain words what it computes, and name one case (such as a blank cell or a missing year) where it could return a misleading value.