For Actuarial Analysts ·
What you'll accomplish
Checking a development exhibit cell by cell is slow, and a transposed figure is easy to miss on a second read. This guide shows how to give ChatGPT a small aggregated triangle, ask it to recompute the age-to-age ratios, and have it list any ratio that differs from the one in your sheet. Your workbook stays the record. The recomputation is a second pair of eyes, and you re-check every flagged cell yourself.
What you'll need
Data caution. Even an aggregated triangle is confidential company data. Whether you may upload one depends on your company's AI policy and on whether ChatGPT is an approved tool for it. Check your plan's data controls for whether chats and files are used for model training and how long they are kept. If in doubt, use an invented triangle with the same shape, which is what this guide does. Never upload policy-level rows, claim numbers or anything tied to a person.
OpenAI's help article recommends structured data with clear column names and one record per row. A triangle fits that if you list it in long form. Create a CSV file named invented_triangle.csv with this content. All values are invented cumulative paid losses.
accident_year_code,age_months,cumulative_paid_invented
AY-A,12,1000
AY-A,24,1250
AY-A,36,1500
AY-A,48,1575
AY-B,12,1200
AY-B,24,1560
AY-B,36,1872
AY-C,12,1400
AY-C,24,1750
AY-D,12,1600
What you should see: Ten rows plus the header. Four accident year codes, with AY-A having four ages, AY-B three, AY-C two and AY-D one. Troubleshooting: If your spreadsheet program adds extra columns, delete them. Save as CSV, not as a workbook with several tabs.
Work out the ratios by hand before you upload. Each is the later value divided by the earlier value, rounded to 3 decimals.
| Accident year | 12 to 24 | 24 to 36 | 36 to 48 |
|---|---|---|---|
| AY-A | 1.250 | 1.200 | 1.050 |
| AY-B | 1.300 | 1.200 | |
| AY-C | 1.250 |
That is six ratios. Now imagine your own sheet shows 1.520 for AY-C from 12 to 24, a transposed 1.250. The check should find exactly that one.
What you should see: Your hand-worked table matching the one above. Troubleshooting: If a ratio does not match, recheck the division. AY-B 12 to 24 is 1560 divided by 1200, which is 1.300.
invented_triangle.csv.What you should see: The file name shown as an attachment in the message box before you send. Troubleshooting: If the attach control or CSV files are not offered, the options depend on the plan and workspace. Ask your IT contact, or paste the ten rows directly into the message instead.
Paste the prompt below. It supplies your sheet's ratios and asks only for a comparison.
The attached file holds an invented cumulative paid triangle in long form.
For each accident year, recompute the age-to-age ratio as the later value divided by the earlier value, rounded to 3 decimals.
Then compare with the ratios in my sheet, listed below, and list every ratio that differs, showing both values.
Do not select factors, average anything, or suggest a tail. Only recompute and compare.
My sheet:
AY-A 12 to 24: 1.250
AY-A 24 to 36: 1.200
AY-A 36 to 48: 1.050
AY-B 12 to 24: 1.300
AY-B 24 to 36: 1.200
AY-C 12 to 24: 1.520
What you should see: A table of six recomputed ratios and a short list with one difference: AY-C, 12 to 24, sheet 1.520 against recomputed 1.250. Troubleshooting: If ChatGPT lists no differences or several, do not trust either answer. Compare its table to your hand-worked table from Step 2. If its numbers are wrong, say so in the chat and ask it to show the calculation for the cell, then verify by hand.
What you should see: For AY-C, 1750 divided by 1400 is 1.250. The sheet cell holds 1.520, so it needs correcting. Troubleshooting: If your sheet has a formula, not a typed number, the slip is in a reference. Click the cell and check which two cells it divides.
The check confirms arithmetic only. It says nothing about whether a development pattern is reasonable. Selecting factors, averaging and a tail belong to you and the reviewing actuary, in your workbook.
Row and column count check
Report how many accident years and how many ages are in the attached file, and whether any accident year has a gap in its ages.
Reasonableness flag without judgment
List any ratio below 1.000 in the recomputed table, without commenting on whether it is acceptable.
Explain a difference
For the flagged cell, show the two values you divided and the result, so I can check them by hand.
Reformat for the sheet
Show the recomputed ratios as a wide table with accident years as rows and ages as columns, so I can compare it to my sheet by eye.