How to analyse financial statements with ChatGPT

Financial analysis is the case where the format of the upload changes the answer, not just the convenience of getting there. The same balance sheet delivered two ways produces two different levels of reliability, and the difference is not obvious from reading the reply.

The annual report is the worst version of the data

A published annual report is a design object. The statements inside are typeset for a printed page, which means multi-column layouts, figures stated in thousands with the unit declared once in a header several pages earlier, negatives shown in parentheses rather than with a minus sign, and footnote markers stuck to the end of numbers.

Text extraction flattens all of that. What comes out is a stream of digits where the column a figure belonged to is a matter of inference. A model reading 1,234 1,109 has to work out which year is which from surrounding context, and it will make a confident guess either way.

The charts do not survive either. text only, except Enterprise That covers the revenue bridge, the segment pie chart, and any table that was pasted into the report as a picture.

Spreadsheets get a different set of limits

Documents and spreadsheets fail in opposite directions, which matters when a filing runs to two hundred pages.

A long report uploads cleanly and then gets read in part. Nothing tells you where it stopped. A spreadsheet either goes through whole or is refused outright, which is the failure you want because you can see it.

Getting from report to CSV

1. Look for the data file first

   Listed companies usually publish an Excel or XBRL version of the financials alongside the PDF. Regulators' filing sites often expose the same statements as structured data. That file is the one you want, and it saves the entire transcription step.

2. One file per statement

   Income statement, balance sheet and cash flow as three separate CSVs. Mixing them into one sheet makes every question ambiguous about which section it refers to.

3. Normalise the units and the signs

   Put the actual numbers in the cells. If the report is in thousands, either multiply through or add a column stating the unit on every row. Replace parenthesised negatives with real minus signs, and strip currency symbols and thousands separators.

4. Label the periods explicitly

   A header row of FY2024, FY2023, FY2022 removes the guesswork. If periods run across columns, say so in your first message.

It reasons about numbers, it does not always calculate them

This is the part that catches people out. With a spreadsheet attached, ChatGPT frequently writes and runs code to produce an answer, and that arithmetic is real arithmetic. When it is working from prose, or from numbers you pasted into the chat, it often produces the answer the same way it produces a sentence, by predicting what should come next. Both replies look identical.

Gross margin of 42.3% derived from a Python calculation and gross margin of 42.3% derived from pattern recognition are indistinguishable in the output. The second one is right most of the time, which is precisely what makes it dangerous.

Three habits fix this. Ask it to show the calculation, not just the result. Ask directly whether it computed the figure or estimated it. And spot check two or three numbers against the source with a calculator, choosing ones where an error would change your conclusion.

What it is genuinely good at

Reading the notes. This is where the useful information in a set of accounts actually lives, and it is prose, which plays to the model's strengths. Ask it to summarise the revenue recognition policy, list every related party transaction, or flag where an accounting estimate changed between years.

Drafting the ratio set. Given clean data it will produce liquidity, gearing, margin and turnover ratios across every period at once, which is tedious to build by hand.

Explaining a movement. "Operating cash flow fell 18% while net income rose. Which working capital line drove that." It will point at the right row and describe the mechanism.

Comparing two companies on structure rather than score. Different revenue recognition, different capitalisation policies, different segment definitions. It reads both sets of notes and tells you what makes the comparison unfair.

Where the PDF still belongs

Upload both files, for different jobs. The CSV is for anything numeric. The PDF, or the relevant extracted sections of it, is for the narrative: management commentary, the notes, the auditor's report, the risk factors.

Keep an eye on how many uploads that costs. The rolling allowance is 80 files every 3 hours, and a full comparative analysis of three companies gets through it faster than expected. Consolidate where you can rather than uploading each year separately.

The practical decision

Numbers go in as CSV, narrative goes in as text, and you verify the arithmetic before it leaves your desk. The transcription work is the price of a reliable answer, and skipping it does not save time so much as move the error somewhere you will not spot it. Treat the output as a fast, well informed first pass by someone who has not double checked their sums.