Forensic Accounting Analysis
Dive Deep into the financial statements to identify if risks lie hidden
Last updated: 20 October 2025 (v2)
Objective:
Perform a forensic accounting examination of a company’s accounts to generate insights beyond traditional income statement trends. Identify and assess risks hidden within accounts and generate monitoring plans.
Explanation:
If you ask professional investors where edge still exists, many will point to the plain-vanilla 10-K (and 10-Q). Not because others don’t read it, but because most readers skim the highlights and miss the connective tissue between footnotes, cash flow idiosyncrasies, and management incentives. Buffet, Munger and other great investors study these documents intensively but its time consuming.
Generative AI massively changes the speed of this work. The risk is that speed without structure turns into superficiality. This prompt lays out a practical, structured way to pair classic financial-statement analysis interpretation that summarizes simply any identified risks. The analysis steps are grounded in the principles from Graham & Dodd’s Security Analysis and Graham’s The Intelligent Investor (quality of earnings, margin of safety), Penman’s Financial Statement Analysis and Security Valuation (clean-surplus and residual income logic), Koller et al.’s Valuation (cash flow over accounting earnings, ROIC vs. growth), White, Sondhi & Fried’s The Analysis and Use of Financial Statements (ties between statements), and Schilit & Perler’s Financial Shenanigans (pattern recognition for red flags). The point isn’t to replace judgment. It’s to let the machine handle extraction, cross-referencing, and first-pass diagnostics so you can concentrate on materiality and causality.
The organizing idea of this prompt is simple: 14 sections, one for each core analysis step that practitioners actually run - accruals quality, cash conversion, revenue recognition tests, capitalization choices, working capital stress tests, leverage and covenant risk, off-balance-sheet items, stock-based compensation and dilution, segment economics, auditor signals, and management discussion & analysis language.
As always, be aware that models can make mistakes. At each step, examine the response and challenge information or conclusions that appear erroneous before proceeding to any subsequent steps. If in doubt use a second model with the same prompt to verify the information and generate challenge questions and answers (CoVe process) to correct interpretations of data.
Link to blog post explanation:
AI vs the 10-K: Forensic Accounting Analysis
Preferred Model:
Gemini 2.5 FLASH. see execution notes
Important Execution Notes:
It is HIGHLY recommended with this complex 14-stage deep analysis task to use Inferential Investor’s Preprocessing and Data Ingest Prompt as the pre-step before this analysis prompt is run. This task needs to extract highly specific financial data from multiple complex 10-K’s and 10-Q’s which stretches a model’s capabilities if done simultaneously with the analysis task itself. The pre-processing, run individually on each document within the same chat window, extracts and brings the necessary data into the context window. The analysis task is then completed (also which the same chat as a follow up prompt) by the model with greater speed, accuracy and insight than otherwise.
The model can do the task without pre-processing with the execution notes below however you may find that it reaches out to alternative sources for the financial data which will need greater verification.
The prompt is by default set for the user to upload the last 2 annual reports (10-K’s) and most recent quarterly report (10-Q) (unless the latest quarter is the year end).
If you find the analysis hangs in the process, it will be due to the complexity of the request and amount of attached information. This usually occurs if you do not pre-process the data first.
This can be solved by 3 methods: (1) using the Gemini FLASH model over PRO, (2) reducing the attachments to a single 10-K (that still has at least 2 years financial information) or 2-3 quarterly reports which are less detailed and/or (3) breaking the prompt down into separate steps (eg. modules 1-7 then 8-14).
This is a longer and more complex analysis exercise using most of a model’s context window and processing capability. Gemini has been shown to handle this better as a single prompt given its encoder-decoder architecture (attends to all information in prompt and attachments prior to answering) and much longer 1m token context window. It will typically provide a more comprehensive response to other models which try and summarize interpretations and analysis to preserve context window.
ChatGPT has struggled with this as a single prompt and is slower than Gemini, often hanging in the response. I have found the need to split this into 2 sections in a CoT conversation using GPT-5 Thinking and run the first 7 steps followed by the last 7 and conclusion. However due to ChatGPT’s autoregressive structure the summarized conclusion up front is best left to the end in order to pull together all the information attended to in the steps.
Insert required info in the square brackets (company name, share price, and optionally any peer companies to compare).
Sample Output:
Copy/Paste Prompt:
Important note: Subscribers can use this prompt for their own analysis. However, the prompt is copyrighted by The Inferential Investor, paywalled, and must not be shared without permission.