<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Vedika Sethi on Medium]]></title>
        <description><![CDATA[Stories by Vedika Sethi on Medium]]></description>
        <link>https://medium.com/@sethivedika.2304?source=rss-3e9b3f42712f------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*h-6ZtE37meu7gVaeC16xbw.jpeg</url>
            <title>Stories by Vedika Sethi on Medium</title>
            <link>https://medium.com/@sethivedika.2304?source=rss-3e9b3f42712f------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 08:07:22 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@sethivedika.2304/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[AI Can Find Financial Data in Seconds, But Can it Verify it?]]></title>
            <link>https://medium.com/@sethivedika.2304/ai-financial-analysts-can-they-catch-the-errors-that-humans-miss-14c9d8d66e45?source=rss-3e9b3f42712f------2</link>
            <guid isPermaLink="false">https://medium.com/p/14c9d8d66e45</guid>
            <dc:creator><![CDATA[Vedika Sethi]]></dc:creator>
            <pubDate>Sat, 03 Oct 2026 17:34:53 GMT</pubDate>
            <atom:updated>2026-11-01T05:54:18.793Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>Exploring the gap between financial data retrieval and verification, and why it matters for AI-powered financial analysis.</blockquote><p>In 2012, a mistake in a risk model helped contribute to a $6.2 billion loss at JPMorgan Chase. The incident became known as the “London Whale” scandal. The loss did not begin with a cyberattack or an unexpected market event. It began when an error in a risk model (a tool used to estimate the chance and impact of a future bad event) was supposed to divide a number by the average of two other numbers. Instead, someone set it up to divide by their sum.</p><p>That mistake likely made the bank’s reported risk look far smaller than it really was. The error went undetected for months while the bank continued adding to a risky trade. Later investigations showed that the formula error was part of a bigger failure involving manual spreadsheet changes, weak controls, and poor review.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qL8PW-XQ_TM-xKtn.png" /><figcaption>What Happened During the London Whale Scandal</figcaption></figure><p>The incident raised a question I wanted to investigate: if AI can retrieve financial information in seconds, why is verifying that information still so difficult?</p><h3><strong>Why This Actually Matters</strong></h3><p>Financial decisions depend on numbers. Investors use them to evaluate companies, banks use them to assess risk, and analysts use them to build financial models. But finding a number is different from knowing whether it is appropriate for the analysis.</p><p>If I ask, “What was Apple’s revenue last quarter?” I still need to know:</p><ul><li><strong>Time period:</strong> Which quarter does the number cover?</li><li><strong>Metric:</strong> Is it total company revenue or revenue from one business segment?</li><li><strong>Accounting basis:</strong> Is it reported under standard accounting rules or adjusted by the company?</li><li><strong>Units and currency:</strong> Is the figure in dollars, millions, or billions?</li><li><strong>Company scope:</strong> Does it cover the entire company or only part of it?</li><li><strong>Version and consistency:</strong> Was the number later corrected, and why might different sources report different figures?</li></ul><p>These are six problems that an AI financial verification tool could help detect. But comparing numbers is not as simple as checking whether they match. Two different figures might both be correct if they cover different periods or measure different things.</p><p>A correction also does not automatically mean the original number was useless. The original figure shows what was reported at the time, while a restated figure reflects a later correction. Which one matters depends on the question being asked.</p><p>Real incidents show why these distinctions matter. In 2003, Canadian energy company <strong>TransAlta lost approximately C$24 million after a spreadsheet copy-and-paste error </strong>misaligned bids with the intended contracts. The mistake went undetected before submission, and the bids could not be reversed.</p><p>My original assumption was that the hardest part of financial research was finding the number. The more I researched, the more I realized that the real challenge is making sure the number still means the same thing when it moves from one document or system to another. That is where AI could help. Not just by finding information, but by checking it against evidence.</p><h3><strong>What Does an Analyst Actually Do?</strong></h3><p>Big public companies, like Apple and Tesla, are legally required to publish detailed financial reports. The yearly report is called a 10-K, and the quarterly report is called a 10-Q. These reports are filed with the U.S. Securities and Exchange Commission, or SEC, which is the government agency responsible for regulating U.S. securities markets.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/862/0*dG3wy6VHbN4O1Ue2.png" /><figcaption>Example of what a 10-K report looks like</figcaption></figure><p>The SEC publishes the reports through a website called <strong>EDGAR</strong>. The filings include financial statements, management explanations, footnotes, risk disclosures, and accounting information. Companies also report financial data using <strong>XBRL (eXtensible Business Reporting Language)</strong>, a structured labeling system that helps computers understand financial information.</p><p>Think of XBRL as a label attached to each financial number. For example, a revenue figure isn’t just a number like $416 billion. Its XBRL data can identify it as revenue, specify the reporting period it covers, and indicate that the amount is reported in U.S. dollars. This context matters because two numbers might look similar but represent different things, such as revenue for three months versus revenue for the entire year.</p><p>XBRL makes financial information easier for software to find, organize, and compare across reports. However, it doesn’t guarantee that every number is correct. A figure could be tagged incorrectly, or two sources could report different numbers for legitimate reasons. This is why checking the meaning and accuracy of financial data is just as important as extracting it. A junior analyst’s job basically looks like this:</p><ol><li>Find the company’s official reports.</li><li>Copy or extract numbers such as revenue, expenses, cash, and debt.</li><li>Put those numbers into Excel or another financial system.</li><li>Build a model that predicts the future.</li><li>Have someone more senior review the work.</li><li>Present it to whoever is making the decision.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/826/1*rEkYg3KupZJABSsrmafdaQ.png" /></figure><p>A financial model is a spreadsheet or software system that uses historical data and assumptions to estimate future performance. It might calculate growth, profit margins, cash flow, valuation, or risk.</p><p>Step five is usually the main checking stage. But “checking” often means looking over the model and seeing whether the output feels reasonable. It does not always mean tracing every number back to the exact page and row where it came from.</p><p>This is what makes financial errors so sneaky. The spreadsheet can look completely normal. The formulas can calculate. The final number can seem believable. A small mistake near the beginning can quietly affect everything that comes after it.</p><h3><strong>Where Errors Enter the Workflow</strong></h3><p>Financial errors can enter a company’s workflow at several stages, from recording information to extracting, calculating, and interpreting it. The challenge is that these errors are not always obvious. A spreadsheet can produce a result even when its formula is wrong, and two financial reports can show different figures without either being inaccurate. This makes financial verification more complicated than simply checking whether two numbers match.</p><p>One source of errors is data extraction. Financial information is often stored in PDFs, tables, spreadsheets, and footnotes, which can make it difficult for AI systems to retrieve the correct figures. Optical character recognition (OCR), for example, can misread numbers, miss negative signs, or separate figures from the columns they belong to. Even if a number is extracted correctly, the system might assign it to the wrong quarter or financial metric.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/646/1*q2pXBtGvmg8JuSyU0rMi5A.png" /><figcaption>Distribution of cell error rates across spreadsheets</figcaption></figure><p>Errors can also occur in spreadsheets through incorrect formulas, missing rows, double-counted values, or data entry mistakes. Research by Raymond Panko has found that spreadsheet errors are common, with one frequently cited study reporting errors in 94% of the spreadsheets examined. These mistakes can be difficult to catch because a spreadsheet can calculate normally while still producing an incorrect result. AI could help identify unusual calculations, but it needs to understand what the formula is supposed to calculate before it can determine whether something is wrong.</p><p>Another challenge is understanding the context behind financial figures. A company might report quarterly revenue, annual revenue, or revenue from a specific business segment. It might also report both GAAP and non-GAAP figures. GAAP (Generally Accepted Accounting Principles) follows standard accounting rules, while non-GAAP figures adjust those results by excluding or changing certain items. Both figures can be accurate, but they should not automatically be compared. If an AI system retrieves a number without identifying its metric, reporting period, or accounting basis, it could compare figures that represent different things.</p><p>Version differences create another problem. Companies can amend filings or restate previously reported financial results, meaning an older figure may no longer be accurate. Retrieval-Augmented Generation (RAG) allows AI systems to retrieve information from external documents, but it does not guarantee that the system has selected the most recent document or the correct figure. It still needs to check publication dates, amendments, and reporting periods.</p><p>These errors can also overlap. An AI system might extract the wrong number, interpret it incorrectly, and compare it with an outdated figure. This is why financial verification requires more than retrieving information or checking whether two numbers match. AI needs to understand where a figure came from, what it represents, and whether it is valid to compare with another figure.</p><h3>Companies Building in This Field</h3><p>Several companies already use AI to work with financial documents, but they solve different problems.</p><h3>AlphaSense</h3><p><a href="https://www.alpha-sense.com/about/">AlphaSense</a> helps financial professionals search across company filings, earnings calls, research reports, and news. Its AI-powered search can find relevant information even when different sources use different wording. However, finding a relevant passage does not automatically verify that a financial figure is correct or comparable with another figure.</p><h3>Daloopa</h3><p><a href="https://daloopa.com/how-our-ai-works">Daloopa</a> extracts financial data from company disclosures and links figures to their original sources. This helps analysts collect historical numbers for financial models without manually searching through every report.</p><p>For example, an analyst researching Apple could retrieve its quarterly revenue figures and trace them back to company disclosures. This makes the data easier to review and verify. However, a source link alone does not establish that the selected figure is appropriate for a specific comparison. The analyst still needs to check the reporting period, metric, units, and accounting definition.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/0*x36rvkxqnjD2E1Xn.gif" /><figcaption>Daloopa’s Product</figcaption></figure><h3>DataSnipper</h3><p><a href="https://www.datasnipper.com/">DataSnipper</a> helps auditors extract information from documents and connect financial figures to supporting evidence. It can reduce manual work when checking records and financial statements. Its focus is primarily on audit workflows, rather than specifically comparing a company’s reported revenue across its SEC filing and earnings release.</p><h3>Koncile</h3><p><a href="https://www.koncile.ai/en">Koncile</a> uses AI to extract structured information from documents, reconstruct tables, and compare data across records. It is relevant to my idea because it can help identify inconsistencies between extracted values. The question I want to explore is whether a specialized tool could go further by explaining why two financial figures differ and determining whether they actually measure the same thing.</p><h3>BlackLine</h3><p><a href="https://www.blackline.com/products/financial-close/">BlackLine</a> provides software for accounting and financial operations, including account reconciliation. Its tools help finance teams match records, investigate discrepancies, and improve the accuracy of financial reporting processes. This addresses an important part of verification, although its focus is broader accounting workflows rather than the specific cross-checking of public-company disclosures I want to investigate.</p><h3>The Gap</h3><p>I think the biggest opportunity is making financial AI more reliable when comparing information across sources. Even advanced systems can retrieve the wrong reporting period or select the wrong financial metric. Recent benchmark testing has shown that AI agents can perform well at retrieving financial numbers but still make mistakes, highlighting the challenges of using AI without human oversight. Research in this area is still ongoing, with researchers exploring ways to improve financial data extraction, accuracy, and reasoning. I want to explore whether a focused tool could go a step further: compare the same financial figure across official sources, check that the figures are genuinely comparable, and explain any mismatch with evidence.</p><h3><strong>Retrieval Versus Verification</strong></h3><p>Retrieval asks, “Where is the information?” Verification asks, “Is this the right information for this purpose, and can I prove it?” If I ask an AI, “What was Apple’s revenue last quarter?” it might return a number and a source link. But verifying that answer requires checking the fiscal quarter, whether the figure represents total revenue, its units and currency, and whether the filing was later amended. A number can come from a legitimate source and still be wrong for the question being asked. The goal is not just to find a number, but to understand its context and have evidence to support it.</p><p>The same distinction applies when two sources report different figures. If one document says revenue was $25.4 billion and another says $26.1 billion, a basic program can flag the mismatch, but it cannot necessarily explain it. The difference might reflect different reporting periods, rounding, accounting definitions, or an extraction error. This is the difference between discrepancy detection and discrepancy resolution: identifying that numbers differ is only the first step; understanding why they differ is what helps an analyst decide what to trust. This changed how I thought about my project. Rather than simply building a tool that compares two numbers, I want to explore one that checks whether they are comparable, explains any discrepancy using source evidence, and acknowledges when it cannot confidently determine the cause.</p><h3><strong>What Research Shows</strong></h3><p>When people say an AI model is accurate, the first question should be: accurate at what? Finding a number, extracting it correctly, checking mathematical consistency, and explaining a financial discrepancy are different skills. A model might perform well at one but struggle with another, so general accuracy scores do not necessarily tell us whether it can reliably verify financial information.</p><p>Research benchmarks highlight these limitations. <strong>FinVerBench</strong> tested whether AI could identify mathematical inconsistencies in financial statements from 43 S&amp;P 500 companies, including statements modified to contain specific errors. Under one testing setup, nine of fourteen model runs produced false-positive rates of 95–100% on clean statements, meaning they frequently flagged correct statements as incorrect. The study also found that rounding made errors harder to detect. Meanwhile, <strong>FinED-Bench </strong>evaluated error detection in long financial documents, including incorrect dates, missing figures, unit errors, contradictions, and financial reasoning. Its best-performing model achieved an overall F1 score of 48.34%, with weaker results on financial-reasoning errors. F1 combines precision — how often flagged errors are real — and recall — how many actual errors the system finds.</p><p><strong>Fin-RATE</strong> examined financial-analysis tasks involving SEC filings, including reasoning within disclosures, comparing companies, and tracking information across reporting periods. Together, these benchmarks suggest that financial verification involves more than extracting a number or finding a source. An AI system must also select the correct figure, understand its context, distinguish genuine errors from harmless differences, and explain its conclusions accurately. This is the specific challenge I want my project to investigate.</p><h3><strong>How Current Verification Can Miss Errors</strong></h3><p>Citations help users trace an AI answer back to its source, but they do not guarantee that the answer is correct. A filing might contain the cited number, yet that number could refer to a different quarter, business segment, unit, or accounting definition. Similarly, comparing an SEC filing with an earnings release does not automatically prove accuracy. Both sources might repeat an outdated figure or use definitions that are inappropriate for the analysis.</p><p>Disagreements between sources also do not necessarily mean one is wrong. Figures can differ because they cover different periods, reflect an acquisition, or use updated information following a restatement. A basic system can flag a mismatch, but explaining it requires understanding the context behind each figure. Human review has limitations too: analysts may check whether a spreadsheet looks reasonable without having time to trace every number back to its original source.</p><p>My goal is not to replace human review, but to make it more targeted. The system I want to explore would flag financial figures with conflicting sources, missing units, unclear definitions, outdated versions, or insufficient evidence. It would then explain the discrepancy where possible, helping analysts focus their attention on the numbers that need closer investigation.</p><h3><strong>My Idea</strong></h3><p>Building a system that verifies every kind of financial data would be too ambitious for a first project. Instead, I want to build an <strong>AI financial cross-checker</strong> that helps analysts verify financial figures by comparing them across official sources. For my first version, I would focus on one company and one metric: quarterly revenue reported in its SEC filing and earnings release.</p><p>The process would begin by identifying the company, fiscal quarter, revenue metric, accounting basis, currency, and relevant document versions. The tool would retrieve both sources, extract the figures, and record their context and exact locations in the documents. Ordinary code would handle precise tasks such as unit conversions and date comparisons, while AI would help interpret footnotes and understand differences in how the figures are reported.</p><p>If the figures matched, the tool would check whether they actually referred to the same metric, period, and company scope. If they differed, it would investigate possible explanations, including rounding, different accounting definitions, restatements, or extraction errors. It would then produce a report showing both figures, the checks performed, the likely explanation, and the evidence supporting its conclusion. If the available evidence were insufficient, it would flag the discrepancy as unresolved instead of guessing. An analyst could review the findings and decide which figure was appropriate for their analysis.</p><p>Ultimately, I want to build more than a program that identifies mismatched numbers. I want to explore whether AI can help explain <em>why</em> financial figures differ, trace conclusions back to reliable evidence, and make the verification process faster and more transparent. I would start by testing the tool on several companies to discover which discrepancies occur in practice, how reliably they can be explained, and where human review remains necessary.</p><h3>What This Has Taught Me</h3><p>I began by thinking the main challenge was finding financial information quickly. My research showed me that the harder problem is preserving its context and proving that a number is appropriate for the decision being made. Two figures can both be correct yet mean different things.</p><h3>The Question I’m Carrying Forward</h3><p>Can I build an AI tool that does more than find financial data? One that verifies its context, explains discrepancies with evidence, and knows when it cannot reach a reliable conclusion?</p><h3>Sources</h3><p>Society of Actuaries. (2016). <a href="https://www.soa.org/globalassets/assets/Library/Newsletters/The-Modeling-Platform/2016/april/mp-2016-iss3-campbell.pdf">https://www.soa.org/globalassets/assets/Library/Newsletters/The-Modeling-Platform/2016/april/mp-2016-iss3-campbell.pdf</a></p><p>Raymond Panko. (2008). <a href="https://mba.tuck.dartmouth.edu/spreadsheet/product_pubs_files/Errors.pdf">https://mba.tuck.dartmouth.edu/spreadsheet/product_pubs_files/Errors.pdf</a></p><p>U.S. Securities and Exchange Commission. (2024). <a href="https://www.sec.gov/xbrl">https://www.sec.gov/xbrl</a></p><p>U.S. Securities and Exchange Commission. (2021). <a href="https://www.sec.gov/search-filings/edgar-search-assistance/accessing-edgar-data">https://www.sec.gov/search-filings/edgar-search-assistance/accessing-edgar-data</a></p><p>U.S. Securities and Exchange Commission. (2016). <a href="https://www.sec.gov/data-research/structured-data/inline-xbrl">https://www.sec.gov/data-research/structured-data/inline-xbrl</a></p><p>U.S. Securities and Exchange Commission Investor.gov. (n.d.). <a href="https://www.investor.gov/introduction-investing/general-resources/news-alerts/alerts-bulletins/investor-bulletins/how-read">https://www.investor.gov/introduction-investing/general-resources/news-alerts/alerts-bulletins/investor-bulletins/how-read</a></p><p>U.S. Securities and Exchange Commission Investor.gov. (n.d.). <a href="https://www.investor.gov/introduction-investing/getting-started/researching-investments/using-edgar-research-investments">https://www.investor.gov/introduction-investing/getting-started/researching-investments/using-edgar-research-investments</a></p><p>FinVerBench. (2026). <a href="https://arxiv.org/abs/2605.29586">https://arxiv.org/abs/2605.29586</a></p><p>FinED-Bench. (2026). <a href="https://aclanthology.org/2026.findings-acl.1481.pdf">https://aclanthology.org/2026.findings-acl.1481.pdf</a></p><p>Fin-RATE. (2026). <a href="https://arxiv.org/abs/2602.07294">https://arxiv.org/abs/2602.07294</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=14c9d8d66e45" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>