Education & Research Methodology

AI Stock Market
Analysis Toolkit

Use AI to research, screen, and evaluate stocks like a professional analyst — without sacrificing accuracy. A practical, checkable workflow built on evidence, not guesswork.

Divy Thakkar

Co-Guide: Nirmal Rabari

Evidence-first AI research workflows for retail investors & finance beginners

Table of Contents

Before You Begin

Preface

I built this toolkit out of a frustration that most retail investors quietly share: AI can summarize anything in seconds, but speed without verification is just confident guessing. I wanted a workflow where every claim an AI tool gave me could be traced back to a real filing, a real transcript, or a real number I checked myself.

This book is the result — five chapters that take you from setting up an evidence-first research system, through reading earnings calls and balance sheets without hallucination risk, to building defensible valuation scenarios and reading sentiment without getting swept up by noise.

This work would not be the same without Nirmal Rabari, who served as co-guide throughout its development. His insistence on practical, checkable frameworks — rather than abstract AI theory — shaped almost every method in this book, from the Evidence-First Pipeline to the TRUTH-GUARD Protocol.

One thing this book deliberately does not do: it does not tell you which stock or fund to buy. Every example, case study, and walkthrough here is built around a methodology you can apply to any company or fund you're already curious about. The goal is to make you faster and more rigorous — not to replace your own judgment, or a licensed financial advisor's.

Divy Thakkar
Author

Part 1: Building Your Research System

AI Stock Research Workflow Setup

Ever paste a company's numbers into a spreadsheet, ask an AI to "summarize the quarter," and then realize you can't tell which part of the summary came from the filing versus the model's guess? That gap creates two problems at once: you waste time chasing inconsistencies, and you lose confidence in the result. The fix is not "use better prompts." The fix is a workflow that forces evidence to travel with every claim.

This chapter sets up a repeatable, end-to-end workflow that turns raw filings and market data into analyst-style research outputs, using a named method called The Evidence-First Pipeline. You will build a system that collects inputs from the right places, extracts evidence in a controlled format, generates your write-up, and verifies outputs back against primary sources before you act.

Disclosure & Risk Note

This is an education and research-methodology book, not investment advice. AI tools can produce factual errors. Always cross-verify any extracted numbers, quotes, and conclusions against primary sources like exchange filings (NSE/BSE), SEC filings, and company investor relations documents before making any investment decision.

The Evidence-First Pipeline

Every statement you keep in your final research output must point to a specific input snippet you extracted from the raw source. You will not ask the AI to "summarize" in one shot and hope for the best. You will split the job into evidence extraction first, then interpretation second.

1. Collect Inputs With a Source Folder Structure

Create folders like 01_Filings, 02_Quotes, 03_MarketData, and 04_Notes. Save the primary document you will trust into 01_Filings, transcripts and press releases into 02_Quotes, and price/volume exports into 03_MarketData. This forces you to separate "raw truth" from "your later AI notes," so every claim can be traced back to a file.

2. Extract Evidence Into a Structured Evidence Table

Use prompts that ask the AI to return a table with columns like Claim, Evidence text, Document name, Page/section, Date, and Confidence. Do not let the AI paraphrase without quoting the exact line it used.

3. Generate Your Output Only From the Evidence Table

Build your write-up from the table, not the original filing again. Ask the AI to write sections such as "Business snapshot," "Financial trajectory," "Balance sheet risks," and "Open questions." If it cannot find evidence, it must say "not found."

4. Verify Every Key Number Back to the Filing

Pick a short set of "must-check" items — total debt, cash, revenue, EPS — and cross-verify them by checking the same row in the original document. This is where you prevent AI hallucinations from entering your research.

Case Study: Riya's First Watchlist

Riya is a finance student building her first watchlist. She downloads a quarterly result and asks an AI to "summarize the quarter," but can't tell whether a risk statement came from the filing or the model's general knowledge. She fixes it by running the Evidence-First Pipeline the same way every time: set up a stock workspace, save the filing, build an Evidence Table with quotes, draft the write-up only from that table, cross-verify 3–6 must-check numbers, and lock the final files in 04_Notes.

Quick Checklist

  • Put the filing PDF into 01_Filings and name it with the date.
  • Extract evidence into an Evidence Table with quotes and document locations.
  • Generate your analyst-style write-up using only the Evidence Table.
  • Manually cross-verify the key numbers in the original PDF.
  • Save both the Evidence Table and final output in 04_Notes for reuse.

Mistakes and Edge Cases That Break Evidence

Mistake 1: One-Shot Summarization

Asking for a full summary in a single prompt invites the model to fill gaps with general finance knowledge. Do this instead: extract evidence first into a table, then generate the write-up from that table.

Mistake 2: Evidence Without Location

Claims without a page/section reference are hard to verify later. Always require Document name and Page/section in the Evidence Table output.

Mistake 3: Mixing Market Data and Filing Claims

Price moves and fundamental facts behave differently. Label your output sections clearly: "Filing evidence" versus "Market data."

Closing Thought

Once you set up The Evidence-First Pipeline, your AI research stops being a one-off task and becomes a repeatable process — turning "AI wrote it" into "the filing supports it."

Part 2: Reading Between the Lines

Earnings Call Summaries With Prompts

You finish reading an earnings call transcript and you feel two things at once: the company clearly signaled what it wants investors to focus on, but the transcript also buries the real changes in dozens of "and then" sentences. That's exactly where LLM prompt templates help — you stop treating the transcript like a novel and start treating it like structured evidence.

Key Insight

You get better results when you force the model to extract "what was said" into a checklist you can verify, instead of asking it to "summarize" in free-form prose.

The CALL-to-Checklist Framework

"CALL" stands for Comments, Assumptions, Limits, and Language — then you convert those outputs into a "Checklist" you can compare quarter to quarter.

  1. Lock the extraction to exact transcript quotes — never let the model infer beyond what was said.
  2. Separate "guidance" from "commentary" and "risk." Guidance sounds like targets or ranges; commentary explains drivers; risk language uses words like "could," "may," "subject to."
  3. Force the model to list assumptions, then mark them as Assumption or Unknown in transcript.
  4. Add limits: "If it's not in the transcript, don't guess."

A Practical Walkthrough

Dev, 34, reads one earnings call transcript per week. He wants a repeatable extraction he can audit, not a long research session.

Prompt Template: CALL-to-Checklist Extraction

1. Extract items into three buckets: Guidance, Management Commentary, and Risks/Uncertainties.
2. For every item include a short quote, who said it, and the topic.
3. If no specific number appears, write "Unknown in transcript."
4. List Assumptions separately and label each as Assumption or Unknown in transcript.
5. Add a Limits section listing what you refused to guess.

Prompt Template: Earnings Call Checklist Builder

Using the extracted items, create a Guidance checklist, Commentary checklist, and Risk checklist — each item as a checkbox with quote and topic. Then add a "What changed vs last quarter" list and "Follow-up questions for next earnings call," only including items explicitly stated or tied to unknowns.

Before vs After

OutputBefore (free-form summary)After (CALL-to-Checklist)
Guidance"They sounded optimistic."Checkbox list with short quotes and topics
Risks"There were some uncertainties."Risk items with quoted language and unknowns
Numbers"They expect growth." (no audit trail)"Unknown in transcript" where numbers never appeared
UsabilityHard to compare quartersEasy to compare, items stay categorized

Troubleshooting Common Failure Modes

The model "summarizes" instead of extracting

Fix: require quote-anchoring, force the three buckets, and add a "refuse to guess" limit.

The model invents numbers or implied ranges

Fix: forbid estimation, push assumptions into an explicit labeled list, and move impact math outside the model using only verified filing values.

Risks mix with commentary

Fix: add strict tagging rules so guidance, risk, and commentary each follow their own wording pattern.

What "Good" Looks Like

Dev ends with a checklist of quoted guidance, a separate commentary list, and an auditable risk list — plus "Unknown in transcript" markers showing exactly where he must verify manually.

Part 3: Spotting Trouble Early

Balance Sheet Red-Flag Detection

"Debt doesn't always break a company. Sudden debt does." Red flags rarely announce themselves with a single scary figure — they show up as changes: debt spikes, promoter pledging, and related-party transactions that look ordinary on paper but can hide real risk. This chapter teaches you to catch those warning signs fast using the RED-FLAG Radar method.

Important Disclaimer

This is an education and research-methodology book, not investment advice. AI tools can misread tables or hallucinate details. Always cross-verify outputs against exchange filings (NSE/BSE, SEC, or the company's official reports) before you act.

Three High-Risk Areas to Watch

  • Debt spikes: sudden increases in borrowings and shifts between short-term/long-term or secured/unsecured.
  • Promoter pledging: when promoter shares get pledged or the pledge increases, pressuring control and liquidity.
  • Related-party transactions: transactions with entities tied to the company that can move cash, revenue, or costs in ways that don't reflect true performance.

Running the RED-FLAG Radar Checklist

Step 1: Build Your Evidence Table

Pull total borrowings, the short-term/long-term split, promoter pledge details, and related-party notes from the latest quarterly report and the prior comparable period. Record the labels exactly as the company prints them.

Prompt Idea

"Scan this balance sheet extract and notes text. Extract lines related to total borrowings, short-term and long-term borrowings, promoter pledging, and related-party transactions. Output a RED-FLAG Radar table with columns: Item, Period, Value, Change vs prior period, Evidence snippet (exact text), Risk interpretation to verify."

Step 2 & 3: Structure Findings and Run the Three Rules

RuleWhat You Check
Debt SpikeBorrowings jump sharply and concentrate in one category (e.g., mostly short-term); new borrowing without clear cash-use explanation.
Promoter PledgingPledged shares increase, or notes show pledge-related actions tied to ongoing debt.
Related-PartyTransaction value grows quickly or shifts toward loans/guarantees rather than normal services.

Case Study: Ananya's Forensic Check

Ananya, 29, reviews a company where total borrowings rise from 100 to 160, mostly in short-term debt; pledged shares increase; and related-party purchases and outstanding balances rise together. She marks each item "Verify in notes" and ends with a pass/fail map: Debt spike — Flag, Promoter pledging — Flag, Related-party transactions — Flag. Her outcome isn't a buy/sell call — it's a clean red-flag map with evidence snippets she can read in minutes.

Common Mistakes

Trusting AI conclusions without evidence snippets

Require an "Evidence snippet (exact text)" for every flagged item, then verify it against the PDF.

Checking only total debt, not the debt shape

A company can keep total debt flat while the risk profile worsens through short-term concentration — always record both splits.

Mixing related-party transactions with normal business

Not every related-party line is a red flag. Track transaction type and direction of change, then check whether the notes explain why.

Closing Thought

AI-assisted red-flag detection works best when you treat the filing as the source of truth and AI as your checklist engine — forcing the same evidence structure every time.

Part 4: From Filing to Multiple

Valuation Modeling With Verified Inputs

A valuation model breaks the moment you feed it "clean-looking" numbers that came from somewhere else. One wrong revenue line, one mismatched share count, or one mistaken debt figure can flip your multiple. Most valuation mistakes come from assumption drift: you change the story you tell the model without updating the source facts.

The ASSUMPTION Ledger

A table you maintain where you list every assumption and input your model uses, attach the source, record the exact figure you extracted, and mark whether you verified it. When you run scenarios, you change only the assumptions you explicitly allow to vary; everything else stays locked.

Before You Run Any Numbers

This is education and research-methodology, not investment advice. If you cannot verify an input, treat it as untrusted and exclude it from your final valuation view.

Step-by-Step: Build Multiples From Verified Inputs

  1. Pull the primary statement set — income statement, balance sheet, and share count basis — from NSE/BSE or SEC filings.
  2. Create the ASSUMPTION Ledger before asking AI for outputs, with columns for Input name, Value, Period basis, Accounting basis, Source file, Verification status, and Notes.
  3. Extract exact figures yourself; let AI echo them back and flag period-basis mismatches.
  4. Ask AI to compute multiples constrained strictly to your ledger values — never let it swap in a different earnings definition.
  5. Run scenarios by changing only one assumption at a time, locking everything else as "Verified."
  6. Reconcile AI results back to primary data signals using a checklist tied to each ledger row.

Constraint Example

"Use only these ledger inputs: Earnings (TTM) from row X, Share count basis from row Y, and Market price as of date Z. Do not use any other numbers." Then ask for the computed multiple, intermediate steps, and a reconciliation note.

Workflow StepWhat You DoWhat AI Does
Ledger setupCreate ASSUMPTION Ledger rowsFormats layout, checks missing fields
Number extractionExtract exact figures from filingsEchoes values, flags period-basis conflicts
Multiple computationProvide locked inputsCalculates the multiple, shows intermediate math
Scenario changesUpdate only allowed assumptionsRecomputes outputs, highlights changed drivers

Case Study: Rahul's Weekend Valuation

Rahul, 41, builds a P/E ledger with Earnings (TTM), Share count basis, and Market price — each marked "Verified" only after extraction from the filing. AI computes the multiple from locked inputs. He then runs base, upside, and downside scenarios by changing only the earnings-growth assumption, writing each one into the ledger and tying it back to guidance language in the filing. He ends his weekend with one base-case multiple, two scenario multiples, and a ledger trail showing exactly which rows drove the result.

Lessons Learned

  • Lock the inputs first, then let AI compute. AI often produces a "reasonable" number based on what it thinks you meant.
  • Reconciliation beats confidence. Check that earnings definition, period basis, and share count basis all match.
  • Change one assumption at a time and record it in the ledger, so you learn which driver actually moves the multiple.
Part 5: Signal Without Noise

Sentiment Signals Without Hallucinations

A single headline can move a stock fast, but it can also mislead you fast. If you use AI for sentiment, you need more than "summarize the news." You need a way to separate useful signals from confident-sounding errors — this is what the TRUTH-GUARD Protocol is built for.

The TRUTH-GUARD Protocol

For every sentiment claim you find, you require a direct quote or timestamped source, a cross-check against the company's primary documents, and a reality check against fundamentals you already track. You rank claims by verifiable strength rather than averaging everything into one "score."

Run Sentiment Through TRUTH-GUARD Every Time

  1. Collect claims, not opinions. Convert shared headlines and social themes into specific claims like "company raised guidance," not vague tags like "bullish."
  2. Extract evidence from the text. Keep timestamps and direct quotes for each claim.
  3. Cross-check against primary sources. Verify against transcripts, press releases, or exchange filings the same day. Untraceable claims get tagged unverified.
  4. Map verified claims to fundamentals. Guidance changes map to revenue/margin outlook; contract news maps to backlog or visibility.
  5. Score confidence with guardrails: Verified, Plausible, or Unverified — and weight claims accordingly.
  6. Decide an action rule. Only update your thesis when at least one Verified claim changes a fundamental driver's direction.

Case Study: Meera's Earnings-Week Run

Meera, 26, runs TRUTH-GUARD every evening for three days around a release date. Day 0: she collects 8–12 specific claims. Day 1: she extracts exact quotes from the transcript and press release, flagging "not found" where evidence is missing. She cross-checks every social claim against primary sources — unverified ones don't influence her thesis. Day 2: she maps verified claims to fundamentals, assigns confidence tiers, and locks a thesis update built only on Verified claims, with a "what would break my thesis" note attached.

Common Sentiment Errors and Fixes

Treating AI summaries as proof

Require a direct quote from a primary source for every claim you want to use, with its date attached.

Mixing verified and unverified claims

Use confidence tiers per claim. Only Verified claims should update your thesis or target direction.

Mapping sentiment to the wrong fundamental driver

Create a one-to-one link between each verified claim and the fundamental driver it should affect; if you can't link them, downgrade the claim's impact.

Final Thoughts

By the end, you won't just "get answers" — you'll run a repeatable research pipeline where AI drafts the work and you validate the evidence. Your first action: within the next 30 minutes, pick one stock you're watching, complete Chapter 1's workflow setup, generate an earnings-call summary using the Chapter 2 prompt, and immediately cross-check the figures.

Research Guardrails — Read This Before You Use AI for Sentiment

This book is an education and research-methodology guide, not investment advice. Always cross-verify critical information and any AI-extracted claims against primary sources such as exchange filings (NSE/BSE) and SEC filings before you make any investment decisions.

Worked Example: Applying TRUTH-GUARD to a Watchlist Stock

This book is methodology, not stock-picking — so instead of naming a "buy," here's how you'd run the full process on any company already on your watchlist, using a fictional example, "Company X":

Claim collected: "Company X raised full-year guidance after strong Q2 demand."
Evidence check: You find the exact guidance sentence in the earnings call transcript and the press release — tag: Verified.
Fundamental link: You check whether revenue growth and order commentary in the filing actually support "strong demand" — tag: Consistent.
Decision rule applied: Because a Verified claim changed a fundamental driver (forward revenue), this is the one signal allowed to update your thesis — everything else stays noise.

Run this same five-step process on a real company you're tracking — your local broker app, screener, or mutual fund factsheet — and let your own verified evidence (not a tip from anyone, including this book) drive the conclusion.

TRUTH-GUARD Pre-Trade Checklist

Before you let any sentiment signal influence a decision, run through this five-point check:

  • Does this claim have a direct quote or timestamped source attached?
  • Has it been cross-checked against a primary filing, transcript, or press release?
  • Is it tagged Verified, Plausible, or Unverified — and not left ambiguous?
  • Does it map to a specific fundamental driver (revenue, margin, backlog, regulation)?
  • Would removing this single claim change your conclusion? If not, it isn't decision-relevant.

Bonus Prompt: One-Line Sentiment Scorecard

"For each claim in my Verified list, output one line in this format: [Claim] → [Fundamental driver] → [Direction: positive/negative/neutral] → [Confidence: Verified/Plausible]. Do not include Unverified claims in this scorecard."

About the Author & Co-Guide

Divy Thakkar wrote this toolkit to give retail investors and finance beginners a practical, checkable workflow for using AI to analyze stocks responsibly — so every output traces back to sources and numbers readers can confirm themselves. The book combines AI tools with disciplined evidence verification to make faster research safer, not riskier.

The methodology in this book was shaped with guidance from Nirmal Rabari, who served as co-guide, helping refine the evidence-first frameworks and real-world workflow structure used throughout.

Final Thought

"Speed only helps when the underlying facts hold up." Thank you for working through this toolkit. Keep your evidence trail tight, keep verifying, and keep researching.