stock-analysis – skill Dữ liệu & Khoa học cho agent AI

Stock Analysis

Produce an evidence-backed fundamental analysis of one company, benchmarked against the right peers, and delivered as a written report plus a sector-relative scorecard.

The principle that governs everything here

A financial metric carries no meaning until you know the sector it came from and the company's own history.

If X earns a 20% operating margin and Y earns 30%, that tells you nothing about which is the better business. Y may be in software (where 30% is mediocre) and X in distribution (where 20% is exceptional). Y's 30% may need three times the capital to produce, so X earns a far higher return on the money invested. Y's margin may be eroding while X's compounds.

Two consequences shape this whole skill:

1. Never rank companies on a single metric. Every judgement combines profitability, returns on capital, cash conversion, balance sheet, growth durability, governance, and price.

2. Compare like with like. Benchmark against sector peers or against the company's own multi-year record — never a raw cross-industry number. For banks, insurers, REITs and miners the standard ratios are not merely less useful, they are *undefined or inverted*; those sectors need their own metric set entirely.

Read references/05-returns-and-dupont.md for why return on capital, not margin, is the metric that actually determines compounding.

Non-negotiables

Never invent a number

This is the failure mode that destroys the value of the whole analysis. A fabricated revenue figure or a hallucinated ROCE produces a confident, well-formatted, *useless* report — and the user may act on it.

• Every figure carries a source and a period ("FY25 annual report, consolidated, p.112" / "10-K FY2024, Item 8" / "Q3 FY26 quarterly results filing, BSE").
• If a number cannot be sourced, write not available and say what would be needed. An analysis with acknowledged gaps is far more valuable than one with invented precision.
• Cross-check headline figures (revenue, net profit, debt, cash) against a second source when possible — at least one of the two must be a primary document.
• Every financial figure in the analysis must trace to a primary document — annual report, 10-K/10-Q, quarterly results filing, concall transcript, investor presentation, DRHP/RHP, exchange filing, or rating rationale. Aggregator websites (screener.in, Yahoo Finance, Tikr, etc.) are navigation aids for locating documents and optional labelled cross-checks — they are never a source of record. The one exception is current share price and market cap, which are inherently sourced from exchange or finance websites and must carry an as-of date.
• State consolidated vs standalone explicitly — for any company with subsidiaries these differ materially, and mixing them silently invalidates every ratio.
• State currency and units. Indian filings use crore/lakh; US filings use millions/billions. Getting this wrong by 10x is a common and embarrassing error.
• Flag stale data. A price or multiple without an as-of date is not usable.

Detailed sourcing routes and a verification protocol: references/01-data-sourcing.md.

Official records are the source — and they hold far more than the financial statements

Two failure modes hide behind a report that looks well-sourced. Guard against both.

First: the source of record is the company's own filings — nothing else is. Rank sources by how many hands the number has passed through, and cite only the primary one:

1. Primary filings — annual report / 10-K, exchange filings (NSE/BSE, SEC EDGAR), quarterly results, the offer document (DRHP/RHP/S-1), audited statements.

2. Company-published secondary — concall transcripts, investor presentations, earnings releases.

3. Regulator / third-party primary — SEBI/MCA/ROC records, credit-rating rationales, exchange shareholding and pledge data.

Third-party research notes, brokerage reports, news articles and data aggregators (screener.in, Tikr, Yahoo/Google Finance, trendlyne) are navigation and cross-check aids only — they exist to help you *locate* the filing and to flag an outlier worth investigating. An aggregator or news figure must never be the thing you cite; when it disagrees with the filing, the filing wins and the disagreement is itself a finding. The one standing exception is live share price and market cap, which carry an as-of date. If a figure exists only in an aggregator and cannot be traced to a filing, it is not sourced — say so.

Second: a filing is not just its three financial statements. Most of what actually decides an analysis is the non-financial disclosure wrapped around the numbers, and it must be read and used as a first-class input — not skimmed on the way to the P&L:

• the business, strategy and risk-factor sections — what is sold and to whom, the stated moat, and the risks management is legally obliged to admit;
• MD&A read across 3–5 years — growth decomposed into volume / price / mix, capacity, capex plans, order book, guidance, and the drift between what was promised and what was delivered;
• the auditor's report, CARO annexure, Key Audit Matters and emphasis-of-matter — the auditor's own map of where the numbers are fragile;
• related-party transactions, contingent liabilities, litigation and capital commitments — the commonest routes for value to leave a minority shareholder, and quantifiable in one sitting;
• governance and ownership — board and audit-committee composition and independence, promoter holding trend and pledge, remuneration versus profit, auditor tenure and any resignation, AGM voting dissent, ESOP dilution;
• segment and operational data — segment-level revenue, EBIT and capital employed (segment ROCE is usually the report's most surprising number), plus the sector KPIs — capacity utilisation, occupancy/ARPOB, ANDA filings, same-store growth, order-book conversion — that never appear in the income statement;
• ESG/BRSR, secretarial audit (MR-3), and subsidiary (AOC-1) disclosures.

references/15-document-diligence.md is the runbook for extracting all of this, with a time-boxed reading order. Treat it as part of the core workflow, not an optional deep-dive: an analysis built only on the income statement, balance sheet and cash flow has read perhaps a fifth of the official record and skipped the four-fifths where the moat, the governance and the landmines live.

Analysis, not advice

Produce analysis, evidence, and a reasoned view of business quality and valuation. Do not produce personalised investment advice, position sizing for the user, or buy/sell instructions framed as recommendations for their money. State clearly that the output is research, not licensed financial advice, and that the user is responsible for their own decisions.

Presenting a bull case, a bear case, a valuation range, and what would falsify the thesis is genuinely useful and stays on the right side of this line. "You should buy 50 shares" does not.

Show the reasoning and the uncertainty

Where an estimate is used (normalised earnings, maintenance capex, mid-cycle margins), say it is an estimate, give the assumption, and show what changes if the assumption is wrong. False precision — a target price to two decimals off a hand-waved growth rate — is worse than an honest range.

Choose a depth mode

Match effort to what the user asked for. Announce which mode you are running so expectations are set.

| Mode | When | What it covers |

|---|---|---|

| Screen | "quick take", "is this worth looking at" | Stages 0–3 plus valuation sanity check. Kill criteria, headline quality metrics, obvious red flags. Short verdict. |

| Standard (default) | "analyse this stock" | All stages, moderate depth per stage, full scorecard and report. |

| Deep dive | "detailed", "thorough", "maximum depth", or a position the user intends to size | All stages at full depth, situation playbook, document-level diligence, forensic pass, scenario valuation, explicit bear case. |

| Forensic | "is the profit real", "are they cooking the books", "cash flow doesn't match profit", "check the accounting" | A different question entirely — *can these accounts bear weight?* Skips business quality, growth and valuation. Follow references/18-forensic-mode.md. |

| IPO | The company is not yet trading — an open or upcoming IPO, a filed DRHP/RHP, "should I apply to X's IPO" | No market price and no public track record, so own-history benchmarking and market-price valuation are both unavailable. Follow references/19-ipo-mode.md. |

The workflow

If you are running Forensic mode, stop here and follow references/18-forensic-mode.md instead — it has its own stages (F0–F5) and its own verdict scale, because "can I trust these numbers?" is not answered by a shorter version of "is this a good investment?".

If the company is not yet listed, stop here and follow `references/19-i

… (nội dung đầy đủ có trong gói tải về)