Methodology · How GMI actually grades stocks

Two methodologies.
One report. Show your work.

We owe you a clear answer to the most important question: how does GMI actually decide a stock is an A versus an F? Below is the full methodology — what we measure, what data feeds the models, what the limitations are, and where the system can be wrong. No black boxes.

The two-layer principle

Numbers and explanation, kept separate.

Every grade is computed by two independent systems with two different methodologies. Both are visible. Both can be audited. Neither one is allowed to grade alone.

The numbers layer.

Microsoft ML.NET

In plain English: a calculator. Feed it the same numbers and it always gives you the same answer.

Built on Microsoft ML.NET — Microsoft's open-source machine-learning toolkit for .NET applications. Runs every time you generate a report and produces four outputs: the ML Health Score (a 0-to-100 fitness rating of the company), four-quarter ML Forecasts (where revenue, debt, cash flow, and equity are projected to go), anomaly detection (flags moments where a company's numbers did something unusual compared to its own history), and a four-week sentiment forecast (the model's read on whether the next month looks bullish, neutral, or bearish).

ii.

The explanation layer.

Anthropic Claude

In plain English: a research analyst who's read the financials and writes you a brief.

Built on Anthropic's Claude — the same AI behind ChatGPT-like tools, used here for a specific job: read the numbers layer's outputs plus the company's financial statements and produce the AI Grade (A+ to F) along with a written explanation of why. Same scoring rubric applied to every stock, every quarter, so a B+ on one stock means the same thing as a B+ on another.

Why two? Because a single model — statistical or generative — can be confidently wrong. When the numbers layer and the explanation layer agree, that's confirmation. When they disagree, that's a signal worth your attention. You always see both.

№ 01 · The AI Grade

A+ to F. Rubric-based. Auditable.

In plain English: the AI reads sixteen specific data points about each company and gives the stock a letter grade — like a report card — along with an explanation of how it got there.

The grade is composed from 16 weighted metrics across four dimensions: financial health, growth, valuation, and momentum. Claude evaluates each metric against a fixed scoring rubric (the same one for every stock) and produces both the grade and the written rationale that explains every input.

Dimension i.

Financial health.

Is this company solvent and well-run?

Profit margins, return on equity (ROE — how much profit a company makes per dollar of shareholder money), return on assets (ROA — same idea but per dollar of total assets), debt-to-equity ratio (how much the company owes versus what it's worth), and interest coverage (whether the company earns enough to pay the interest on its debt).

Dimension ii.

Growth.

Is it getting bigger over time?

Revenue CAGR (compound annual growth rate — the steady annual growth rate that would get you from where the company started to where it is now), earnings growth, and free cash flow growth across the four-year window we look at.

Dimension iii.

Valuation.

What are you paying for it?

P/E ratio (price-to-earnings — what you pay per dollar of profit), P/B ratio (price-to-book — what you pay versus the company's stated net worth), EV/EBITDA (a fancier price-to-earnings that includes debt), and free cash flow yield (how much cash the company spits out each year, as a percentage of its price).

Dimension iv.

Momentum.

What does the market think right now?

Analyst-target consensus (what Wall Street analysts think the stock should be worth), price-action signals (which way the stock has been moving and how steadily), and the trajectory of the sentiment forecast described later on this page.

Data window

Last 4 fiscal years (annual financials)

Consensus data

Latest analyst-target as of report time

Grade scale

A+, A, A−, B+, B, B−, C, D, F

Rationale

Written for every grade; click to expand

№ 02 · The ML Health Score

A 0–100 composite. Computed, not guessed.

In plain English: a single number from 0 to 100 that says "this company's financials look healthy" or "they don't." Computed by math, not opinion — so the same numbers always produce the same score.

The ML Health Score is computed by Microsoft ML.NET (Microsoft's machine-learning toolkit) from the company's historical financial statements. It's a second opinion that doesn't share a methodology with the AI Grade — when both agree on a stock, that's confirmation; when they disagree, that's a signal worth investigating.

Revenue stability.

How steady the company's top-line revenue has been across the multi-year window — straight line up, choppy zigzag, or in decline.

Debt trajectory.

Which direction the company's total and long-term debt has been moving, and how fast — getting deeper in the hole, paying it down, or holding steady.

Cash flow trend.

Whether the company is actually generating cash (free and operating cash flow) and whether that cash is growing or shrinking over time.

Equity growth.

How fast the company's book value (net worth on paper) is compounding, and whether they're issuing new shares that dilute existing owners.

85–100 Strong

60–84 Solid

40–59 Mixed

0–39 Concerning

№ 03 · ML Forecasts

Four quarters forward. 95% confidence intervals.

In plain English: for each company, the system looks at the past few years of financials, sees the trend, and projects four quarters into the future. Every projection comes with a "best case" and "worst case" so you know how confident the system is.

The forecasts are produced by SSA — short for Singular Spectrum Analysis, a time-series technique that breaks the historical numbers into trend, seasonal patterns, and noise, then projects each piece forward. Implemented in ML.NET. Same inputs always produce the same forecast.

Revenue forecast.

Where the company's quarterly revenue is projected to go over the next four quarters.

ii.

Debt forecast.

How the company's total debt is projected to move over the next four quarters.

iii.

Cash flow forecast.

Where free and operating cash flow are projected to go over the next four quarters.

iv.

Equity forecast.

How the company's net worth (book value) is projected to grow or shrink.

Every forecast ships with three numbers: a best-guess midpoint, an optimistic upper bound, and a pessimistic lower bound — so you can see how confident the model is. (The bounds are 95% confidence intervals — the range the model thinks the true answer probably sits in, leaving 5% room for being wrong.)

№ 04

Anomaly detection.

In plain English: the system watches each company's financial history and flags any moment when the numbers did something unusual compared to that company's own normal pattern.

Example: a company has had steady debt for four years and then debt suddenly jumps 40% in one quarter — we flag it. We don't tell you why it happened. We tell you that something statistically out-of-pattern occurred and you should go figure out whether it was a strategic move (fine) or fundamentals breaking down (not fine).

What gets flagged: spikes, reversals, and sudden shifts that don't fit the company's own multi-year pattern. What doesn't: broad market events, sector-wide rotations, or anything that wouldn't show up in this company's own data.

№ 05

Sentiment forecast.

In plain English: the system's read on whether the next four weeks for this stock are likely to lean positive, negative, or roughly flat — based on technical price patterns and the company's underlying health.

Produces one of three labels — Bullish, Neutral, or Bearish — plus an intensity score and a confidence rating. Built on common technical signals like RSI (Relative Strength Index — measures how overbought or oversold a stock is), MACD (Moving Average Convergence Divergence — measures whether the stock's recent trend is strengthening or weakening), and volatility, layered on top of the fundamental health signal.

Horizon: 4 weeks. Cadence: recomputed on every report run. Honest note: a four-week outlook is the noisiest, least-certain number on the page. Treat it as context, not a directive.

№ 06 · Data sources

Where every number comes from.

Primary provider

Financial Modeling Prep (FMP).

All fundamental financial data — income statements, balance sheets, cash-flow statements, key metrics, analyst targets, price history, company profiles — comes from Financial Modeling Prep, a licensed institutional data vendor. We use 13+ endpoints across their /stable and /v4 API surfaces.

AI provider

Anthropic Claude.

The natural-language layer is Anthropic's Claude, accessed via the official API. Three model tiers are exposed per request: Haiku (fast/cheap), Sonnet (default), and Opus (deepest). The actual cost is shown before every Claude invocation.

ML framework

Microsoft ML.NET.

Health scores, SSA forecasts, anomaly detection, and sentiment classification are all built on ML.NET, Microsoft's open-source machine-learning framework for .NET. Models run in our infrastructure; no external ML provider is involved.

Honest about what we can't do

What this methodology will not tell you.

What to buy. GMI grades the data. We do not issue personalized buy / sell / hold recommendations, and Claude is prompt-engineered to avoid them. The decision is always yours.
The future. ML forecasts are projections based on historical data. Companies pivot, markets break, black swans show up. The 95% confidence interval exists because the model is, by design, uncertain.
Anything we can't see. Methodology runs on financial statements, analyst data, and price history. Insider knowledge, unannounced acquisitions, pending litigation, geopolitical risk — none of it is in the data we have access to.
AI infallibility. Claude can be wrong. We mitigate by grounding it in your real report data, requiring it to cite specific metrics, and pairing every AI Grade with an independent ML Health Score. But a confident wrong answer is still possible.
Day-trading or options signals. Our horizon is days, weeks, quarters, years — not minutes. Fundamentals don't move that fast, and neither do we.

If a methodology page doesn't have a "here's what we can't do" section, treat it as marketing. Treat this one as engineering documentation.

See the methodology in action

Run a real report. For nothing.

Start free — no credit card required. Up to 8 tracked stocks with live grades and one AI report a month, free for life. All the methodology on this page, applied to your actual holdings.

A system by

Disclaimer: Grade My Investments is not a registered investment advisor and does not provide financial advice. All grades, scores, forecasts, and AI-generated content are for informational and educational purposes only — not personalized recommendations. Methodologies may evolve; this page reflects the current implementation. Full Terms.