How to Use AI to Analyze Stocks: A Step-by-Step Workflow

Why Pasting a Ticker Alone Fails

To use AI to analyze stocks well, give the model real data before you ask it anything. A prompt like "analyze NVDA" with no attached numbers forces the model to answer from memory, and memory is the weakest part of a language model when the subject is a balance sheet.

The evidence is specific. A CoLM 2025 study (arXiv 2504.00042) tested LLMs on more than 197,000 questions about the financial data of U.S. public companies. The models knew less about past performance than recent results, and they were more likely to hallucinate figures for larger companies, especially for recent years.

So the workflow below has one rule: the model computes and reasons over data you supply, never data it recalls. The six steps are (1) real data in, (2) fundamentals, (3) technicals, (4) sentiment and news, (5) a bull/bear debate, and (6) a checklist rather than a trade.

Step 1: Get Real Data Into the Model

There are three practical ways to put market data in front of a model.

MCP (Model Context Protocol) is the fastest route for an ongoing workflow. Alpaca's Trading MCP Server exposes 65 tools across its Trading and Market Data APIs, including get_stock_bars, get_stock_snapshot, and news lookups, and works with Claude Desktop, Claude Code, Cursor, VS Code, and Gemini CLI (Alpaca docs, 2026). Two settings matter for analysis-only use: ALPACA_PAPER_TRADE defaults to true, and ALPACA_TOOLSETS can restrict the server to read-only market data. Tradewink runs a hosted MCP server with 39 tools on the same idea, including get_technical_analysis, get_earnings, get_news, and analyze_ticker.

Without a connection, a CSV upload works. Export 120 daily bars (date, open, high, low, close, volume) and attach the file. Every figure the model cites can then be traced to a row.

Confirm the data landed: ask the model to print the first and last three rows plus the date range. If it cannot, nothing downstream is trustworthy.

Step 2: Fundamentals Pass With SEC Filings

The fundamentals pass reads what the company filed, not what a summary site says it filed. SEC EDGAR Full-Text Search covers the full text of every electronic filing since 2001, including exhibits, and supports exact phrases, exclusion with a hyphen or NOT, OR clauses, wildcards, form-type filters, and a NEAR() operator for two terms within a set number of words (SEC EDGAR FAQ). It does not support natural-language queries.

A practical sequence:

1. Open the latest 10-K and 10-Q. Copy the income statement, cash flow statement, "Risk Factors," and "Management's Discussion and Analysis."
2. Search EDGAR for the exact phrase "going concern" plus the company name, filtered to the last two years.
3. If an earnings call transcript exists, include the Q&A. Analyst questions are where the friction shows.

Paste those excerpts and use this prompt:

Analyze the attached 10-K and 10-Q excerpts for [COMPANY].
Rules:
- Use only numbers in the attached text. Quote the exact line for
  every figure you cite.
- If a number is not in the text, write "NOT IN SOURCE". Do not
  estimate it.
Tasks:
1. Revenue, gross margin, operating margin, and free cash flow for
   the last two fiscal years, with year-over-year change.
2. Debt maturing within 24 months versus cash on hand.
3. The three risk factors that changed most versus the prior filing.
4. Any language on customer concentration, going concern, or
   material weakness in internal controls.
Output a table, then a 100-word summary of what changed.

The "NOT IN SOURCE" rule is the most important line in the prompt. It gives the model a permitted way to say it does not know, which cuts the rate of invented figures. The guide to reading earnings reports covers which lines to extract.

Step 3: Technical Pass: Compute, Do Not Recall

Language models are decent at arithmetic on data they can see and poor at recalling indicator values from memory. A model that receives 120 closes and computes a 14-period RSI step by step is doing math you can check.

Prompt template for the technical pass:

Attached: daily OHLCV bars for [TICKER], [START DATE] to [END DATE].
Compute, showing intermediate values:
1. 20-day and 50-day simple moving averages as of the last bar.
2. 14-day RSI with Wilder smoothing. Print the average gain and
   average loss used.
3. 14-day ATR and ATR as a percent of the last close.
4. Three highest swing highs and three lowest swing lows in the last
   60 bars, with dates.
5. Average volume over the last 20 bars versus the last bar.
Use no price or indicator value not derived from the attached bars.
If the data is insufficient, say so.

Spot-check one number by hand. If the 20-day SMA is wrong, discard the pass. The RSI entry has the formula the model should follow.

Step 4: Sentiment and News Pass

News is where a knowledge cutoff is most dangerous, because the model will describe "recent" events that are months old. The fix is to supply the headlines yourself, with dates, and require the model to attribute every claim to a supplied item.

A 2024 live experiment in Finance Research Letters found that ChatGPT-4 with internet access produced stock "attractiveness ratings" that correlated positively with later earnings announcements and returns, and updated them after earnings surprises and news (Pelster and Val, 2024). The narrow lesson: a model given fresh information can classify it usefully. A model given nothing cannot.

Attached: [N] news items for [TICKER], each with date, source, and
headline. Today's date is [DATE].
For each item:
1. Classify as positive, negative, or neutral for the stock.
2. Classify as company-specific, sector-wide, or macro.
3. Rate durability: one-day noise, one-quarter effect, or multi-year.
Then:
4. List any item that contradicts another item.
5. Name the most important scheduled event in the next 30 days that
   appears in the items. If none appears, say "none in source".
Do not add events or facts that are not in the attached items.

Step 5: The Bull/Bear Debate Prompt

A single model asked for "both sides" produces a bull case and a bear case that share one underlying view, because both come from one forward pass. Research presented at ICML 2025 (Kim, Garg, Peng, and Garg, "Correlated Errors in Large Language Models") evaluated more than 350 models and found that when two models both erred on a leaderboard question, they gave the same wrong answer about 60% of the time, with higher correlation between models from the same developer or on the same base architecture.

Three practical consequences:

1. Run the bull and bear cases in separate sessions so neither sees the other's draft.
2. If you can, give the bear case to a different model family (Claude and GPT, or Claude and Gemini).
3. Add a neutral third session whose only job is to list where the two cases disagree on facts.

You are the [BULL / BEAR] analyst for [TICKER]. Attached are the
fundamentals table, technical summary, and news classification.
Argue the [BULL / BEAR] case only.
- Every claim must cite a row or line from the attached material.
- Give exactly three arguments, ranked by how much they would move
  the stock if true.
- For each, state the number or event that would prove it wrong
  within 90 days.
- End with a confidence score from 0 to 100 and one sentence on what
  would raise it by 20 points.
Do not mention the opposing case.

The falsification line ("what would prove it wrong") matters more than the arguments. It becomes the checklist in Step 6. The multi-agent AI trading entry covers how separate sessions and a disagreement veto fit together.

Step 6: Turn the Output Into a Checklist, Not a Trade

The output of an AI stock analysis is a list of conditions, not an order ticket. The final prompt converts everything above into checkable items.

From the attached bull case, bear case, and disagreement list,
produce a pre-trade checklist for [TICKER]:
1. Facts both sides accept (with source line).
2. Facts in dispute, and the single data point that would settle each.
3. Five conditions that must all be true before any position is
   considered, each a yes/no question with a number.
4. Three invalidation triggers: a price level, a fundamental event,
   and a time limit.
5. Sizing inputs only: entry range, stop distance in dollars and ATR
   multiples, and risk per share.
Do not recommend a buy or sell. Do not state a price target.

Take the sizing inputs to the position size calculator. If the numbers survive, run the idea in a paper trading account before any capital is involved.

Worked Example: Northwind Robotics (Hypothetical)

Every number in this section is invented. Northwind Robotics (fictional ticker NWRB) is a made-up industrial automation company.

Step 2 flags what most summaries miss: $95M of debt maturing against $58M of cash means the company needs a refinancing or roughly two more years of the current $31M free cash flow. The model marks the refinancing terms as "NOT IN SOURCE" because the 10-Q does not disclose them.

Step 3 shows price above both moving averages with RSI at 63, below the usual 70 overbought threshold. ATR of $1.90 means a 2-ATR stop sits $3.80 below entry.

Step 5, run in two sessions, produces a bull case built on free-cash-flow growth and a bear case built on the debt wall. The neutral session finds the two agree on every number and disagree only on whether a refinancing will close. That disputed fact becomes the top line of the Step 6 checklist: "Has the company announced refinancing terms for the $95M maturity, yes or no."

Sizing inputs from Step 6, still hypothetical: entry $46.50 to $47.50, stop $3.80 below (2 ATR), risk per share $3.80. On a hypothetical $25,000 account risking 1% per trade ($250), that is 65 shares, about $3,070 of exposure. Nothing here is a recommendation; it shows how the arithmetic flows into a sizing decision.

What LLMs Are Good and Bad At for Stock Analysis

Limitations You Cannot Prompt Away

Four constraints survive every prompt template:

• Knowledge cutoff. The model knows nothing after its training date. Treat any "recent" fact stated without a supplied source as stale until you verify the date.
• Hallucinated figures. The CoLM 2025 study counted a stated figure more than 10% off the true value as a hallucination and found the problem concentrated in the best-covered companies. Attach the source and require quotes.
• No live prices. Without a connected tool, any price the model mentions is a memory, not a market. The MCP route in Step 1 exists to solve this.
• Correlated model errors. Two sessions of one model, or two models from one developer, share blind spots. Different model families reduce the overlap but do not remove it (Kim et al., ICML 2025).

None of these are fixed by better prompting. They are fixed by supplying data, requiring citations, splitting sessions, and keeping a human on the final decision.

How Tradewink Handles This

The workflow above, automated, is the design. The MCP server supplies quotes, bars, earnings, filings, and news to Claude or any MCP client through tool calls, so the model never recalls a number. Indicators are computed server-side from stored bars, not by the model. AI conviction scoring applies an additive, capped boost to rule-based screener output; the AI conviction scoring guide explains why the model stays out of the primary signal. Everything defaults to paper mode.

This article is educational. Tradewink is a publisher of trading tools and research, not a registered investment adviser, and nothing here is financial advice or a buy or sell call on any ticker. Stock trading involves substantial risk of loss, AI output can be confidently wrong, and past results do not predict future performance.