Back to Home

Methodology

This page describes how each kind of content on StockAlpha.ai is actually produced, and what each one cannot tell you. It is deliberately unglamorous. Every content type here is machine-generated — see our editorial policy and AI disclosure for who is accountable for it and how the sources are ranked.

The common pattern across all of them: we pull structured data from our databases, hand that data to a language model with instructions to reason over the supplied numbers rather than recall its own, and publish the result on a schedule. The model writes the prose. It does not source the numbers.

Ticker Reports

Pages under /tickers. Each report covers one security from a fixed research universe of roughly 500 US-listed symbols. Reports are regenerated on a weekly batch schedule over the weekend, with a catch-up pass early in the week for any batch that failed.

Each report is assembled from:

  • Company profile and quote data from our market data providers
  • Fundamentals and analyst estimates from Financial Modeling Prep
  • SEC filing data — 10-K and 10-Q figures, Form 4 insider transactions, 13F institutional positions
  • Recent news and events matched to the symbol
  • A short-interest reading where one is available

What a ticker report is not:

  • It is not a price target, a rating, or a recommendation.
  • It is not current. A report refreshed on Sunday does not know about Wednesday. Check the report date before you use it, and check the live quote rather than the figure in the prose.
  • It is not a substitute for the filings. Where a report and an SEC filing disagree, the filing is right and we want to hear about it — see corrections.

Sector Briefs

Pages under /sectors. We publish a brief for each of 12 sectors, twice daily on weekdays (morning and after the close). Each brief summarizes the news and price action for that sector inside a defined lookback window — roughly the prior 12 hours for a morning brief and the prior 8 hours for an evening one, widened to 48 hours on weekends and about 60 hours on a Monday morning so the window reaches back past the weekend.

A sector brief is a summary of what was already reported, grouped by sector. It contains no original reporting and no forward view. If nothing much happened in a sector during the window, the brief will say relatively little — that is the intended behavior, not a failure.

Market Recaps

Pages under /alpha-recap and /alpha-breaking. Recaps are generated after the US close on trading days from the day's index moves, sector performance, and the news items our ingestion picked up during the session. Breaking items are generated at fixed intervals through the day rather than in reaction to a newsroom decision, so our coverage of an event begins at the next scheduled run, not at the moment it happens.

Recaps describe what moved and what was reported alongside it. Where a recap suggests a reason for a move, treat that as a plausible association assembled from the day's news, not as an established causal explanation. Markets move for reasons that are not in the news.

Alpha Picks

Alpha Picks is our quantitative ranking model. It is in limited internal preview and is not generally available. We describe it here because it is the part of the site most likely to be mistaken for something it is not.

How it works:

  • Universe: S&P 500 constituents only. Roughly 500 large-cap US securities. Nothing about the model has been validated outside that universe.
  • Features: daily end-of-day prices and derived technical measures, fundamentals from filings, insider and institutional activity, short interest, search-attention and sentiment signals, and a small set of macro readings.
  • Model: a gradient-boosted tree model trained on that feature set, blended with a transparent rule-based composite score. Picks are produced per horizon — roughly one day, one week, and one month of holding period.
  • Evaluation: we measure rank information coefficient (how well the ranking orders subsequent returns), win rate, and return net of an assumed round-trip transaction cost, against an equal-weighted mean of the same universe over the same window.

What the model has not demonstrated

  • No statistically significant edge. At our current sample size, the bootstrapped 95% confidence interval around the model's rank information coefficient still includes zero. The point estimate is positive; that is not the same as evidence. We are not going to describe it as a proven signal until the interval clears zero.
  • Confidence scores have been poorly calibrated. In evaluation, higher-confidence picks did not reliably outperform lower-confidence ones. We apply a post-hoc calibration to the displayed confidence, and even so it should be read as a rough ordering, not a probability.
  • The benchmark is the universe, not the index. Reported outperformance is relative to the equal-weighted mean of the same 500 securities. It is not an S&P 500 index comparison and should not be read as one.
  • All results are hypothetical. Picks are recorded to a paper ledger. No customer capital is or has been traded on them. Hypothetical results do not include the slippage, fills, borrow costs and behavior of real trading.
  • Transaction costs are an assumption, not a measurement. We deduct a flat round-trip cost; real costs vary, and short positions carry borrow costs we do not model.

Podcasts and Audio

Our audio is text-to-speech narration of the written article it accompanies, produced by a synthetic voice. It is not a recorded human host, an interview, or a live broadcast, and it introduces no information beyond what the source article contains.

Limitations That Apply to Everything Here

  • Language models make confident mistakes. Supplying the data reduces fabrication; it does not eliminate misreading, wrong emphasis, or a plausible sentence with a wrong number in it.
  • Everything runs on a schedule. Freshness is bounded by the cadence described above, not by the news cycle.
  • Coverage is a fixed list, not a judgment. A security appearing on this site means it is in our generation universe. It does not mean we think it is worth owning, and the absence of a security means nothing at all.
  • Our third-party data can be wrong or stale, and errors in it propagate into our output.
  • Sentiment ratings and outlooks are model opinions. They are not measurements, not consensus, and not advice.
  • None of it is personalized. Nothing here accounts for your holdings, tax position, time horizon or risk tolerance. See our disclaimer.

Changes to This Methodology

We revise this page when the underlying process changes — a new data source, a model change, a different cadence. Material changes are also logged on our corrections page.

Methodology owner: [[PENDING: named editor]], [[PENDING: title]]

Questions: support@jeffersonequityintelligence.com

Last updated: August 13, 2026