Anatomy of a Framler signal:
how the engine reads NVDA
The Framler engine produces a single 0–100 score per ticker per day across 1,000+ US, European, and Asian equities. Behind that one number sits a thirteen-factor decomposition, a regime-aware composition layer, a conformal prediction interval, and an honest disclosure of how mature the calibration is for that ticker's sector. This page walks through what each of those parts is and how to read them. For the actual live numbers today, open /stocks/NVDA — this page teaches the vocabulary.
What you see at the top of a ticker page
Each ticker page opens with three pieces of information set apart from everything else — the score, the verdict, and the agreement bar. The score is the headline number (0–100). The verdict translates the score into a qualitative call: bullish, mixed, or bearish. The agreement bar measures how many of the independent factor streams point in the same direction. A high score with high agreement is structurally stronger than a high score with conflicting factors.
The card above is an example layout, not a live read. Open /stocks/NVDA for today's actual score, factor breakdown, regime, interval, and confluence pattern.
The thirteen factor families
Each of the thirteen factors traces to a peer-reviewed paper — Novy-Marx 2013 for quality, George & Hwang 2004 for momentum, Fama-French 1992 for value, Bernard-Thomas 1989 for post-earnings drift, Loughran-McDonald 2011 for sentiment on 10-K filings, Sloan 1996 for accruals, Cohen-Frazzini 2008 for cross-asset spillover, Pan-Poteshman 2006 for options flow, Asquith-Pathak-Ritter 2005 for short-interest, Amihud 2002 for microstructure, Seyhun 1998 for insider activity, Moskowitz-Grinblatt 1999 for sector-relative momentum, and Asness-Moskowitz-Pedersen 2013 for the interaction factor that rewards stocks where quality, value and momentum simultaneously agree.
For NVDA today, the dominant contributors are typically quality (gross profitability premium), momentum (52-week-range position), and PEAD (post-earnings drift window after the most recent print). None of these is a view about Nvidia specifically — they are the same factor measurements applied to every ticker in the universe. The edge, if there is one, comes from the composition, not from any single factor being newsworthy.
Full per-factor descriptions live at /methodology; click any factor name on the ticker page to jump to that paper's explainer.
The 90% prediction interval
Most stock-rating tools give you a point — a single number. Framler gives you a range: the 90% conformal prediction interval (Vovk 2005). The score sits in the middle; the interval brackets the realistic outcome band. A narrow interval means the calibration data on similar stocks (regime × tail-dependence bin) shows the engine has been tight on this kind of setup historically. A wide interval means the opposite — recent calibration data is noisy and the engine is not yet certain about the right band.
The Mondrian binning (Vovk 2003) means different stocks get different interval widths even at the same score. A quality-aligned mega-cap in a stable risk-on regime gets a tighter band than a high-momentum biotech the week before a Phase 3 readout. This is how the engine tells the reader "same score, different confidence" honestly.
The regime context
Every Framler score is computed under the assumption of one of three macro regimes — risk-on, transition, or risk-off — inferred each day by a Bayesian Online Change-Point Detection model (Adams-MacKay 2007) reading SPY price action, VIX implied volatility, breadth, the yield curve, high-yield versus investment-grade credit spreads, and the dollar index. In risk-on, momentum and quality amplifiers run at full strength; in risk-off, they damp sharply (Hamilton 1989; Ang-Bekaert 2002).
The regime label and the underlying posterior probabilities are visible on each ticker page as a small badge near the score. They explain why the same factor configuration on the same ticker can produce a different headline score depending on what the market is doing.
The sector-calibration honesty layer
This is the part most retail tools omit. Different sectors have different rolling-window Information Coefficients on the engine — Financials, Industrials, and Materials have historically scored predictively; Technology, Energy, and Communication Services have run closer to zero or negative. When you open a Technology ticker, the page displays a calibration-warning strip alongside the score that says something like "Sector calibration noisy — engine accuracy in Technology has been close to zero over the last 14 days". A high bullish score on a sector where the engine itself confesses low predictive accuracy should be discounted accordingly.
This is the discipline most fintech tools choose not to ship — telling the reader where the model is uncertain costs conversion. We do it because the alternative is a pitch that the engine refuses to make.
The forecast surface
Below the headline score, ticker pages render a forecast row for four horizons — 1, 7, 30, and 90 trading days. For each horizon the engine publishes an expected return, a 90% confidence band, and a probability of a positive return. The 1d and 7d horizons have been calibrated out-of-sample against the engine's own walk-forward measurement since 2026-05-17. The 30d horizon completes around 2026-06-26; the 90d horizon around 2026-09-15. Until each window matures, the engine falls back to literature-prior amplifiers and widens its interval accordingly — visible on the calibration badge.
The full forecast surface for NVDA lives at /stocks/NVDA/forecast.
Confluence patterns — when independent factors align
On top of the linear factor composite, the engine runs a confluence-pattern overlay — about two dozen named multi-factor configurations like QUALITY COMPOUNDER (quality + value + accruals + risk-on), SHORT SQUEEZE SETUP (Asquith-Pathak-Ritter 2005 — high short interest, high momentum, high options-flow), or PRICED FOR PERFECTION (Shiller 2000 asymmetric-downside). When a pattern fires, the engine adds a small calibrated boost or dampener consistent with that pattern's historical effect size in the literature.
Pattern firing histories are visible on /patterns with hit-rate and mean-return statistics per pattern; the confluence-pattern label that fired on a particular ticker today shows on the ticker page itself.
What the page does NOT do
The Framler score is not a price target, a recommendation, or a trade signal. It is a research signal — one input to a reader's own analysis, alongside fundamentals, valuation, position-sizing rules, and risk tolerance the engine cannot know. The engine cannot tell you what to do with money. It can show you how thirteen independent factor streams read a ticker today and how confident the calibration is on that reading.
Live measured per-horizon Information Coefficients across the universe sit on /track-record. The full pipeline architecture sits on /methodology. Mathematical coherence checks (factor weights sum to one, BOCPD posterior sums to one across regimes, bounded correlations, positive interval widths) run live on /coherence.
Read a real ticker
The walkthrough above describes the shape of every ticker page in the universe. Open one and read it for yourself — the structure is identical, the numbers are live.
Or read /methodology for the academic stack behind every signal, or /learn for plain-English definitions of every term used in this walkthrough.