How Framler scores stocks

0. System overview

Framler is organised as a pipeline where one online posterior over the current market regime feeds several downstream decisions simultaneously — factor weighting, covariance shaping, prediction-interval width, and confidence scaling. The same posterior powers all of them, so the outputs are coherent rather than independently averaged.

Thirteen academic factor families are composed into a per-ticker 0–100 score. A confluence engine then applies literature-backed multi-factor pattern overrides (e.g. Sloan + gross-margin crack; Asquith short-squeeze setup). Every override carries a published effect size and is conditionally fired only when the participating factors show genuine joint tail-dependence — not merely high average correlation.

The engine is Bayesian end-to-end: priors come from the literature, posteriors update from live residuals (once forward returns accumulate), and uncertainty is reported alongside every point estimate as a conformal interval.

0.5 How to read Framler cards

Each stock card carries five independent badges. They sit on different axes — none of them is a re-skin of another — and beginners often read them as duplicates of the Framler score number. This section maps every label to what it actually encodes.

Framler score (0–100) — the headline

The single number computed by the 13-factor composite. Every other label below is supplementary; if you only look at one thing, look at this.

Tier — size of insider activity (separate axis from Framler)

Insider-trading magnitude over the last 90 days, classified per Lakonishok-Lee 2001 (larger trades carry more information). Independent of Framler — a stock can have Framler 80 and still be tier 3 (no insiders bought), or Framler 50 and tier 1 (big insider buy on a mediocre composite).

Signal label — Framler × insider, in plain words

The right column of each card combines Framler and insider size into one phrase. Format: PRIMARY · MODIFIER — primary carries the verdict colour, modifier drops to brand cyan as a secondary-evidence cue.

Bearish side mirrors these — STRONG BEARISH, INSIDERS EXITING · MIXED, BIG SELLING · WATCH, MODERATE BEARISH, FACTORS REJECT, FACTORS BEARISH, QUANT TILT BEAR — when the verdict is SELL.

TRIPLE badge — three sources at once

A rare configuration. Fires when all three are true on the same ticker on the same day: insider net-buy ≥ $1M, at least three positive news catalysts (with positive > negative), and conviction ≥ 60%. The TRIPLE badge appears in the card header and pushes that card to the top of the feed.

CONVICTION % — supplementary evidence bar

A separate Bayesian posterior over five supplementary signals (insider flow, news sentiment, fund accumulation, analyst consensus, 52-week price position) combined as log-likelihood ratios. Not the Framler score — this measures how strongly the supplementary stack agrees with the engine, not the engine's own conclusion.

X / 6 evidence streams — pill count

The small line under the signal label. Counts how many of six fixed supplementary streams have content for this ticker: 13F-fund accumulation, insider buy/sell ≥$100K, ≥2 positive news catalysts, ≥5 covering analysts, 52-week price in the top/bottom 20%, and Framler ≥75. Not the 13 academic factors — those are the engine internals; these are the pills shown on the card.

Confluence pattern (green plaque, e.g. CONFIRMED BEAT +22 · tail 0.08)

When a specific combination of the 13 factors matches a documented academic setup, the matching pattern fires and overrides part of the composite. 19 patterns total — VALUE TRAP, SHORT SQUEEZE, CONFIRMED BEAT, EARNINGS VALIDATED, PHARMA CATALYST NEAR, ADCOM NEAR, and others. Each is anchored to a published paper (Bernard-Thomas, Asquith-Pathak-Ritter, Sloan, etc.).

The number after the pattern name (e.g. +22) is the score shift the pattern applied. The tail 0.08 figure is the copula tail-alignment for the pattern's factor set — higher means the factors in the setup co-crash reliably in history, low means the joint behaviour is closer to independent. Patterns only fire when their factor set shows genuine joint tail behaviour, not merely average correlation; the specific cutoffs for amplifying vs damping a pattern are calibrated and proprietary — see §2 Tail-dependence layer.

Other small things

14-day sparkline — recent Framler trajectory. Green up, red down, grey flat. Shows FRESH when there are fewer than two days of history.

Market regime badge (top of dashboard) — the BOCPD regime label: Expansion / Normal / Slowdown / Contraction (or risk-on / transition / risk-off in the older naming). Factor weights shift by regime — momentum gets more weight in expansions, quality more in contractions. See §4 Online regime detection.

Sector tag — Healthcare and Biotechnology unlock a pharma-tilted weight vector and special pharma signals (Phase 3 failure probability, AdCom-event proximity, trial enrollment). P/E is ignored in biotech mode; the engine reads cash runway, gross margins, and pipeline momentum instead.

1. Thirteen factor families

Each factor returns a 0–100 score per ticker per day; higher is more bullish. Factor scorers are self-normalising against the live universe so a 70 means "top-quintile today", not against a stale literature baseline.

Thirteen factors is a headline number; the effective breadth after removing redundancy is lower. We report that explicitly — the composite optimiser penalises correlated signals rather than double-counting them.

1.5 Macro & commodity context

The thirteen factors above are per-ticker: they describe what a single name looks like today. Two universe-wide context layers sit alongside them.

2. Tail-dependence layer

Linear correlation describes average co-movement; in crises what matters is what factors do together specifically in the tails. A Gaussian copula with ρ = 0.5 has zero tail-dependence; a Student-t copula at the same ρ has meaningful joint-crash probability. We estimate lower- and upper-tail dependence non-parametrically for every factor pair, per regime, and shrink toward an independence prior so thin-sample pairs do not dominate.

The Confluence Engine uses this directly: a pattern's confidence is amplified when its participating factors also show genuine tail-alignment in history, and damped when the alignment is weak. This prevents "textbook pattern, no real co-movement" overrides.

Method: Schmidt & Stadtmüller 2006; Frahm, Junker & Schmidt 2005; Embrechts, McNeil & Straumann 2002.

3. Regime-aware risk model

Standard mean-variance (Markowitz) optimisation solves for weights that minimise factor-score variance for a given information-coefficient vector. Using Spearman ρ as the covariance input under-estimates correlated crash risk — exactly the risk that matters for sizing. Framler blends the correlation matrix with the tail-dependence matrix, with the blend weight determined by the current regime posterior: crash-dominated in risk-off, squeeze-dominated in risk-on, unbiased in calm regimes.

A Tikhonov ridge auto-tunes to guarantee positive-definiteness even when the empirical tail-dependence matrix is near-singular on thin samples. The same regime posterior that decides the blend also supplies the factor weights themselves — one posterior, two coherent uses.

Method: Grinold & Kahn 1999; Embrechts, McNeil & Straumann 2002; Higham 2002 on nearest-correlation projection.

4. Online regime detection

Hidden Markov Models give discrete regime labels that flip abruptly. Framler uses a Bayesian online changepoint detector that maintains a full posterior over how long the current regime has been running. From the run-length posterior we derive a probability vector across three market regimes — risk-on, transition, risk-off — so the engine is honest about uncertainty near turning points rather than forcing a binary decision.

Fresh regime (low expected run-length) ⇒ the engine admits ignorance and leans on transition priors. Mature regime ⇒ directional bias takes over. The same posterior powers factor-weight blending, covariance blending, interval width, and dynamic exposure updates (sections 3, 5, 6). One posterior, four coherent decisions.

Method: Adams & MacKay 2007; Hamilton 1989 for baseline hazard rate.

5. Bin-conditional prediction intervals

Marginal conformal intervals guarantee coverage on AVERAGE across the universe — over-covering easy tickers and under-covering hard ones. Framler uses a bin-conditional (Mondrian) variant that partitions tickers by tail-alignment strength and regime, then calibrates one width per partition. The result: the stated 90% interval is close to 90% inside each bin, not only on average.

Until enough forward-return observations accumulate per partition, Framler falls back to a marginal interval inflated by average tail-dependence. This is the single place where the engine publicly admits the difference between "calibrated" and "prior" — every ticker page shows it.

Method: Vovk, Lindsay, Mammen & Vovk 2003 (Mondrian Confidence Machine); Angelopoulos & Bates 2021 for introduction.

6. Dynamic factor exposures

Static factor weights lag regime shifts. Framler tracks factor exposures as a Kalman dynamic linear model where the per-state random walk is accelerated on days the regime detector signals a change-point. Process noise is driven directly by the change-point posterior, so exposures pivot fastest on precisely the days they need to — again the same posterior powering yet another decision.

The DLM output is blended with the regime-aware Markowitz solution. Early in calibration the blend leans on Markowitz priors; as the DLM accumulates observations it takes over. Both are reported; the blend is transparent.

Method: Kalman 1960; West & Harrison 1986, 1997 on Bayesian forecasting and dynamic models.

7. Factor attribution

When the Confluence Engine overrides the linear composite (e.g. a value-trap pattern pulls a superficially-cheap ticker down), which factor deserves the credit? The naïve linear answer is misleading — overrides fire only under joint conditions, so part of each factor's contribution comes from its role in enabling the pattern, not from its standalone score.

Framler uses a Shapley decomposition of the score attributable to each factor. Every contribution splits cleanly into a linear share (the coefficient-style piece) and an interaction share (the pattern-enabling non-linearity). The four uniqueness axioms guarantee a consistent story across tickers. Per-ticker attribution is available in the authenticated research API.

Method: Shapley 1953; Štrumbelj & Kononenko 2014 for ML interpretability.

8. Multiple-testing correction

A daily scan of the universe across the factor stack runs several thousand simultaneous hypothesis tests. Under a naïve per-test significance threshold a meaningful fraction of those tests would be flagged as "significant" under pure noise — hundreds of spurious high-conviction signals per day if uncorrected. Framler runs every daily emission through a Benjamini-Hochberg false-discovery-rate correction, controlling the expected proportion of false discoveries among rejections rather than naïvely rejecting everything that crosses a per-test threshold. The exact significance level and target FDR rate are calibrated and proprietary.

The null-distribution scale uses a robust estimator (median absolute deviation with a Gaussian-consistency multiplier) rather than the empirical standard deviation. The robust scale is unaffected by the strong signals we're trying to detect — a plain σ would inflate when real signals are present, hiding the very signals it's meant to find. The live FDR-adjusted rejection set is published at /api/diag/fdr alongside the robust scale and the BH cutoff p-value, so the correction is observable, not just claimed.

Method: Benjamini & Hochberg 1995, Controlling the False Discovery Rate, JRSS-B 57(1); Bonferroni 1936 (FWER fallback); Huber 1981 on robust scale; Storey 2003 review for finance.

9. Risk-parity factor blending

Equal-weighted factor blends ignore that factors have radically different time-series volatility. A jumpy factor that swings on every earnings announcement contributes more variance to the composite than a sticky balance-sheet factor that drifts slowly, even at identical nominal weights. The user reads "20% PEAD, 20% Quality" as equal influence; the math says otherwise.

Framler inverse-volatility-scales each factor's effective contribution before blending. Per-factor σ is estimated as the average per-ticker temporal standard deviation over a rolling window, refreshed once per cron run. The window length and minimum sample threshold are calibrated and proprietary. The result: each factor contributes proportional varianceto the composite, not just proportional weight. The estimator is silent on insufficient history — until signal history accumulates a stable per-ticker σ, the blender falls back to the plain weighted average automatically.

Method: Maillard, Roncalli & Teïletche 2010, The Properties of Equally-Weighted Risk-Contribution Portfolios, JPM 36(4).

10. Modern-phrase NLP augmentation

The Loughran-McDonald 2011/2018 financial sentiment dictionary is the academic standard but was built on 10-K corpora through ~2018. Modern earnings-call and MD&A vernacular — "AI capex", "design win", "valuation reset", "guidance cut", "hyperscaler buildout", drug-class names like GLP-1 — does not tokenise to any LM category. Tickers heavy in this language scored zero NLP signal even when the sentiment was unambiguous.

the engine runs a curated multi-word phrase scanner over the raw text before LM tokenisation. Phrase matches increment the same category counters (positive / negative / uncertainty / litigious) as single-word LM hits, so the downstream percentage formulas and z-score normalisation are unchanged. The phrase list is curated against 10-K MD&A and earnings-call transcripts circa 2024-2026; speculative meme vocabulary is intentionally excluded — it belongs in a separate retail-sentiment lexicon.

Anchor: Loughran & McDonald 2011, JoF 66(1); Tetlock 2007 for tone-vs-return causality. The phrase set is curated, not derived from a Word2Vec / BERT cluster — those approaches require a labelled 2024+ corpus we do not yet license.

11. Sigmoid-Platt confidence calibration

Confidence numbers exposed to users have to be probabilities, not heuristic point sums. Linearly clipping a heuristic additive total at [10, 97] saturates without information — two tickers with raw totals of 100 and 200 both render "97" despite the second carrying twice the conviction. Framler passes the heuristic through a logistic (sigmoid) transform centred at the bullish-action threshold. The output saturates asymptotically at 0% / 100% rather than clipping, monotone in the heuristic sum, so cross-ticker rankings are preserved while the spacing becomes calibrated.

This is heuristic-Platt: the sigmoid scale and centre are set by hand to match the legacy verdict-tier semantics — exact values are calibrated and proprietary. True Platt scaling — fitting the sigmoid parameters by maximum likelihood on realised forward-return outcomes — kicks in horizon-by-horizon as samples mature. 1d and 7d horizons calibrated on 17 May 2026; the 30d horizon clears its ≥30-sample gate around 26 June 2026, and 90d around 15 September 2026.

Method: Platt 1999, Probabilistic Outputs for Support Vector Machines; Niculescu-Mizil & Caruana 2005 for calibration comparison.

12. Proper scoring rules + reliability diagram

Information coefficient (IC) tells you whether scores predict returns. Proper scoring rules tell you whether the model's probability statements are honest. A model can post IC = 0.10 but if it says "90% confidence" on signals that come true only 70% of the time, the calibration is dishonest and position-sizing built on it overshoots.

The engine computes three calibration metrics on persisted forward returns:

• Brier score (Brier 1950) — squared error of P(positive) vs realised binary outcome. Lower = better. Decomposes into reliability − resolution + uncertainty (Murphy 1973).
• CRPS (Matheson-Winkler 1976) — continuous generalisation of Brier for the full Gaussian forecast, not just P(>0). Closed form for our N(μ, σ²) projections.
• Reliability diagram (Murphy-Winkler 1977) — predictions bucketed by probability, observed positive rate vs predicted plotted. Honest calibration lies on y=x. Expected Calibration Error (ECE) summarises deviation in one number.
• PIT histogram + KS test(Gneiting-Balabdaoui-Raftery 2007) — Probability Integral Transform of every forecast against its realised outcome distributes Uniform[0, 1] under perfect calibration. Histogram shape diagnoses the failure mode: U-shape = over-confident (σ too tight); inverted-U = under-confident; tilted = biased mean. The Kolmogorov-Smirnov statistic tests uniformity formally; p < 0.05 means the engine is statistically inconsistent with honest calibration.

The output of this layer flows back into the engine via recommendHalfwidthScale(): when observed coverage in a Mondrian bin sits below the nominal target, the next calibration pass widens that bin's halfwidth in proportion to the gap. Calibration is not a one-shot operation; it's a feedback loop.

Method: Brier 1950; Murphy 1973 decomposition; Matheson & Winkler 1976; Gneiting & Raftery 2007.

13. Knockoff filter — FDR control without independence

Section 8 (Benjamini-Hochberg) controls the false-discovery rate under independence or PRDS. On the equity cross-section factors co-move — momentum and spillover share Spearman ρ ≈ 0.33 — so BH is theoretically incorrect even though it's the industry default.

Barber & Candès 2015 introduced the knockoff filter as a rigorous solution: for every predictor X_j the engine generates a synthetic "knockoff" X̃_j with the same first two moments and the same correlations to X_−j, but conditionally independent of the response. A factor is retained if its correlation with the realised forward return is meaningfully larger than its knockoff's. The filter holds under arbitrary dependence between predictors.

Concretely: if a factor cannot beat its knockoff for ≥ 30 consecutive calibration days, its weight in the Bayesian blend decays toward zero. We use the equi-correlated form (basic variant in section 3 of the paper) — appropriate for a 13-factor cross-section. Model-X knockoffs (Candès et al. 2018) become interesting if and when we extend to alternative-data factor pools.

Method: Barber & Candès 2015 (Ann. Stat. 43(5)); Candès, Fan, Janson & Lv 2018 for the Model-X extension.

14. Crisis stress testing + permutation overfit detector

Average IC across rolling windows misses the question that actually matters for institutional capital: does our 90% confidence interval cover reality during a crisis? A backtest that posts 0.05 IC on average but where the CI collapses to 30% coverage during a crash is a model that fails when capital is most exposed.

The engine reruns itself on historical crisis windows (currently 2022 rate-hike bear market; 2020 COVID and 2008 financial crisis once Yahoo history is extended) using only data available at the "as-of" date — no leakage. Reported per crisis: composite IC, observed CI coverage rate, regime asymmetry (bullish-flagged vs bearish-flagged tickers), and a permutation-test p-value contextualising the IC against a null distribution from 1,000 random label shuffles.

Permutation testing (Edgington 1969; Good 2000) is the only honest answer to the small-sample IC problem. With 26 tickers, the Spearman rank standard error is 1/√25 = 0.20, so an IC of 0.12 is only 0.6σ from zero. Without the permutation test we would mistake noise for signal.

Live results: /backtest/crisis-stress. Test source: src/lib/data/__tests__/crisis-stress-2022.test.ts and permutation-test.test.ts.

References