Professional-grade quantitative equity research, built on public data, available without a $25,000/year Bloomberg terminal.
Why this exists
The math behind professional quant investing isn't secret. Fama-French 1992 is public. Sloan 1996 is public. Asquith-Pathak-Ritter 2005 is public. The papers that describe the edges institutional funds trade on are literally citable from this page. What's not public is the engineering - the pipeline that turns thirty years of academic findings into a daily score you can actually use.
Framler is that pipeline, built once and exposed for free. A multi-factor stack, a confluence pattern library, Bayesian regime detection, Bayesian Phase 3 failure prediction, conformal prediction intervals. The code implementations exist because we wrote them from the primary literature, not because we licensed them from anyone.
The product thesis: if retail investors had the same math institutional funds use, they'd size positions more honestly and lose less money to overtrading, consensus-chasing, and narrative investing. The best way to test that thesis is to put the math in front of them and see what happens.
Who built it
Framler is independent quantitative research, founded and self-funded by Illya Kuznetsov — built since April 2026 to do one thing the incumbents won't: publish honest, measurable stock research at a retail price and let anyone check the record. Every factor is grounded in its primary academic paper, accuracy is measured and published as it matures, and every day's calls are anchored in Bitcoin so the track record can't be quietly rewritten. The moat is that discipline — and the refusal to overclaim — not pedigree.
Bootstrapped by choice so far — no outside capital yet. We'd rather build the live track record than raise on promises. That said, we're open to a small, aligned pre-seed with an investor who shares the honesty-first model — one that funds engine breadth and data depth without pressuring the honest pricing and timeline on the pricing page.
Independent does not mean unreviewed. Every commit runs through 800+ unit tests, the engine is covered by behavioural invariants checked live against production data (visible at /coherence), every factor traces to a peer-reviewed paper on the methodology page, and external math review is planned before the paid Pro tier launches. Until then we publish what works and what doesn't - including the live calibration progress and the module-by-module status page.
Amihud 2002 illiquidity premium as the final factor. NLP scoring (Loughran-McDonald), short-interest factor, Fama-French regression in production. Engine v6-v8 series.
Two-round audit closed 4 CRITICAL + 13 HIGH + 24 MEDIUM findings. CSP nonce with strict-dynamic, Kelly sizing moved server-side, RLS hardening. 688 invariant tests.
Full domain cutover from deepvane.com. Brand v5 wordmark (FRAMLER with violet l-and-r accents). Counter-AI repositioning across hero, methodology, story.
14 May 2026
Forecaster v2.1 / v2.2 + version consolidation
Piecewise alpha (Moskowitz 2012), VIX-conditional variance, pattern-conditional dampener, width-aware mean. Internal v11 → public v2 consolidation. 696 invariant tests. Programmatic research pages expanded (ticker, sector and glossary pages) with canonical controls.
isPriorMode flips at 00:00 UTC. Conformal calibration window opens — 30-day forward returns from the April snapshots become measurable. UX-marathon: ModeSelectorHero, sparkline saga, terminology unification (Conviction/Confidence → "Agreement").
17 May 2026
First adaptive weights persisted
1d and 7d horizons calibrate on 1091 samples across 3 regimes (all / risk_on / transition). Measured ICs 0.055–0.10 replace literature priors via inverse-covariance Markowitz with copula-blend covariance. Conformal halfwidth calibrated on 485 samples (7 days); the constant itself stays private. Engine LIVE.
1 Jun 2026
HAR-RV variance forecast activated
Corsi 2009 heterogeneous-autoregressive realised volatility blended into the variance pipeline — a measured ~10% out-of-sample MSE improvement over the Huber baseline across 56 independent months; the blend coefficient stays private. Flag-gated activation with a one-command rollback tag.
3 Jun 2026
Forensic screens — Altman Z + Beneish M
Bankruptcy-distress (Altman 1968) and earnings-quality (Beneish 1999) screens computed live from SEC filings on every ticker page. Display-only and deliberately independent of the Framler score — extra context, not wired into the engine.
●Jul 2026
30d horizon maturing
The 30-day sample threshold is reached and adaptive weights extend to the medium horizon — but only a couple of independent 30d windows exist so far, too few to call the measured IC trustworthy. Still maturing.
Projected
~15 Sep 2026
Full 90d calibration
All four horizons (1d / 7d / 30d / 90d) on measured ICs across all regimes. Engine fully on out-of-sample alpha measurement. First honest 90-day walk-forward Sharpe ratio publishable.
Q3 2026
Pro tier + API
Alerts, historical depth, Shapley per-ticker, REST API access. Enterprise tier for funds and quant teams.
Principles
Honest before impressive
If a number isn't measured, we say so. The interval width is published alongside the score. The calibration state is displayed on every ticker. We'd rather under-promise on track record and deliver than the opposite.
Composition, not opacity
Every factor traces to a peer-reviewed paper. The citations are on the methodology page, on every pattern detail page, on every blog post. The moat is how the layers compose, not that the ingredients are secret.
Free as a weight on the industry
Good quantitative research gated behind institutional paywalls is a market inefficiency. Making it free - and well-engineered enough to actually use - is the bet.
Lean by design
No institutional budget, no bloat. The whole stack is engineered to deliver professional-grade research on public data - and it is held to that standard on every page.