Preregistered 3-month measurement published
- The methodology page now carries the locked protocol for scoring engine v8.1 model estimates and analyst targets against prices realized 90 days later: sample, series, two primary hypotheses (bias; combination value with a weight fixed in advance), a company-clustered bootstrap decision rule, descriptive secondary analyses and the readout dates (interim 12 Nov and 12 Dec 2026, primary after 11 Jan 2027). Locked before any v8.1 forecast had a 90-day outcome; later changes are recorded only as dated amendments.
ADI excludes cap-pinned companies; DeepSeek series note
- AI Dispersion Index (ADI) now excludes companies where every model is pinned to the same engine cap. Their dispersion is zero by construction — an engine constant, not model agreement. In the v8.1 period this applied to TSLA, TIETO and BRK-B and had lowered ADI by up to 1.2 pp; the full history is recomputed with the same definition.
- Series note: DeepSeek has run with its reasoning mode disabled since 14 September 2026 (marked "ds2" in the bias chart); the provider also renamed the served model on 10 September. Relative to the other four models, DeepSeek did not return to its pre-31 July position: its median offset from the rest of the panel was −7.4 pp before 31 July, −0.5 pp with reasoning and +1.5 pp since. The lower absolute level in September is shared by the whole panel.
Hardening: security headers, strict types, lighter pages, cleaner data windows
- Security headers on every response (HSTS, nosniff, referrer policy, frame-ancestors, permissions policy); the framework version header is no longer sent.
- Frontend now type-checks under strict TypeScript, and CI runs the web typecheck and build on every pull request.
- Lighter dashboard: the companies table receives only the fields it renders instead of the full five-model payload for every company.
- Every server-side data request now has a timeout; a slow endpoint degrades that one section instead of stalling the whole page.
- Open Graph metadata on all pages, so shared links show the page's own title and description.
- Copy: the daily-signal wording now reads "all 5 models above/below spot" and index descriptions say "model estimates" — descriptive, not advisory.
- Accessibility: loading and chart animations are disabled when the browser's reduced-motion preference is on.
- Aggregate endpoints (indices, bias history, traits, calibration, accuracy) now exclude days whose pipeline run did not complete.
- Subscriber preferences can no longer be changed without the personal link.
- Pipeline: resumable runs, a run deadline guard, DeepSeek retries no longer repeat an exhausted reasoning budget, and the prior-day anchor uses the latest available trading day (series note: Mondays previously had no anchor).
- Accuracy horizon extended to 90 days.
Agent access: MCP server + /agents
- New: A remote MCP (Model Context Protocol) server lets AI assistants query the barometer directly as structured tools — daily snapshot, company deep dives and history, model behavior profiles, disagreement indices, the AI-vs-analyst benchmark, weekly reports, and a methodology explainer. Read-only, no API key.
- Every tool response carries the as-of date, engine version and a research-observatory disclaimer, and history tools carry machine-readable series-break metadata (engine v8/v8.1 boundaries, the 2026-07-31 DeepSeek served-model change) — so agents inherit the same interpretation guardrails the site uses.
- New: /agents page with setup instructions (Claude Code, claude.ai connectors, generic MCP clients), the tool list and example prompts; llms.txt and the API's OpenAPI description now point agents to it.
Engine v8.1 · Pure-model headline · Prompt v11.5
- Published estimates are now the pure model view: the 70/30 analyst blend was removed from the headline and preserved as a separate research series (calibrated blend). The headline never contains the benchmark it is compared against.
- Engine v8.1 conventions: mid-year discounting; targets are true 12-month values (DCF present value rolled forward one year at the cost of equity, minus dividends); the CAPM anchor's erroneous flat +1pp debt premium was removed.
- Net effect: published targets shifted mechanically upward (~10–25pp depending on market) at the version boundary — see the v8.1 markers on charts. Both series remain stored; no history was rewritten.
- Prompt v11.5: corrected WACC method (cost-of-equity ceiling, no flat debt premium), symmetric forward-looking guidance, terminal-growth placeholder clarified, null-field rules split for core vs optional fields.
- Model lineup: gpt-4o-mini → gpt-5.4-mini, gemini-2.5-flash → gemini-3-flash. GPT now runs at temperature 0.4 like the other models (it had been running at the provider default ≈1.0), and the actually-served model version is logged on every call.
Clearer numbers: which gap is which, and what 'spot' means
- Clarified: The Market Pulse and index pages now say explicitly that their gaps use pure model estimates (before the 30% analyst anchoring), while the companies table and company pages show calibrated estimates — the same stock can legitimately show two different gaps, and the labels now say why
- Clarified: Spot prices are the previous close captured at the morning pipeline run, not live quotes — marked on company pages and in the table's Spot column tooltip, so comparisons against your live broker quote make sense
Market indices: context, drivers, and an honest momentum window
- New: Each index card now carries context — 7-day change and the min–max range within the current engine regime. A within-regime percentile appears automatically once enough v8 history has accumulated (~September)
- New: 'What drove today's readings' — the companies furthest below and above the market and those with the most model disagreement, as clickable links to their pages
- Changed: Sentiment Shift (ANM) now measures the change over five trading days instead of day-over-day — single-day differences were dominated by model run-to-run noise, not sentiment
AI vs Analysts: the relative benchmark
- New: A Benchmark tab on the AI vs Analysts page scores the AI panel's own estimates (before analyst anchoring) against analyst consensus targets on realized prices — 30/60/90-day horizons, per market
- The design is deliberately relative, not a who-was-right contest: systematic bias (analysts run optimistic in US large caps, the AI panel runs conservative), the value of combining both (the optimal blend beat either alone out-of-sample in the US), and a split by up- vs down-market windows
- Every number ships with its caveats on the page: measurement window (engine v6+v7; v8 folds in as its forecasts mature), in-sample vs out-of-sample blend weights, Yahoo analyst-target lag, and 12-month targets scored at short horizons
Coming soon: vote on what we cover next
- New: /coming-soon page lists 47 expansion candidates across six markets (S&P 500, OMXH25, OMXC25, OMXS30, DAX, OMXT). US and Helsinki candidates get a weekly single-model preview estimate through the same deterministic valuation engine as the live universe; other markets are locked until the engine supports them
- New: Anonymous voting on both market packs and individual companies — votes guide what we expand to next. Promotion to full 5-model coverage stays a manual editorial decision, and expansions aren't guaranteed
- Note: Preview estimates come from a single AI model (DeepSeek) and are clearly labeled as previews — they are not comparable to the live 5-model consensus and never enter the accuracy backtests
- Behind the scenes: the preview universe runs in a fully isolated weekly pipeline and separate database tables — it cannot touch the live production data
Weekly deltas on company pages, sitewide polish & AI-citability
- New: 'What Changed (7 days)' on every company page — week-over-week change in the AI consensus estimate, median WACC, terminal growth, revenue CAGR, EBIT margin target, and model dispersion, as a sentence and a table. The same deltas are published in the page's structured data so AI search engines can cite them directly
- New: 'Explore the data' links on the About page, and the dashboard stats strip now says 'Updated <date>' explicitly
- Improved: Accessibility pass — keyboard focus management and Escape in all modals and expandable charts, descriptive labels on every chart, proper heading structure, larger touch targets and better table layouts on small screens, and a contrast pass on muted text
- Improved: Shared page headers, footers, empty states and number formatting across all pages — the same figure now renders identically everywhere; accuracy and calibration info boxes state exactly what is measured and how the scores are computed
- Fixed: Company social-share images can no longer fail with a server error — a fallback card is served if image generation hiccups; the homepage Finnish-language alternate is now declared correctly (via the sitemap) instead of pointing to the same URL as the English one
- Behind the scenes: llms.txt guide for AI crawlers, per-hit AI-crawler telemetry, hardened subscribe/email path, automated database backups, and pinned build dependencies
Measurement integrity: pure model estimates
- Changed: Accuracy, model traits, and bias history now measure each model's own estimate (before analyst anchoring) and are scoped to the current engine version — the backtest grades the models, not the analyst blend
- Changed: Dispersion definitions unified (population standard deviation) — the Disagreement index (ADI) level shifts slightly as a result
- Fixed: The model-range band on company history charts now matches the plotted lines, so the lines no longer exit their own labeled range
Easier-to-navigate research, sector links, and a fuller earnings page
- New: Weekend Read essays now link to the companies, models, and other essays they discuss, and the AI Signals page groups the long-form essays separately from the weekly reports with a jump menu — so the deeper research is easier to find and move around in
- New: Company pages now link to peer companies in the same sector, so you can move between related stocks directly
- Improved: The earnings page keeps showing the most recent reporting season between seasons, instead of slowly emptying out as reports age
- Behind the scenes: the indices, accuracy, disagreement, and model-comparison pages now publish structured data so AI search engines can cite their figures directly
AI Market Indices: one definition, one source of truth
- Fixed: The market-index charts (Valuation Gap, Disagreement, Sentiment Shift) showed a misleading drop at the latest day and could disagree with the headline figures. Charts and headlines are now calculated the same way and stay consistent across the full history
- Fixed: The dashboard's AI Market Pulse and the Indices page now always show the same values. Sentiment Shift is the day-over-day change in the valuation-gap index — read it over several days, since single-day moves include model noise
2026-06-09v1.26 · Engine v8 · Prompt v11
Analyst-cap reform: model disagreement stays visible
- Changed: We removed the caps that limited how far each model's estimate could sit from the analyst consensus. They were quietly forcing the five models to agree — on roughly a fifth of company-days all five landed on the same number. Removing them lets the models' real disagreement show, which is what this platform exists to measure
- New: A wide safety bound (roughly 0.4×–2× analyst consensus) still catches extreme outliers. Expect wider spreads between models and somewhat more negative average gaps for now — this is the reform working, not a regression
- Changed: Updated the model instructions to give clearer, non-conflicting guidance on discount rates and to stop nudging estimates toward the analyst target
Site polish & navigation consistency
- New: Subscribe call-to-action now appears on every content page (Models, Disagreement, Accuracy, Earnings, Indices, Model detail, Methodology) instead of only the dashboard, signals index, and signal detail. Same compact pattern everywhere — one click jumps to /signals#subscribe
- New: Methodology is now in the 'Switch page' dropdown so you can jump straight into the DCF documentation from anywhere on the site (it was previously only reachable from the dashboard tab bar and footer links)
- New: Model name in each company-page model card now links to that model's detail page with a small arrow — readers can jump from a company directly into the model behind a specific estimate
- New: /models hub got an ItemList JSON-LD that points search engines and AI agents (SGE, Perplexity) at all five /model/[id] subpages. Header also gained a date chip and a model count chip so the freshness signal isn't only on the dashboard
- Improved: Article detail pages (/signals/[id]) now have a proper h1 at the top instead of the article title being only an h2. Better for SEO on dedicated article URLs that already carry NewsArticle JSON-LD and per-article OG images
- Improved: Article detail pages also gained the 'Switch page' dropdown — readers landing on an article from a Twitter share can now jump to the rest of the site instead of only back to the signals index
- Improved: /disagreement and /indices used to have two stacked footer rows — a stats line ('N companies · date') above the disclaimer footer. Moved the freshness chips into the page header where every other content page shows them, so the disclaimer footer stands alone
- Improved: Heading structure cleaned up on the dashboard and on the company, earnings, and signals pages. Several content sections (Daily Brief, Movers/Top3/Flop3, Model Breakdown, Reported/Upcoming, etc.) were styled as headings but rendered as plain divs — now they're proper h2 or h3 elements. Better for screen readers and document outline
- Improved: Daily Brief's 'Generated by Claude' attribution at the bottom of the brief was hard to read (10 px, 50 % opacity) — bumped to 11 px and 65 % so the AI source is visible without dominating
- Fixed: A daily background error in the trend-streak detection step that had been logging silently. No effect on displayed data
AI-citable company pages and Weekend Read #8
- New: Dataset schema.org JSON-LD on every /company/[ticker] page with PropertyValue entries for AI consensus target, spot price, gap, model agreement, dispersion sigma, analyst consensus, analyst count, and the AI-vs-analyst difference. AI agents (SGE, Perplexity, AI Overview) can now extract structured numbers directly without parsing the page
- New: Per-ticker descriptive sentence above the page intro — '5 AI models estimate TICKER median target X (gap vs spot Y, model agreement Z). Analyst consensus A (N analysts). Experimental — not investment advice.' Bilingual; same numbers we already display, now in prose form an LLM can quote verbatim
- New: Weekend Read #8 'Which AI is best at investing?' — bilingual ~1500-word essay covering each model's behavioural wins, the effective-N herd finding (1.21 in March → 1.10 in May, with the engine-cap caveat that 19 % of company-days have all five models forced to the same number), and the AR(1) anti-anchoring discovery (all five LLMs negative, opposite of the +0.3 to +0.5 anchoring documented in human analysts). Three explicit caveats up front: calibration is WIP, backtest horizon needs months, and the panel is mid-cost tier not each provider's flagship
- Improved: Signals article markdown parser now silently skips horizontal-rule lines (---) and body-level h1 (# Title) — '##' headings provide enough section structure and the page-level title is shown above the article. Fixes every weekend-read past and future
Expandable Charts
- New: Small expand button on every chart in the unified design system (bias index, accuracy, indices ACDI/ADI/ANM, disagreement map, TP history). One click opens the chart in a portal modal up to 1300 px wide so you can read it close up. ESC or backdrop click to close, body scroll locks while open
- Polish: Expand button placement adapts to layout — bottom-right on desktop where the legend is centered, top-right on mobile where the legend wraps to the bottom edge. Avoids covering chart content on either form factor
- Polish: Disagreement Map shows only the scatter chart and legend in the expand modal — the quadrant tickers stay in the inline view as supporting context, not duplicated in the enlargement
- Polish: Bias index zoom widget (− All +) shrunk to match the 7d/14d tab buttons. X-axis date labels at the right edge no longer collide when the second-to-last natural tick lands within striking distance of the forced final tick
Unified Chart Design System
- Improved: Bias index chart redesigned (originally shipped 6.5.) — no dots, thinner 1.5 px lines, leader labels at the right edge so you can read which line is which without checking the legend, dynamic Y-axis (no wasted bottom space), +/− zoom (7d/14d/30d/60d/All), hover crosshair shows all 5 models on the same date in one tooltip, click a line or its legend to focus that model (others dim to 0.15), tap empty space to reset, mobile-anchored tooltip in the bottom-right with touch support
- Improved: Accuracy chart on /accuracy and /model/[id] redesigned to match the bias chart — no dots, thinner lines, hover crosshair shows all 5 models for the same date in one tooltip, click a line or its legend to focus that model (others dim), tap empty space to reset
- Improved: Company-page TP history chart now supports click-to-focus on individual models (line goes 2.4px and others dim to 0.15) — spot, consensus, and the bear/bull band stay at full opacity as reference context
- Improved: Index charts on /indices (ACDI / ADI / ANM) — dots removed, hover crosshair with viewport-clamped tooltip showing the index label and value, touch-friendly
- Improved: Disagreement Map tooltip now stays inside the viewport on small screens and supports tap on mobile (anchored to the chart's bottom-right)
- Internal: Extracted shared chart helpers (date/percent formatters, Y-axis tick picker, leader-label collision solver) and a reusable crosshair-tooltip component, used by all five chart types — keeps style and behaviour identical across the site
Models Hub & Weekend Read #8
- New: /models hub page — side-by-side scorecard for all 5 AI models, with personality cards, cross-model bias chart, and the stocks they disagree about most. Tabbar entry added
- New: Weekend Read #8 — 'Same prompt, five answers' — a five-clue investigation into why identical prompts produce divergent answers across models, even after heavy standardisation. Bilingual
- New: /models-performance API endpoint — composite scorecard data for the models hub
Easier Navigation & Stronger Sharing
- New: Company picker dropdown on company page header — jump from one stock to another without going back to dashboard
- New: 'Switch page' dropdown on Indices/Earnings/Accuracy/Disagreement/Signals/About — same lateral navigation pattern site-wide
- New: Per-company social share images — when you share /company/META, LinkedIn/X show a rich card with ticker, AI estimate, gap, and agreement
- New: Business summary on company pages — short description of what the company does (placed below fundamentals)
- Improved: Whole row clickable on dashboard companies table — not just ticker name
- Improved: Earnings page — company names clickable, redundant arrow column removed
- Improved: Disagreement Map quadrant tickers now link to company pages
- Fixed: Sitemap was using wrong API field, never reflecting actual universe — now updates correctly
Signal Purity & Index Interpretations
- New: Raw vs calibrated agreement on company pages — distinguishes genuine model consensus from cap-induced agreement (when engine caps pulled estimates together). ⚠ flag when agreement is partly artificial
- New: Plain-language interpretation under each AI Market Pulse card (ACDI, ADI, ANM) — one sentence explaining what the value means
- New: Top-3 disagreement badge on company page header — surfaces stocks where AI models genuinely can't agree
- Fixed: Trading days count now consistent between dashboard and indices page
- Fixed: Methodology page now consistently labels engine v7 (was v6 in subtitle)
Reverse DCF & Earnings Hit Rate
- New: 'What Would Need to Be True?' on company pages — solves backwards what assumptions (CAGR, margin, WACC) the current stock price implies, side by side with AI consensus
- New: Earnings hit rate summary on /earnings — total reported, converged/diverged/neutral counts, directional accuracy at a glance
- New: Stock 3d column in earnings reported table — 3-day post-earnings reaction (more stable than 1-day per Bernard & Thomas 1989); when 1d and 3d directions disagree, both are shown
Today's Signals & Earnings Reactions
- New: Today's Signals on dashboard — daily prioritized list of the most interesting moves (earnings, unanimity, big gaps, movers)
- New: Earnings reactions — see whether stock moved toward or away from AI estimate after reporting (converged/diverged with 1-day price reaction)
- New: Weekend Read #6 — 'Do AI Models Think or Repeat Patterns?' analyzing round-number bias, model correlation, and herd behavior
- New: Each Signals article now has its own page and URL for sharing
- Improved: Dark mode readability and visual hierarchy across the dashboard
AI Earnings Calendar
- New: /earnings page — see how AI models position before earnings and react after. Upcoming dates, AI gap, model agreement, EPS/revenue estimates
- New: Earnings data in pipeline — next earnings date, analyst EPS and revenue estimates fetched from Yahoo Finance daily
- New: Dashboard navigation link to Earnings page
Page Intros & Research Background
- New: Page intro cards — dismissible first-visit intro on 6 pages explaining what each page shows
- New: Research Background section on about page — academic context with linked references
- Fixed: Daily Brief now works on weekends and holidays
- Fixed: Tooltip overflow scroll eliminated across all pages
Weekend Read #5 & Frontend Audit
- New: Weekend Read #5 — 'How Often and How Much Do AI Models Change Their Minds About Stocks?' with temporal variance analysis per stock and sector
- Improved: Disclaimer footer added consistently to all content pages
- Improved: Full bilingual coverage — remaining hardcoded strings moved to translation system
- Fixed: Dashboard crash prevention when API is unavailable
Onboarding & Weekend Read #4
- New: 4-step guided onboarding — 'Start here' now opens an interactive walkthrough on the dashboard instead of linking to FAQ
- New: Weekend Read #4 — 'One Month In: What 2,760 AI Valuations Taught Us' with XOM/Iran case study and model personality analysis
- Onboarding highlights model personalities (Claude conservative, GPT volatile, DeepSeek reliable) and AI vs Analyst Disagreement Map
- Fixed: Daily Brief pipeline bug (PostgreSQL UPDATE syntax, JSONB cast) — briefs now generate correctly
AI Daily Brief & Smart Alerts
- New: AI Daily Brief — Claude-generated 2-3 sentence narrative summary on dashboard, updated with each pipeline run
- New: Narrative alert types — quadrant changes ('NOKIA moved to Consensus'), 5-day streaks, and Top 3 entries
- New: Weekly email cap — max 2 alert emails per subscriber per week, prevents inbox fatigue
- New: Manual editorial notification — notify_editorial.py script to trigger subscriber emails for new articles
- Improved: Alert thresholds raised — consensus shift 15→20%, disagreement 40→50%, unanimous gap 20→25%
- Improved: Disagreement Map ghost dot toggle — cleaner legend, checkbox at end, opacity-based on/off
Design System & Disagreement Map Fix
- Fixed: Disagreement Map now shows real data — X-axis changed to AI vs Analyst gap
- Unified design language across all pages: consistent tabs, buttons, badges, and hover effects
- Mobile: language switcher moved to title row, improved tab layout
Disagreement Map & User Experience
- New: AI vs Analyst Disagreement Map — scatter plot comparing AI model disagreement with analyst consensus spread (/disagreement)
- New: Subscribe prompt after AI Market Pulse for easier signup
- New: Premium interest option in subscription form
- New: 3 FAQ questions linking to Methodology, Changelog, and Accuracy pages (38 total)
- New: Weekend Read #3 — 'When the Market Moves Toward AI'
- Improved: Alert thresholds raised (consensus shift 8→15%, disagreement 30→40%), bias bars added to emails
- Improved: Social sharing fixed — OG image now resolves correctly (was pointing to localhost)
- Improved: v7 engine markers on Bias Index, Indices, and TP History charts
- Fixed: Title duplicates removed from all pages, model card number overflow, ### markdown headings in signals
2026-03-24Engine v7 · Prompt v10
Pipeline Calibration & Reliability Upgrade
- Engine v7: sector-specific ROIC calibration, healthcare margin bounds widened (45%), technology margin bounds widened (55%)
- Prompt v10: dedicated sector guidance for all 10 sectors (was 3) — healthcare, technology, energy, telecom, consumer, industrials, materials
- Temperature harmonized to 0.4 for all models — enables fair comparison of model volatility
- JSON repair + retry with 5s delay: model validity improved from ~83% to ~100%
- DRY_RUN mode: safe pipeline testing without affecting production data
- Cap-induced agreement detection: distinguishes genuine model consensus from safety-cap convergence
Dashboard Redesign
- New stats strip: live counters for Models, Trading Days, and Estimates replace the static trust bar
- Cleaner header: removed redundant chips (AI CONSENSUS, tickers), simplified to 4 visual rows
- Tab dividers between model tabs and navigation links for visual clarity
Shareability & Navigation Polish
- Share buttons (X, LinkedIn, copy link) on dashboard, company pages, and signals — framed as AI market insights, not stock picks
- Award badge tooltips on model pages: hover to see what each badge means (EN+FI)
- Navigation consistency: methodology page now links back to dashboard (was linking to About)
Data Quality, SEO & New Features
- Pipeline retry logic + JSON repair: model validity improved from 83% to 91%
- Q&A page with 35 bilingual questions across 6 categories (incl. Research & Findings)
- Currency symbols (€/$) on all price displays
- SEO: structured data (FAQPage, BreadcrumbList, NewsArticle), dynamic meta descriptions with 'disagreement' keyword
- Ticker search in companies table
- ADI (Model Disagreement) time series chart — was placeholder, now shows real data
- Shrinkage disclaimer + valid model count (X/5) on company pages
- Weekend Read #2: Claude vs GPT — Two AI Analysts Compared
- Bias Index badges clarified: 'Most Positive/Negative' with time window context
- Accuracy page info tooltip restructured with per-column explanations
- Fixed: Gap % display (×100), broken RSS feeds cleared, FAQPage schema duplicate removed
UX Polish & Mobile Optimization
- Hero tagline explains the site's purpose to new visitors
- Dashboard sections are now collapsible (▾) — cleaner layout, less scrolling
- Mobile: chart sections collapsed by default, consistent card widths, tab bar split into model tabs + navigation
- SEO: case-redirect middleware (301), hreflang tags (EN+FI), sitemap now includes all 24 company pages
- Fix: Gap % on company page now displays correctly (was showing 0.4% instead of 43%)
2026-03-17Engine v6 · Prompt v9
Bayesian Calibration & Signal Architecture
- Bayesian calibration: 70% DCF + 30% analyst consensus blending reduces systematic bias
- Deterministic terminal growth per market and sector (replaces LLM guesswork)
- Tiered analyst TP caps: mega-cap US +70%/−35%, cyclical +55%/−40%, standard +50%/−40%
- AI indices (ACDI, ADI) now computed from pre-calibration DCF signal
- Forward-looking prompt: LLMs instructed to forecast, not copy historical figures
- Market-specific equity risk premium: Finland 4.5%, USA 5.0%
Weekend Read & Editorial System
- Bilingual editorial articles (Weekend Read) on AI Signals page
- Dynamic sitemap, canonical URLs, SEO improvements
AI Market Indices & Enhanced Signals
- Three AI indices: ACDI (valuation gap), ADI (disagreement), ANM (sentiment shift)
- Weekly AI Signals reports with trend streaks and sector rotation
- Confidence bands (bear/bull cases) on company pages
GDPR, Subscriptions & Production Hardening
- Email subscriptions with double opt-in (Resend API)
- Rate limiting, ISR caching, CORS restriction
- GDPR-compliant privacy policy
Valuation Engine v5 & Model Accuracy
- ROIC-based FCF conversion, TV share cap, CapEx normalization
- Model accuracy backtesting with directional accuracy scoring
- Calibration scoring (per model, /100 format)
Platform Launch
- 5 LLM models (Claude, GPT, Gemini, DeepSeek, Grok) generating valuation assumptions
- Two-stage FCFF DCF + Excess Return on Equity for financials
- Dashboard with model tabs, company pages, and bilingual UI (EN/FI)
All content is generated by AI models and may contain errors. This is an experimental tool — not investment advice, research, or recommendation.
Terms of Use ·
Privacy Policy