A field guide to AI stock scores

Why this is worth reading carefully

In April 2026, Investing.com surveyed retail investors and found that roughly 62% now use AI tools in their research process. Only 23% trust the outputs without verifying them against other sources.¹ The remaining 39% — the use-but-verify group — describe the audience this essay is written for. If you are paying for one of the five products below, or thinking about it, the question is not whether the AI score is “good” in the abstract. The question is what it measures, how to read it, and what to ask before treating it as decision-relevant.

Five scoring systems dominate the retail AI-scoring market in 2026: Danelfin's AI Score, Kavout's Kai Score, Seeking Alpha's Quant Rating, Simply Wall St's Snowflake, and TipRanks' Smart Score. Each costs €10 to €600 a year. Each promises a single, scannable read on a stock. Each measures something materially different. Treating them as interchangeable — five flavours of the same idea — is the most common mistake retail investors make in this category.

Six things vary across the set. The numerical range and what high means; the input mix (fundamentals, technicals, sentiment, analyst data, alt data); whether methodology is public; whether weights are disclosed; how performance is reported; and whether anyone outside the company has validated it. The next sections work through each tool.

Danelfin — AI Score, 1–10

What it measures. Danelfin ingests roughly 10,000 features per stock per day across three families: about 600 technical indicators (RSI, MACD, momentum, volatility), 150 fundamentals (CROCI, EBITDA, P/E, ROA), and 150 sentiment features (analyst price targets, insider transactions, short float). A tree-ensemble model trained on the trailing 252 trading days produces a daily score between 1 and 10, representing the model's estimated probability that the stock will produce returns higher than the S&P 500 over the next 90 days.² A 10 is labelled in the company's terminology as the strongest bullish tier; a 1 the strongest bearish.

Transparency. Mixed. Danelfin publishes factor families but not feature-level weights — the tree ensemble is a structural black box. What it does publish, unusually, is an independent audit page with Newey-West HAC-adjusted t-statistics, Harvey-Liu-Zhu multiple-testing thresholds at |t| ≥ 3.0, survivorship-bias-corrected delistings at −100%, and Fama-French 5-factor residual alpha.³ Live tracking began 1 July 2025. Numbers from before that date are backtests across three non-overlapping windows (2017–19, 2020–22, 2023–25). This is the most methodologically self-aware audit page in the set.

Published performance. The headline marketing number is Danelfin's flagship long-only strategy returning +376% versus the S&P 500's +166% over January 2017 to June 2025. AI Score 10 US stocks have a stated 21.05% annualised alpha after 3 months; Score 1, −33.28% annualised alpha. The win rate of Score ≥ 7 beating the S&P over 90 days is reported as 70.24% across a 900-stock US universe.²

Independent validation. None in the peer-reviewed sense. No academic paper evaluates Danelfin's AI Score as of May 2026.

Coverage. ~900 US-listed stocks historically, with the European universe expanded in February 2026 from the STOXX 600 to 5,500+ stocks across European exchanges.

Pricing. Plus €29/mo, Pro €79/mo, Elite €179/mo, with roughly 30% off on annual billing.

Kavout — Kai Score, 1–9

What it measures. Kavout describes 200+ factors: fundamentals (earnings quality, revenue, balance-sheet items), pricing and volume data, technical indicators (RSI, Z-Score), and alternative data including news and social sentiment, analyst upgrades, and insider transactions. The model is a combination of regression, classification, deep learning, and reinforcement learning. Reinforcement learning, the company says, allows the model to dynamically adapt as new data arrives.⁴ The output is an integer 1–9 indicating estimated outperformance probability over one to three months.

Transparency. Lower than Danelfin. Kavout discloses factor categories and model families but not weights, model architecture, training-set composition, validation procedure, or holdout test results. The performance claim is gated qualitatively: an “estimated alpha of 4.84%” appears on the public page with no period, sample, or confidence interval. The footnote reads: “K Score alpha is an estimate only. Actual alpha results may vary.”⁴

Independent validation. None academic. FactSet distributes the K Score datafeed, which provides some institutional credibility but is not validation of predictive accuracy.

Coverage. US equities and ADRs (history since 2003), China A-shares, UK, Germany. Non-US coverage is thinner.

Pricing. Free tier with 10 research credits per month; Pro $16/mo; Premium $39/mo. Annual billing reduces the headline numbers.

Seeking Alpha — Quant Rating, 1.0–5.0

What it measures. Five factor grades — Value, Growth, Profitability, Momentum, EPS Revisions — each graded A+ to F using 100+ metrics compared sector-relative (not absolute). The model is recalculated daily before market open. The result is a 1.0–5.0 numerical score paired with categorical labels.⁵

Transparency. The most explanatory methodology document in the set short of Simply Wall St. The FAQ states explicitly that the rating is not a simple average of the five factor grades; certain factors carry higher weight; and disqualifying thresholds exist (D+ or worse in Growth, Momentum, or EPS Revisions disqualifies a stock from the strongest-bullish tier; D− or worse in Value or Profitability does the same). Exact weights are not disclosed. Notable choices the methodology explains: no DCF (rejected as too sensitive to terminal-growth assumptions) and no historical multiple comparisons (rejected because multiples shift with rates and sentiment). The model is sector-relative — it ranks within a sector, not across sectors.

Published performance. 2024 calendar year: Quant strongest-bullish stocks returned +37.15% versus the S&P 500's +12.75%. 2025: +28% versus +11%. Long-term, 2010–2024: +1,754% cumulative versus +385% for the S&P 500.⁶

Independent validation. Unique in this set: a working paper by Russell Jame at the University of Kentucky (April 2024) analysed 10,000+ stocks rated 2016–2022 and found the Quant Ratings strongly predict future returns, with “very bearish” stocks producing returns approximately 32 percentage points below the S&P 500 annually.⁷ One caveat: the paper was disseminated through Seeking Alpha itself, and whether it lands in a top journal remains to be seen.

Coverage. ~5,600 US-listed stocks. Excludes stocks without analyst coverage and foreign exchange listings.

Pricing. Premium $299/yr (often discounted to $269); Alpha Picks $449/yr; Bundle $639/yr; Pro $2,400/yr.

Simply Wall St — Snowflake, five axes

What it measures. Not a single number — a five-axis radial visual: Value, Future Performance, Past Performance, Financial Health, Dividends. Each axis is composed of six binary checks (the share trades at >20% below DCF; revenue growth exceeds the weighted market; D/E < 40%; etc.). A stock passing more checks on an axis paints a larger green sector; failing them paints a smaller red one.⁸

Transparency. The full model is open-source on GitHub — every check, threshold, and industry-specific substitution is documented inMODEL.markdown.⁹ This is the highest transparency in the set. A user can replicate it.

Published performance. None. Simply Wall St positions the Snowflake as a diagnostic visualisation rather than a return-predictive score, and so does not publish “high-Snowflake” backtest returns.

Independent validation. Not formally peer-reviewed, but the open model has been replicated and critiqued in public forums.

Coverage. Broadest in the set — 120,000+ stocks across 90+ markets, sourced from S&P Global. This includes US, Canada, UK, all of European exchanges, Australia, Japan, Hong Kong, and dozens of smaller markets.

Pricing. Premium ~$120/yr; Unlimited ~$180/yr. Annual-only on most tiers.

Known weakness. The checks are industry-agnostic. The same 6×5 grid is applied to a bank, a utility, a software company, and an emerging-market miner — but healthy capital structures differ materially across those industries. A user reading the Snowflake without context can confuse “passes thresholds set for the average firm” with “actually well-positioned for this industry.”

TipRanks — Smart Score, 1–10

What it measures. Eight inputs (not ten — a common confusion): Wall Street analyst consensus weighted by track record; corporate insider activity; financial bloggers; individual investor sentiment; hedge fund manager activity; news sentiment; technical factors; fundamentals.¹⁰ The output is a 1–10 score.

Transparency. Conflicting on the public page. TipRanks help content describes the eight inputs as contributing equally at 12.5% each. Other public-facing pages and reviewers describe the score as weighted by historical predictive value, with weights undisclosed. The discrepancy is itself a transparency red flag.

Published performance. Smart Score 10 stocks are stated to have produced cumulative returns between +217.9% and +430.3% since 2016 (the range depends on which dated page is read), versus ~246% for the S&P 500 over a comparable period. The claim is backtested, with TipRanks explicitly disclaiming live future performance.

Independent validation. None academic.

A reflexivity concern. One of the eight inputs is “financial bloggers.” Some of those bloggers publish their analysis directly on TipRanks. This means part of the input feeding the Smart Score is generated on a platform whose business depends on the score's apparent credibility. The loop is small and probably second-order, but it exists in no other score in this set.

Coverage. ~75,000 stocks, plus tracking of 7,500+ analysts and 96,000+ financial professionals. ETFs, crypto, options, and FX also covered.

Pricing. Smart Portfolio $215/yr; Premium $360/yr; Ultimate $600/yr.

The transparency spectrum

Across the five scores, methodology transparency ranges from “full model on GitHub” to “we will tell you the factor categories.” Ranking from most to least open:

The ordering is not the same as the ordering by quality. Simply Wall St is the most transparent but does not claim to predict returns. Danelfin claims the most rigorous out-of-sample design but is opaque at the feature level. Reading transparency and predictive validity as the same axis is the second most common mistake in this category.

Six methodological weaknesses they share

Five different scoring approaches, but the same underlying problems show up across all of them. Worth being clear-eyed about each.

1. 1. Backtest contamination. Every score publishes a backtested headline number. The academic literature on machine-learning stock scoring — and the Stanford GSB simulation of an AI fund manager published in June 2025¹¹ — indicates that even with careful out-of-sample design, live-deployment edge tends to be 10–20% smaller than the backtest. Danelfin's live tracking began on 1 July 2025. For the others, “live” performance is harder to disentangle from reconstructed history.
2. 2. Survivorship bias. When a tested universe drops delisted or acquired stocks, returns improve mechanically. Of the five, only Danelfin's audit page documents −100% treatment of delistings explicitly. The others do not disclose their treatment of survivorship.
3. 3. Sector-relative versus absolute confusion. Seeking Alpha and Simply Wall St rank within sector — meaning there is always a top-quartile energy stock, even when energy is bleeding. Danelfin and Kavout aim for absolute outperformance probability. TipRanks is somewhere in between. Users frequently treat all five as today's highest-rated names without registering the difference.
4. 4. Forward-return horizon. Danelfin trains against 90-day forward returns. Kavout against 1–3 months. Seeking Alpha against unspecified forward windows. TipRanks against an ambiguous horizon. Simply Wall St is not a forward-return predictor at all. If your holding period is two years, a score trained on three months is not measuring what you need.
5. 5. Equal-weighted versus market-weighted backtests. A backtested portfolio of top-decile stocks weighted equally tilts heavily toward small-cap names and inflates apparent alpha. None of the five fully discloses its weighting convention in public marketing materials.
6. 6. Factor independence (or the lack of it). Subscribing to two or three of these scores feels like diversification but mostly is not. They all consume similar upstream factor universes — analyst revisions, momentum, valuation multiples. The correlations between their top-decile choices on any given day are higher than the marketing implies.

Seven questions before paying

1. 1. When did live tracking start? Backtested numbers and live numbers are not the same animal.
2. 2. What forward horizon was the model trained against? A 90-day model is not the right instrument for a five-year position.
3. 3. Sector-relative or absolute? Does the score always point you at the best of energy in a sector rout, or does it stand down?
4. 4. Equal-weighted or market-weighted backtests? Equal-weighted tilts small-cap and overstates alpha.
5. 5. What does the tool refuse to score? Coverage gaps reveal model limits. Seeking Alpha drops foreign listings and stocks without analyst coverage. Stock Rover drops European exchanges entirely.
6. 6. Can I see point-in-time historical scores? Without per-day historical scores, you cannot verify that the backtest is not reconstructed history.
7. 7. Has anyone outside the company evaluated it? Of the five, only Seeking Alpha has independent academic work (Jame, Kentucky, 2024) — and even that was disseminated through Seeking Alpha itself. Treat this asymmetry seriously.

What an honest score looks like

Simply Wall St publishes its model on GitHub. Anyone with three hours can read it, find the binary thresholds, and decide whether the framework matches their investment style. They do not publish forward-return claims because the score is diagnostic, not predictive — and they are honest about that. This is the bar. Most of the others are not at it yet.

Seeking Alpha's Quant Rating is the only score in this set with independent academic validation, with the caveat that the paper was disseminated through Seeking Alpha. The next time an AI score's company points at a backtest, ask which university produced the working paper.

The Stanford GSB simulation, published June 2025, is worth reading in its own right. A model was trained on roughly 170 simple, public variables (Treasury rates, credit ratings, sentiment from earnings calls) and asked to rotate half of each fund's holdings quarterly across ~3,300 actively managed US equity funds over a 30-year window. The model added roughly $17M of quarterly alpha across the set — six times the human alpha — and beat 93% of human managers.¹¹ The interesting thing is not the magnitude. It is that the AI's edge came from extracting predictive value from boring, public variables using better statistical machinery, not from exotic data. Most retail AI scores claim the opposite — a unique alternative-data edge. The Stanford finding indicates the marketing might have the story upside down.

Acutic's posture

Acutic does not produce a single composite score. Instead, five named analyst agents — quality, valuation, technical setup, news, risk — analyse the same instrument independently, and the user sees all five outputs side by side, with the evidence and the disagreements. The methodology page is public. Every analyst output is labelled as AI-generated at the point of display. The product produces research, not personalised investment advice.

If a single-number score is what you want, several of the products above will do the job — provided you read it with the seven questions above ready. If you want to see the reasoning, see where the analysts disagree, and decide for yourself, that is a different shape of product. See the methodology page and the AI transparency page.

Notes

1. ¹Investing.com, How Retail Investors Are Using AI in 2026, April 2026 retail-investor survey.
2. ²Danelfin methodology and reported performance — danelfin.com/how-it-works.
3. ³Danelfin audit page — audit.danelfin.com.
4. ⁴Kavout K Score description and pricing — kavout.com/k-score and kavout.com/pricing-plans.
5. ⁵Seeking Alpha Quant Ratings FAQ — help.seekingalpha.com / Quant Ratings FAQ.
6. ⁶Seeking Alpha Quant performance — about.seekingalpha.com / Quant performance.
7. ⁷Jame, R., et al., Quant Ratings and Future Returns, University of Kentucky working paper, April 2024. Coverage on Seeking Alpha: Seeking Alpha summary.
8. ⁸Simply Wall St Snowflake explainer — support.simplywall.st / Snowflake.
9. ⁹Simply Wall St Company Analysis Model — github.com / SimplyWallSt / MODEL.markdown.
10. ¹⁰TipRanks Smart Score glossary — tipranks.com / glossary / smart-score.
11. ¹¹deHaan, E., Noh, S., Lee, C., Liu, M., Stanford GSB Insights, June 2025 — gsb.stanford.edu / AI analyst.

Related reading on the Acutic blog: the previous essay, what changes for AI investing tools on 2 August 2026, covers the regulatory checklist that complements the methodology checklist above. Request early access.

Acutic provides investment research and educational content. It is not investment advice. Acutic operates as a non-personalised investment research and analysis service under MAR Art. 20 / § 85 WpHG. Past performance does not predict future results.