How Sideravia works

1 · The Sideravia Score™ (0–100)

Every stock is ranked against its whole universe (e.g. the S&P 500) as a percentile blend of six pillars: Quality 25% (ROE/ROIC, margins, Piotroski F-Score), Strength 20% (Altman Z, leverage, coverage), Valuation 20% (P/E, EV/EBITDA, FCF yield — inverted, cheaper = higher), Growth 15%, Momentum 15% and Income 5%. A 75 means "better than ~75% of the universe on the weighted blend" — it is a relative rank, not a price target.

2 · The verdict (Strong · Constructive · Mixed · Weak · No view)

A deterministic weight-of-evidence read across independent signals (valuation gap, quality, trend, insider activity, analyst view). It is rule-based — the same inputs always produce the same verdict — and every supporting/opposing signal is shown next to it.

Public thesis cards summarise the same analysis with a stance label — Strong · Constructive · Mixed · Weak — derived from the composite score and the verdict together. Every label on this site, on cards and everywhere else, describes the model's conclusion, produced by the same rules for every covered stock. None of them is an instruction to you: we do not know your situation, and what you do with your money is your decision alone.

3 · Opportunities & conviction

Eight independent signal families (value gap, quality, trend, momentum breakouts, post-earnings drift, insider buying, seasonal patterns, low crowding) are checked daily. Conviction (0–100) rises with confluence — how many unrelated families agree — and the weights are tuned by our own measured signal accuracy (rank IC), not opinion.

4 · The evidence loop — we grade ourselves

Every night the ranked opportunities and scores are frozen (point-in-time, before the returns they'll be judged against — no look-ahead). We then report, on live data: signal rank-IC per metric and forward returns by tier vs SPY in the signal lab, and every published call — win rate, profit factor, alpha, best and worst — in the public track record. When a sample is too small to trust, we say so instead of showing it.

5 · The AI layer is grounded

Every AI feature (decision brief, red-team, chat, filing digests) reasons only from the stock's own data shown on the page — scores, financials, company filings — and must cite the figure it uses. It never invents numbers; when the data can't answer, it says so.

6 · How ETFs are rated

A fund is not a company, so it is not scored like one. A fund's Screen Rank and its star rating are a percentile position inside its own peer group — a category of comparable funds built from asset class, region, size and style — so a short-duration bond fund is ranked against other short-duration bond funds and never against a global equity fund. Four inputs go in: cost 35% (the published ongoing charge), one-year return 25%, three-year risk-adjusted return 25% and distribution yield 15%. Cost carries the largest weight because it is the only one of the four that is a fact about next year rather than a description of the last three.

Return and risk carry half the weight between them only because we do not yet hold tracking-difference or bid-ask-spread data — the two measures that should replace them. The research this method is built on is explicit that past return predicts future net return weakly, and we would rather write that here than let the weights imply otherwise. When those inputs arrive, this is where the weight comes from.

A rating is withheld rather than lowered whenever we cannot evidence it: less than three years of history, no published ongoing charge, a leveraged or inverse product, fewer than five comparable peers, or fewer than 60% of the four inputs present. A withheld rating reads No view. It is an absence, not a low score — it says nothing about the fund, only about what we can measure, and it never sorts or ranks as the worst value. A fund that distributes nothing has its yield dropped from the blend rather than counted as a zero.

7 · Where the data comes from

Company fundamentals, dividends, corporate actions and a growing share of prices come from one commercial market-data provider, EOD Historical Data, under a written display permission that covers publishing them here. Prices are moving across to it market by market, and a market only moves once it passes our own validation, so the markets that have not yet moved are still served by our previous price source while that work finishes.

Identity and disclosure data are public and free: ESMA's FIRDS instrument register, the GLEIF legal-entity index, SEC EDGAR filings, ESEF inline-XBRL annual reports for EU issuers, FINRA short-interest publications, and the St. Louis Fed's FRED economic series.

We compute the financial ratios and the dividend-adjusted price histories ourselves, from raw financial statements and raw closing prices, rather than publishing a provider's pre-computed figure. That is slower, and it is the reason the next section can be specific. Data is refreshed six nights a week — some fields weekly, some on their publisher's own cadence — and each card shows its freshness. Sideravia is research, not financial advice: the verdicts are inputs to your decision, and the track record exists so you can judge how much weight they deserve.

8 · What we don't know, and what we withhold

A methodology page that lists no limits is hiding them. These are ours, and each one was measured rather than estimated.

• Italy has no venue-native price feed we can license, so 183 Italian companies keep a separate price source. We do not quietly substitute a foreign secondary listing for them: those lines often trade almost no volume, and a chart built from one would look right and be wrong.
• Vendor dividend-adjusted prices are not reliable enough to publish. Across a 146-event audit of one provider's own adjusted price series, roughly one name in six carried an adjustment that was wrong, mis-dated or missing — silently, in a way no plausibility check catches. So we rebuild the adjusted series ourselves from raw closes, dividends and splits, and we check every dividend against the price gap on its own ex-date.
• Forward estimates are not available to us for most European companies. No provider we can reach sells them at European coverage. Those fields are therefore empty on those names rather than filled with a substitute, and anything built on them — a forward multiple, a forward growth rate — is empty too.
• Some ratios cannot be trusted across currencies. When a company reports its accounts in one currency and trades in another, a provider's pre-computed ratios can divide one currency by the other without converting. We found the same company's dividend yield published as 2.7% on its London listing and 0.2% on its Stockholm listing, from the identical dividend. We now recompute that yield from the underlying figures instead of publishing the provider's. The price multiples on those names are being rebuilt the same way.

Where a number is missing on this site, it is missing on purpose. An absence is never rendered as a zero, and a withheld rating never ranks as the worst one.

9 · How we check the data, every night

Every night we re-measure the whole served universe field by field — around 7,000 instruments, roughly 140 fields each — and compare the result against the previous night. Only a regression is reported: a clean night is silent, so a message means something actually moved. From launch that automated check is joined by a monthly audit whose job is to attack the data rather than confirm it, and whatever it finds is either fixed or written down here.

10 · Pre-registered retirement criteria

A track record only means something if the rules for failure are written down before the results are in. So we pre-register them: a published pick strategy is retired if, once it has at least 20 measured cohorts at its primary horizon (Long-Term: 90 days · Medium-Term & Quant: 60 days · Short-Term: 30 days), its average excess return vs the equal-weight S&P 500 (RSP) is zero or negative over the trailing 26 weeks. A retired strategy stops being published, and its full record stays on this site — archived, never deleted. Construction changes (like the sector cap added 14 Jul 2026) are disclosed the day they ship and never applied retroactively.