Prototype dashboard · trend recommendations and growth forecasts built from real Yahoo Finance historical data across major global markets.
Every asset can be analyzed with any of these eight independent models plus an Ensemble that blends them — pick one on the Dashboard's "Forecast models" section to overlay it on the chart, or rank the whole universe by it on the Rankings tab. All of them work on log-price (ln of the close), so a straight line in the model means constant percentage growth, not constant dollar growth.
The original model. Fits a linear regression to log-price over the selected lookback and projects it forward — equivalent to extrapolating the average compound growth rate. Also drives the Bullish/Bearish/Neutral recommendation and the composite ranking score, by comparing this long-term trend against recent momentum.
Forecast horizon = max(30, 25% of the lookback's span in days). Annualized growth = exp(b·365.25) − 1.
Compares a short (20-bar) and long (50-bar) simple moving average of log-price. The forecast extrapolates the short average's recent slope — smoother and less sensitive to single-day outliers than the raw linear-trend fit.
"Golden cross" when SMA₂₀ ≥ SMA₅₀ at the last bar, otherwise "death cross" — reported as context, not used in the forecast math itself. Needs at least 55 bars (longWindow + 5).
A simplified ARIMA: the log-price series is differenced once, then a first-order autoregression is fit on those differences (no moving-average term, no automatic order selection). A full ARIMA fit needs numerical optimization that's out of scope for a client-side prototype — this is a deliberate, labeled simplification of the standard method.
Forecast iterates d̂ₜ₊₁ = φ·dₜ + c one day at a time and accumulates it onto ln(price); needs ≥30 bars. No MA(q) term at all — that's the honest gap vs. a real ARIMA fit.
Double exponential smoothing on log-price (a level component and a trend component, no seasonality — irregular trading calendars don't have a clean period to model). The trend is damped (multiplied by a factor below 1 at each step forward) — without damping, a locally strong trend compounds linearly forever and produces wildly overconfident long-horizon forecasts.
The geometric-series closed form lets the model handle fractional h (sub-day chart steps) without a loop. Needs ≥10 bars.
Forecasts volatility, not price direction. Rather than adding its own line to the chart, its volatility estimate sets the width of the Ensemble's uncertainty band, and is shown as a standalone "current annualized volatility" figure.
Daily vol = √σ²ₙ; annualized vol = daily vol × √252. Persistence α+β = 0.98 < 1, so forecast variance mean-reverts toward the unconditional variance as h grows — never explodes or flatlines instantly. Needs ≥30 daily returns.
A small linear regression predicting next-day return from engineered features — the prior day's return, the gap between a 5-day and 20-day moving average, and a 14-day RSI-like momentum measure — then projected forward recursively, day by day. A genuine feature-engineering + linear-model approach computed fresh in-browser; not a stand-in for a trained Random Forest, LSTM, or Transformer, which don't fit a static client-side prototype.
Each day's predicted return is clamped to ±15%, then fed back in as the next day's "lag return" — a recursive, AR-like walk using the fitted features instead of the raw series. Needs ≥40 bars, ≥15 training rows.
No real news or text data source is connected. This score is derived purely from a shift in recent price momentum, framed the way a sentiment score would read. It exists to show where a real news/NLP signal would plug into the Ensemble — not as a genuine sentiment measure.
Label: sentiment > 0.3 → "Positive", < -0.3 → "Negative", else "Neutral". The ×0.5 factor deliberately keeps this a modest tilt on top of recent momentum, not a dominant signal — honestly, there is little real math here to be rigorous about, since the underlying signal is a stand-in for data this prototype doesn't have.
Simulates hundreds of random price paths under geometric Brownian motion, using the asset's historical drift and GARCH-estimated volatility, then reports the 10th-90th percentile of simulated outcomes — a distribution of possible futures instead of one line.
Rankings uses a lighter 80-path run without keeping per-path chart data, for speed across the whole asset universe. Needs ≥20 bars.
A weighted average of every directional model's forecast (the Sentiment Proxy is weighted at half since it's simulated, not real data). Its uncertainty band comes from GARCH's volatility forecast rather than any single model's own fit.
Whichever directional models are actually available for the asset/lookback contribute (e.g. ARIMA drops out under 30 bars) — the weights above apply to whichever subset is present, not a fixed 7-model set.
The Rankings tab scores every asset in the selected filter (class, country) and lists the top or bottom 1/5/10/25/100 by whichever metric is chosen under "Rank by":
Use Top to see the highest-scoring assets, or Bottom to see the lowest-scoring ones — useful for spotting names the model is bearish on, not only bullish on. Clicking any row expands its full Dashboard-equivalent detail — chart, recommendation, stats, forecast models, backtest — inline, without leaving the Rankings tab.
A separate ranking, built on different data than everything else in this app: real historical dividend payments, annual and quarterly financial statements (free cash flow, revenue), and analyst EPS estimates for ~900 large-cap companies across all 20 countries in the asset universe (top ~50 per country, fewer for smaller markets that don't have 50 real liquid large-caps) — not price history, and not the full ~940-asset universe.
Every company can be scored by any of 9 independent models, selectable under "Rank by" — matching how the Dashboard offers several price-forecast models rather than one blended number. Click into any company to see all 9 compared side by side.
The original model: 40% dividend growth consistency (from real payment history, penalized for any cut >10% year-over-year) + 40% sustainability (60% earnings-based payout ratio + 40% the level of FCF margin — not just its trend, so a company stuck at a deeply negative margin can't hide behind a decent payout ratio — then nudged by FCF margin trend and recent quarterly FCF momentum) + 20% current yield.
How many consecutive years has this company avoided a cut, from real payment history? A 15-year unbroken streak scores at the top of the scale — distinct from the Composite's CAGR, since the same growth rate over 3 years vs. 15 years is a very different track record.
A "cut" is any year-over-year drop >10% in the annual dividend total (see the shared cut-detection logic below). The first and last calendar years on record are dropped before counting — they're usually partial years from the fetch window, which would misread as a cut.
Recomputes the payout ratio directly from real numbers — total dividends paid (estimated from per-share amounts × shares outstanding) ÷ free cash flow — rather than using the earnings-based ratio the Composite relies on. Some companies look sustainable on earnings but are straining cash flow, or the reverse; this can disagree with the Composite when that happens. A year where FCF was negative while a dividend was paid scores at the bottom of the scale, rather than being excluded from the ranking.
| FCF payout ratio | Tier | Score |
|---|---|---|
| ≤ 50% | comfortable | 90 |
| 50–80% | moderate | 65 |
| 80–100% | tight | 35 |
| > 100% | exceeds FCF | 10 |
| FCF ≤ 0 | dividend not covered at all | 0 |
Current yield divided by the stock's own annualized price volatility (last 24 months, the same measure used in the Dashboard's forecast models). A 4% yield from a steady utility scores higher than the same 4% from a volatile miner.
Blends debt-to-equity (60%) with the payout-ratio tier (40%) — high leverage can make an otherwise-comfortable payout fragile. Not computed for Financial Services companies (banks, insurers), who don't report debt-to-equity the way industrials do — shown as not applicable rather than guessed.
| Debt/equity (%) | Leverage score |
|---|---|
| ≤ 50 | 90 |
| 50–100 | 65 |
| 100–200 | 35 |
| > 200 | 10 |
Ranks a company's yield as a percentile against other companies in the same sector within this ~900-company dividend dataset (not the whole market) — a 3% yield is mediocre for a bank but strong for a tech company, so this reframes yield by what's typical for that sector here.
Needs ≥3 sector peers with known yield in the dataset, otherwise scored as insufficient data.
Yahoo has no direct analyst dividend forecast field — only forward EPS estimates. This projects next year's dividend as (forward EPS estimate × current payout ratio), assuming the payout ratio holds steady, and nudges the result by the recent balance of analyst EPS estimate revisions (more raising vs. cutting in the last 30 days). Labeled as a derived projection, not a genuine analyst dividend call.
A single CAGR over the whole dividend history can't tell a company that grew fast early and has since slowed down apart from one that's held a steady pace throughout. This compares dividend growth in the earliest third of the years on record to the most recent third — catching a slowdown a long-run average would smooth over. Needs at least 6 years of dividend history to compute.
| decelerationPts | Trend label |
|---|---|
| ≤ -1 | accelerating |
| -1 to 1 | holding steady |
| 1 to 5 | decelerating |
| > 5 | sharply decelerating |
Compares net income growth to free cash flow growth over the same annual window, both from real financial statements. A company can report growing earnings while cash flow doesn't keep pace (working capital changes, capex timing, one-off gains) — that's a subtler, earlier warning sign than the payout ratio or FCF margin alone catch, since it questions whether the earnings behind the payout ratio are backed by real cash.
| gap (net income CAGR − FCF CAGR) | Read |
|---|---|
| ≤ -2 pts | FCF outgrowing earnings — conservative, cash-backed |
| -2 to 2 pts | growing in line with each other |
| 2 to 8 pts | earnings outpacing FCF — worth watching |
| > 8 pts | earnings far outpacing FCF — real divergence |
If either CAGR can't be computed because net income or FCF was negative at an endpoint, the model scores 15 when net income still grew (a red flag: earnings up, cash not even positive) rather than silently abstaining; if both signals are unusable it returns insufficient data.
In the main table, the "Paid in" column shows the distinct months (across the company's full history on record) it has typically paid a dividend in — e.g. "Feb, May, Aug, Nov" for a quarterly payer. Clicking into a company shows every individual payment with its yield at payout: that payment's amount ÷ the average closing price over the 3 months before the payment date, using the company's own real price history — not today's price applied retroactively.
Trailing P/E and EPS remain visible as valuation context in each company's detail view, not as part of any score — a low P/E doesn't predict dividend behavior the way payout ratio and FCF trend do.