Classical quantitative techniques, applied transparently to US equity data. We don't sell secret sauce because there isn't one — what we sell is honest execution of well‑understood methods, walk‑forward validated, with real accuracy numbers visible to every user.
Most stock‑prediction apps treat their algorithms as a mystery. We do the opposite. The methods below are well‑published in quantitative finance — k‑nearest‑neighbors on time series, HAR‑style volatility models, cross‑sectional ranking on standard factors. Our differentiation isn't a hidden trick; it's that we measure how well these methods actually work, publish that number, and refuse to overclaim. If a panel doesn't show real edge over a naive baseline, the app says so.
For a chosen forecast horizon H (5, 10, or 22 trading days), the lookback window length is set to L = max(20, 4H). The last L daily log‑returns are z‑normalized (centered and divided by their own standard deviation) so two windows are compared on shape, not absolute magnitude. The same z‑normalization is applied to every historical window in the past five years for that stock; Euclidean distance between the current window and each historical window gives the similarity. The k closest historical windows are the "analogs."
For each analog, we record what actually happened to that stock over the H trading days that followed the analog window. The collection of those forward returns is the output — a small empirical distribution, summarized by its mean, its 25–75% interquartile range, the proportion that finished positive ("up‑rate"), and a histogram. A kernel similarity score exp(−d² / 2σ²) with σ set to the median candidate distance gives each analog a comparable "closeness" weight.
A distribution of forward outcomes from the closest historical analogs, plus the stock's overall base rate for comparison. The honest read is the spread (25–75% range), not the mean.
Walk‑forward rank‑IC across 200+ US large‑caps and three horizons. The raw‑path variant beat a snapshot‑based alternative at every horizon — that's why it's what ships.
k times this stock looked similar" — not as a prediction. The mean of the distribution is not a forecast; the spread is the honest content.
This is the most trustworthy of the three. We compute realized volatility over rolling windows at three frequencies — daily, weekly (5‑day average), and monthly (22‑day average) — the classical HAR ("Heterogeneous Autoregressive") feature set introduced by Corsi (2009). These features feed a logistic regression that outputs the probability that realized volatility over the next H trading days will be high (above this stock's trailing median) versus calm.
Training is strictly walk‑forward: at every test point the model uses only data available up to that point, never the future. We re‑train on an expanding window and evaluate on the next out‑of‑sample period. The accuracy you see in the app is what the model actually achieved on data it never saw during training — not in‑sample fit.
A probability of high vs calm volatility, the model's real out‑of‑sample accuracy, an AUC score (0.5 = no skill, 1.0 = perfect), and the two naive baselines it must beat for a "real edge" verdict.
Beats two baselines: Persistence ("today's regime continues") and Always‑calm ("guess the more common regime"). If we don't beat both, the verdict is not green. AUC is reported alongside accuracy because accuracy alone can mislead on imbalanced regimes.
For every stock in your loaded universe we compute five standard factors, then rank each stock cross‑sectionally as a percentile. The factors: price momentum (12‑month minus 1‑month return), trend (price vs trailing moving averages), low‑volatility (inverse realized vol), RSI‑14, and 52‑week‑high proximity. The output is a position on a 0–100 scale per factor, not a single composite score.
The factors themselves are textbook technical indicators with decades of published literature; we make no claim that any one of them, in isolation, predicts returns reliably for an individual stock. What ranking provides is context: where does this stock sit versus its peers right now, and how does that position compare to its own history?
Five percentile dots showing where the stock sits on each factor versus the loaded universe. Clickable leaderboards expose the top and bottom names on each factor.
Percentile ranks are by construction — they reflect the universe you load. We don't backtest these as standalone trading signals (and you shouldn't either). Ranking is best read as descriptive context for the other panels.
A short, honest list — because what an analytics tool omits is as important as what it includes.
The methods we use are well‑documented in the quantitative finance literature. A few starting points if you'd like to read more:
Markets are hard. Honest analytics on top of public price data can give you useful context — distributions, regimes, rankings — but they cannot reliably tell anyone what to buy. Anyone who claims otherwise is selling something else. Stock View's value is in showing you what the data actually says, measuring how well it knows, and refusing to overclaim. That's the whole product, and that's why the methodology is published rather than hidden.