Learn DeepVane in plain English

A 0-100 number that summarises how attractive a stock looks across 12 different academic angles.

Each of the 12 factor families (quality, value, momentum, etc.) scores the stock 0-100. We z-score them across the universe, weight each by its prior reliability, and combine into one composite. 50 = neutral / average. 65+ = strong-buy region (top decile). Below 35 = strong-sell region. The number does not say "this stock will go up by X%" — it says "this stock currently looks more attractive than this percent of comparable names".

Plain-English bucket of the score: BUY, MIXED or SELL.

BUY = score ≥ 60 (factor stack agrees stock is attractive). MIXED = 40-60 (factors disagree, no clear edge). SELL = ≤ 40. We deliberately use 3 buckets, not 5 or 10 — finer granularity gives false precision when forward returns aren't calibrated yet.

The engine's best guess for how much the stock will return over the next N days, with a confidence band.

Computed by APEX v10 Forward-Return Engine as μ_h = drift + α × regime × pattern × tail, where α is the alpha contribution of the composite signal at horizon h, and the multipliers compose every other engine layer. We show 4 horizons (1d, 7d, 30d, 90d) with mean, 90% CI, and P(positive return). Prior-mode until 2026-05-16 — read it as "engine's theoretical projection from published factor literature", not "guaranteed outcome".

A label for what mood the overall market is in: risk-on (bullish), risk-off (bearish), or transition.

Adams-MacKay 2007 Bayesian Online Changepoint Detection — we maintain a posterior distribution over which regime the market is currently in, updated daily on SPY log-returns. Each factor has different efficacy in each regime (e.g. momentum works great in risk-on, gets crushed in risk-off). The regime modulates how we weight factors and how we amplify the alpha contribution.

A 90% confidence band around the score — accounts for how uncertain the engine is about this specific ticker.

Vovk-Gammerman-Shafer 2005 conformal prediction provides finite-sample, distribution-free coverage guarantees: if the model says [56, 82] is a 90% interval, then 90% of similar future predictions WILL contain the realised value. We use Mondrian conformal (different intervals per regime + verdict bin). Wider interval = more uncertainty about this name.

How often a pattern's factors actually move together in extreme conditions, not just on average.

Schmidt-Stadtmüller 2006: an empirical estimator for whether two factors co-move in their tails (extreme upper or lower decile). λ ≥ 0.25 means the factors historically co-move when the market is stressed — structural pattern, not coincidence. λ < 0.10 means they're independent in tails — the apparent setup is probably random alignment. We use this to dampen pattern confidence when tail-alignment is weak.

A way to update factor weights smoothly over time as their informational content drifts, rather than re-fitting from scratch.

Kalman 1960 / West-Harrison 1997 dynamic linear model: each factor's coefficient on forward returns is treated as a slowly-evolving random walk. We update the posterior on each new observation. This avoids the choice between "use the prior forever" (rigid) and "re-fit on a rolling window" (jumpy). Activates as forward returns accumulate post-2026-05-16.

A fair way to split the credit/blame for the score among the 12 factors — each factor gets exactly its marginal contribution.

Shapley 1953 game-theoretic value: across all permutations of factor inclusion, what's factor X's average marginal contribution to the final composite? This gives a cleaner attribution than simple "factor i has weight w_i × score" because it accounts for factor interactions. The Score Composition waterfall on every ticker page shows Shapley contributions.

A named setup the engine recognises when multiple factors align — like "Quality Compounder" or "Short Squeeze".

When a specific combination of factor scores fires, we label it with one of 18 academic-style patterns (Quality Compounder, Value Trap, Short Squeeze Setup, Earnings Quality Crack, etc.). Each pattern has a published source effect size. The pattern overrides part of the raw composite — that's the +δ shown on the banner. Tail-dependence determines whether the override is full-weight (structural co-movement) or half-weight (coincidence).

Companies that consistently make money on what they own — high gross profit per dollar of assets.

Novy-Marx 2013 showed that gross profitability (revenue − cost of goods sold, divided by total assets) predicts cross-sectional returns better than book-to-market value alone. High-quality firms compound; low-quality firms grind. We score 0-100 with 100 = highest GPA in our universe.

Stocks that look cheap relative to their fundamentals (low P/E, low P/B, etc).

Fama-French 1992 famously documented that high book-to-market stocks ("value") outperform low book-to-market stocks ("growth") in long horizons. We blend P/E, P/B, EV/EBITDA, and forward-PE divergence into one z-scored value factor. 100 = cheapest in our universe by composite, 0 = most expensive.

Stocks that have been going up recently tend to keep going up; the ones falling tend to keep falling — for medium horizons.

Jegadeesh-Titman 1993: 12-month-minus-1-month price return is one of the most replicated anomalies in finance. We use 12m−1m return + revenue acceleration. The "minus 1m" part avoids the well-known short-term reversal effect (Jegadeesh 1990). High momentum = recent winner. Low momentum = recent loser.

After a company reports earnings that beat expectations, the stock keeps drifting upward for weeks. Same downward after a miss.

Bernard-Thomas 1989 documented that stocks under-react to earnings surprises — the price moves on the day, but more drift continues for 60+ days. We measure standardized unexpected earnings (SUE) and recent earnings revisions to capture this. PEAD is one of the longest-standing anomalies in finance, persistent through decades of arbitrage attention.

Companies whose reported profits depend heavily on accounting accruals (rather than cash) tend to underperform.

Sloan 1996: investors fixate on reported earnings without distinguishing the cash component from the accrual component. Firms with high accruals (lots of "earnings" that haven't turned into cash yet) systematically underperform. We compute working-capital accruals scaled by total assets — high accruals = bearish signal. This factor is INVERTED in scoring (lower accruals = higher score).

Stocks heavily shorted by institutions tend to underperform — the shorts are usually right on average.

Asquith-Pathak-Ritter 2005: short interest above 5% of float predicts negative future returns. But there's a flip side — extreme crowded shorts can squeeze higher. We use short-interest-to-float ratio + days-to-cover. Score is INVERTED: low short interest = high score (bullish), high short interest = low score (bearish). The "SHORT_SQUEEZE_SETUP" pattern flags potential squeezes.

When company executives buy their own stock with own money, that's a strong bullish signal. Selling is weaker as a bearish signal (insiders sell for many non-information reasons).

Seyhun 1998 (the canonical text on insider trading): open-market purchases by officers and directors predict positive abnormal returns. We pull SEC Form 4 filings, weight by transaction size, scale by company market cap. Score uses net dollar flow over rolling 90-day window — insider purchases push score up, sales push it down.

Smart-money option-buying shows up before stock-price moves. Heavy call buying = bullish lead; heavy put buying = bearish.

Pan-Poteshman 2006: option order flow contains private information that hasn't reached the equity market yet. We track put/call ratio (volume) + IV skew. Low P/C ratio with high call demand = bullish. High P/C ratio = bearish.

When a key supplier or customer makes news, the linked company's stock moves a few days later — investors are slow to update.

Cohen-Frazzini 2008: economic links between firms (supplier-customer relationships) predict returns because the market fails to instantly reflect news from one firm into the price of its linked firm. We track sector-peer momentum and industry-peer earnings surprises as proxy for these links. High peer momentum that the focus stock hasn't yet absorbed = bullish spillover.

Reading the company's 10-K filing for negative-sounding words — companies whose MD&A is unusually negative tend to underperform.

Loughran-McDonald 2011 built the standard financial-sentiment dictionary (after showing that general English sentiment dictionaries fail catastrophically on financial text — "liability" is not a negative word in finance). Li 2008 showed that 10-K MD&A negativity has predictive power. We score the tone of the most recent 10-K MD&A section using LM-2018 dictionary; high negativity = bearish.

A formula for how much money to bet on a high-confidence setup so you grow capital optimally without going bust.

Kelly 1956 derived the bet size that maximises long-run log wealth. In practice, full Kelly is too aggressive (a single mis-estimate of edge tanks the portfolio). Practitioners use ½-Kelly. We compute Kelly from inferred edge (composite divergence from 50, scaled by 3% prior directional advantage) × confidence boost × interval-width penalty, and cap at ±5% of portfolio. Output: dollars per $10k allocated.

For pharma stocks: how likely the company's upcoming Phase 3 clinical trial readout will fail.

Hay et al. 2014 showed Phase 3 trial success rates anchor around 58% (so failure rate ~42% as a base). We Bayesian-update this base rate with live signals: enrollment velocity, endpoint amendments, mechanism evidence from PubMed, sponsor cash runway. Output: P(failure) per active Phase 3 + days to readout. Phase 3 failures move stocks 30-50% overnight, so timing matters as much as probability.

How many days until the next big binary event (Phase 3 readout, FDA decision, etc.) that could move the stock 30%+ in a day.

Pulled from ClinicalTrials.gov primary completion date for active Phase 3 trials. <90 days = imminent (size positions smaller). 90-365 = mid-term. >365 = distant (binary risk far enough that current pricing already reflects most of it).