How the DeepVane APEX score is calculated — a plain-English walkthrough
Twelve academic factor families, one 0-100 composite, honest uncertainty intervals. What each component contributes and how to read the number on any ticker page.
If you've opened any ticker page on DeepVane you've seen a number like 67 / 100 with a small range underneath, e.g. [52 – 74], and maybe a pattern label like QUALITY COMPOUNDER. That single number is the APEX composite. This post walks through what's inside it, what the range means, and why you can read the pattern label as a shortcut to the underlying story.
The 12 factor families
APEX runs twelve separate 0-100 scorers over every ticker every day. Each scorer is a direct implementation of a published academic factor — no black-box features, no "alternative data" we can't describe.
- Quality (Novy-Marx 2013) — gross profitability and free cash flow yield. Answers "is this actually a good business?"
- Value (Fama-French 1992) — price-to-earnings, price-to-book, forward divergence. Answers "is this priced reasonably?"
- Momentum (Jegadeesh-Titman 1993) — 12-month minus 1-month price return, plus revenue acceleration.
- PEAD (Bernard-Thomas 1989) — earnings surprise magnitude and the subsequent price reaction.
- Insider flow (Seyhun 1998) — size-normalised net SEC Form 4 dollar flow over 90 days.
- NLP tone (Loughran-McDonald 2011 + Li 2008) — Management Discussion section sentiment and readability.
- Short interest (Asquith-Pathak-Ritter 2005) — crowded-short indicator with a directional inversion.
- Options flow (Pan-Poteshman 2006) — put/call imbalance and implied-volatility skew.
- Industry spillover (Cohen-Frazzini 2008) — peer industry performance with a time lag.
- Accruals (Sloan 1996) — earnings quality proxy from cash-flow vs reported income.
- Sector momentum (Moskowitz-Grinblatt 1999) — industry group vs broader market.
- Factor interaction (Asness-Moskowitz-Pedersen 2013) — the multiplicative Q×V×M term.
Each returns higher = more bullish. The twelve raw scores are combined with weights derived from each factor's published information coefficient (IC), shrunk Bayesian toward a uniform prior when out-of-sample data is thin.
Why combining works better than cherry-picking one factor
Any single factor has an IC of maybe 0.03 — meaning 3% of next-period return variance is explained by the factor score today. That sounds small because it is. The magic is that twelve mostly-independent factors combined with appropriate weights give a composite IC closer to 0.06-0.08 — because the noise components cancel while the signals stack. That's the entire case for multi-factor investing, going back to Grinold-Kahn's 1999 fundamental law of active management.
DeepVane measures this explicitly and publishes it on the coherence audit page as the "effective breadth" — how many genuinely-independent factors you actually have out of twelve raw ones. Live value is currently around 6-7, which is what you want.
What the [lower – upper] range means
The number in brackets under the score is a 90% prediction interval — not a forecast range, not a confidence band, specifically a conformal interval. It says: if the method is calibrated properly, 90% of similar tickers will land inside this range when we check their realised forward return 30 days from now.
Where most quant dashboards show a point estimate and stop, we publish the interval alongside it because lying about precision is the oldest mistake in applied statistics. Interval width itself shifts by regime — tighter in settled bull and bear phases, wider in transition — and that shift is driven by the same Bayesian regime posterior that drives the factor weights. One piece of math, multiple uses.
The pattern label as a shortcut
On top of the 12-factor composite, a pattern engine checks whether any of fifteen multi-factor academic setups fire on today's factor values. Examples: VALUE TRAP (cheap + low quality + bad accruals), SHORT SQUEEZE (crowded short + rising price + bullish options), QUALITY COMPOUNDER (high profitability + reasonable valuation + clean accounting).
When a pattern fires, the composite gets overridden by the pattern's published effect size — that's why the number on a ticker with a pattern label may not match the simple weighted average of its factors. The pattern is telling you the interaction is the story, not the list of twelve numbers. The full catalogue of patterns with citations lives on the Pattern Library.
Reading the number in practice
A sensible first pass:
- Score above 60 — bullish lean. If the pattern label is one of the bullish variants (QUALITY COMPOUNDER, PEAD DRIFT, SHORT SQUEEZE SETUP), the story is consistent. Check the interval width before sizing.
- Score between 45 and 60 — neutral to mildly bullish. Pattern is likely NO EDGE or CONFLICTING SIGNALS. The ticker is unremarkable today; wait for the picture to clarify.
- Score below 40 — bearish lean. If the pattern is one of the bearish variants (VALUE TRAP, QUALITY CRACK, PRICED FOR PERFECTION) the story is coherent — expensive or declining without fundamental support.
None of this is investment advice. It's a starting point for asking the next question — does the current catalyst calendar, position sizing, and regime environment support the thesis? The composite is a research tool, not a trading signal on its own.
Where we're honest about limitations
The stack deployed to production in April 2026. Forward-return tracking — actually measuring IC, Sharpe, conformal coverage against realised prices — begins 16 May 2026 after the first full 30-day window accumulates. Until then we show live pipeline diagnostics and literature effect sizes, not out-of-sample results we don't have yet. The Track Record page lists exactly what we will measure, with which benchmarks, starting when.
Small-cap coverage is thin today; the universe is ~234 curated names, weighted toward the S&P 500 plus selected semis, biotechs, and fintech exposure. International coverage (20-F foreign issuers) is at ~82% NLP coverage and improving. We'd rather publish a shorter list well than a long list poorly.