Decile Tracker

What this measures

On the 10th of every month we lock the entire ranked universe into ten equal deciles by score (decile 1 holds the highest-ranked stocks, decile 10 the lowest) and then measure how each decile actually performs until the next vintage forms a month later. If the ranking works, decile 1 should beat decile 2, which should beat decile 3, and so on down the staircase. That monotonic staircase, or the lack of it, is the single most honest test of a ranking system.

We track seven rank dimensions independently: the composite StockRank (QVM), each factor alone (Q, V, M), and the three pairwise blends (QV, QM, VM). The blends answer “which factors carry the signal?” If QM tracks QVM closely, Value is adding little. That finding gets published either way.

The rules we cannot break

• Walk-forward only, no restatement. A vintage is formed from the scores that existed on that date and is never re-formed. Measured returns are frozen once written. No retroactive edits, no deleting bad months. The scoring weights this series validates were locked on 2026-07-06; if we ever re-weight, this series freezes and a new one starts under a new regime label. The old one does not quietly absorb the change.
• No survivorship laundering. Every stock formed into a vintage stays in it. A stock that delists or suspends mid-month is taken to its last available price; a stock with no price at all is held flat at 0%, never silently dropped. Per-vintage coverage is published so you can see exactly how much of each cohort was measurable.
• Total returns, one price basis. Returns are equal-weight within each decile, USD-normalised, and computed as total returns (dividends reinvested) from a single consistently-adjusted price series pulled at measurement time, so splits and dividends cannot fabricate or hide performance.

The investable gate

The headline series only includes stocks with a market cap of at least $200M and at least $1M of average daily traded value, both evaluated when the vintage forms (roughly 12,800 stocks, about 1,280 per decile). Thin, illiquid names carry stale prices that inflate decile returns, so we exclude them from the claim we are asking you to trust and show only the investable series on this page. The ungated series over all ranked stocks is still computed and stored alongside it. The gate applies at formation only: a stock that later shrinks below it stays in its vintage to its last price, so the gate itself cannot smuggle in survivorship bias.

Why monthly, on a fixed date

Factor effects, momentum especially, are studied at monthly horizons across the academic literature (Fama-French, Jegadeesh-Titman), so monthly is the standard unit of evidence. Daily or weekly deciles are dominated by market noise and imply churn no real investor would trade, while quarterly snapshots would spend three times the calendar to reach the same sample size, and statistical power here comes from the number of vintages, which starts at zero.

Our scores refresh nightly but change slowly (monthly financials, a data-quality gate), so a monthly cadence matches how the scores actually evolve. Each vintage forms on the 10th of the month, a fixed calendar anchor, so the schedule can never drift over the years. Calendar months run 28-31 days, so the staircase averages period returns geometrically normalised to a 30.44-day month; the vintage table always shows the raw figures.

How the combos are built

QV, QM and VM average the two component factor percentile ranks (each 0-100, universe-relative as of the vintage date) and re-decile that average. We deliberately do not average decile numbers (that throws away resolution) and do not reuse the composite’s internal weights (the point is an unweighted two-factor blend). A stock missing either component is excluded from that combo’s deciles for that vintage.

The AI overlay, tested

Alongside the quant deciles we track five AI-verdict cohorts. At each vintage, every stock is bucketed by its latest AI news review as of that date: AI Approve (deep review) (the weekly full-news review of the top-ranked shortlist), AI Approve (light screen) (the broader monthly tier-two screen), AI Flag, AI Remove, and Not AI-reviewed. The bucket is locked at formation (a verdict issued after the vintage forms cannot reach it), and each cohort’s forward return is then measured exactly like a decile: equal-weight, USD, total return, delisted stocks to their last price. The series is walk-forward from the first vintage formed after the AI review-tier system shipped; nothing is backfilled onto earlier vintages, and the tracker states the actual start date.

Two things must be said plainly. First, the cohorts are wildly uneven by design: deep-review approval is a tiny cohort (tens of stocks) while the unreviewed bucket is most of the universe, so per-cohort counts are shown everywhere and the deep-approve series carries a permanent tiny-sample label. Second, deep review is rank-gated: the AI reviews a top-of-rank shortlist, so the deep-approve cohort inherits top-rank membership and its returns are partly a rank effect, not pure AI signal. We publish each cohort’s average StockRank next to its returns so you can see exactly how much rank it started with, and the more informative comparison is Approve versus Flag/Remove within the reviewed set, stocks the AI saw and judged differently. Cohort membership follows the review schedule (a stock can move between deep and light coverage between vintages), so this measures the overlay as operated, not a frozen panel.

Country slices

The country filter answers “how do the global ranks perform among, say, Japanese stocks” by slicing the global deciles by each member’s listing country. We deliberately do not re-rank within a country: membership, decile boundaries, and scores are identical to the global series. Only countries with at least 100 investable members at formation get a published slice (about ten stocks per decile cell on average; thinner slices are noise, so they stay absent rather than published), and a country whose price coverage fails in a given month shows an explicit gap rather than an interpolated line. Because a country’s members are not evenly spread across the global deciles, some decile cells in smaller markets are thin; the per-decile counts under the staircase are there so you can weigh each bar accordingly.

Reading the numbers honestly

With fewer than 24 monthly vintages, everything on this page is labeled early evidence: enough to watch, not enough to prove. We publish the D1−D10 spread with its t-statistic and sample size rather than a bare average, and the monotonicity ρ (Spearman rank correlation between decile number and return) per vintage. Early staircases will be jagged; the cumulative view smooths as vintages accrue. Returns are gross of transaction costs and taxes, and no position sizing is applied; this is a test of the ranking, not a portfolio.