Matty Ice

The house thesis — a regime memo

Measured on the last regime; positioned for the next

This page takes a position, and states it as a position: the next five to ten years favor value over growth, in the United States and more so outside it. A regime change is underway — the long growth regime is giving way to one that pays for earnings, assets and cash returned, and charges full price for promises. Year to date through November 9, 2026, value is beating growth across the board — most dramatically in large caps, where Russell 1000 Value is up 26.2% against a 2.5% decline for Russell 1000 Growth, a spread of nearly 29 percentage points on a total-return basis. The same pattern holds down the cap table: S&P 500 Value +14.6% against S&P 500 Growth +8.3%, and Russell 2000 Value +20.4% against Russell 2000 Growth +10.9%. The scorecard at the top of this page tracks these races as they develop.

Total-return figures through the close of November 9, 2026: FinanceCharts daily series, cross-checked against PortfoliosLab and MyPlanIQ.

Read every backtest on this page with that in mind. A backtest can only sample the regime it is given, and every test below samples 2016–2026 — the growth regime itself. The one window that would test value on its home ground, 2000–2009, cannot be replayed honestly at all: as-filed fundamentals do not exist in machine-readable form before 2009, a limit the 4.0 study documents in full below. So the measured results stand exactly as measured — and they measure the last regime. They are not evidence about the next one, in either direction.

There is no edge in the consensus trade. Every portfolio already owns all the Nasdaq and US growth it can hold; pitching that trade again pitches what everyone already has. The edge, when a regime turns, is in the change itself — owning what the turn re-rates, bought before the crowd rushes in. That is what the screens and studies below are for, and it is why the international section on this page carries the weight it does: the deepest discounts, and the strongest version of this thesis, sit outside the United States.

Expressing the thesis

Positioning for the regime change

The thesis is not a reason to abandon the sectors this firm is bullish on for the long term — technology, AI, compute, the themes the DARTS database tracks. It is a reason to change the price discipline inside them. Two tools do that here.

The first is Matty’s screen itself — the GARP bridge. It hunts inside the buildout with a value investor’s tests: profitable growth, a price cap, a PEG ranking. Tonight it holds the buildout’s own giants at value prices — Micron at 4.2 times forward earnings, NVIDIA at 17.9 times. A regime change does not require leaving the buildout; it requires refusing to overpay for it.

The second tool is the deep-value discipline — the regime’s native instrument, ranking on price against book and dividend yield. The 4.0 test below transplanted it to US indexes, and it did not earn a live list here. But inside the Nasdaq-100 — the growth index itself — the plain value screen fought QQQ to a risk-adjusted near-tie (Sharpe 0.89 against 0.91) with about ten points less drawdown. Value is already surviving inside the growth index. As the regime turns, that is the beachhead.

The same two disciplines, other water

Other ponds

Neither discipline is American by nature. A screen that ranks price against assets, earnings and yield travels to any market where those numbers are filed and believed — and the deepest discounts in the developed world currently sit outside the United States.

Japan is the re-rating in progress. Governance reform since 2023 — the exchanges pressing listed companies on capital efficiency and price-to-book — has been unwinding a market where a large cohort of companies still trades below book value. The transmission mechanism is not hope; it is shareholder returns — buybacks and dividends paid by companies sitting on more cash than their market prices credit them with. A value ranking in Japan today fishes a pond that policy is actively restocking.

South Korea is the deepest discount among major markets. Its companies trade at roughly 0.45 times the global market's price-to-book level; about 52% of Korean listed companies sit below book value, against about 12% in the United States. Part of that is a family-control discount — founding families holding listed groups through small direct stakes — and the Value-Up reforms now attack it through the controlling families' own tax incentives, which is why the reform has teeth where earlier campaigns did not. The 15-year study of that market carries the strongest evidence on this page for the deep-value discipline: a price-to-book and yield ranking returned 12.2% a year; the same ranking behind the stability overlay returned 14.6% a year with a Sharpe ratio of 0.63 — in a market where the average stock returned 3.8% a year.

Beyond those two, the same pattern — profitable, asset-heavy, out of fashion — recurs in pockets of the UK and continental Europe, and in commodity-linked markets whose whole sectors trade under book in the quiet years of the cycle.

The home pond has its own left-for-dead water. Below the mega-cap complex that dominates the index, small and mid caps and entire unfashionable sectors trade at fractions of the multiples paid for the buildout — and tonight's lists already fish there. Between the two pass lists, the names span memory chips, enterprise hardware, education, footwear, apparel and security technology. The screen hunts price, not themes; it found an education company and a security-technology maker in the same net as the memory makers because the net is woven from numbers, not narratives.

Screens do not predict regimes. They harvest what each regime misprices — growth at a reasonable price in one decade's water, assets under book in another's — and they hold their opinions cheaply enough to be graded in public. The nightly forward grading is the referee: it shows, quarter by quarter, which pond is paying.

Scope, stated plainly: the screens on this page run today on the US-listed DARTS universe. The international ponds are the stated direction of travel — the same two disciplines pointed at the markets where the discounts are deepest — and the scorecard at the top of this page already tracks the international race, from behind.

Nightly · Run over the entire classified database

The Matty Ice Screen

Every night, this screen re-runs one set of value tests over the entire classified database — 1,456 companies after duplicate share classes are removed: return on assets at least 10%, price at or below its 200-day average, forward P/E at most 25 on the next fiscal year's consensus EPS, and 2-year EPS growth at least 10% a year, ranked by PEG. Tonight, 12 names pass — as of Nov 10, 2026 (committed snapshot). Refreshed nightly — the screen job runs at 2:45 AM ET.

A screen is a lens, not a recommendation. Passing means a name cleared the value tests tonight — it says nothing on its own about whether the company's own insiders are buying or selling; every name below links to its full public record, where that tape sits next to the fundamentals.

The tests, and what each one reads

• Return on assets (TTM) at least 10%Net income (TTM) / total assets — nightly fundamentals feed, SEC companyfacts
• Price at or below its 200-day moving averageYahoo daily closes, latest close vs the mean of the last 200 closes
• Forward P/E at most 25Latest close / consensus EPS for the fiscal year ending in 2027 — StockAnalysis forecast table, normalized EPS
• 2-yr EPS growth at least 10% a year(FY2027E EPS / FY2025 actual diluted EPS) ^ (1/2) - 1 — actual from SEC filings (StockAnalysis table as labeled fallback)
• One share class per companyDuplicate classes of the same issuer (same CIK) collapse to the larger market cap in the nightly cap feed

Tonight's list — 12 pass

How the database narrows — tonight's funnel

1. 1,468Classified universe→
2. 1,456One share class each→
3. 1,216ROA measured→
4. 221ROA at least 10%→
5. 108At or below 200-day avg→
6. 54P/E and growth on file→
7. 12Pass every test

Basis: ROA is net income (TTM) over total assets from SEC filings (the nightly fundamentals feed). Forward P/E is the latest close ÷ consensus EPS for the fiscal year ending in 2027 (StockAnalysis, normalized basis); the growth rate compares that estimate with the filed diluted EPS for the fiscal year ending in 2025. A name that cannot be measured on a test is a data gap in the nightly feed — never estimated, never passed. Tonight's gaps: 240 names carry no ROA in the fundamentals feed · 132 have no FY2027 consensus estimate · 35 have no forecast table · 1 has a non-positive FY2025 EPS base. The full nightly file: darts-screen-latest.json.

By email

The Matty Ice Brief

The screens in your inbox every weekday morning — what entered, what dropped out, and the full lists.

2016–2026 · Quarterly formations · Pre-registered variants

The Backtest

Study dated November 10, 2026.

Before this screen earned a nightly run, it was tested: ten and a half years of quarterly portfolios against the S&P 500, with every variant written down before any result was computed.

Two of the screen's tests run on analyst consensus estimates as of the day the screen is run — the forward P/E reads an estimated earnings figure, and the growth test ends at that same estimate. No point-in-time archive of those estimates exists in our data, so the historical screen cannot be replayed exactly; buying such an archive is a spending decision that has not been taken. What was tested instead is the closest declared reconstruction: trailing P/E in place of forward P/E, trailing two-year EPS growth in place of the estimate-based growth test, return on assets and the 200-day test unchanged. Fundamentals came from SEC filings, using only what had been filed by each formation date. Portfolios formed at each quarter-end from March 2016 through June 2026 — 42 formations — held one quarter and equal-weighted; a formation with no passers sat in cash for that quarter, and those quarters are counted, not smoothed away. The benchmark is the S&P 500 through SPY, dividends included. No trading costs or taxes are counted; measured turnover was low enough that a standard cost assumption moves nothing in the conclusions.

One honest consequence of the substitution: trailing numbers are stricter than estimates. Fast growers that pass a forward P/E test routinely fail a trailing one, so the reconstructed screen passed far fewer names than the live screen does — three names at the most recent formation, against the live screen's eighteen on Nov 3, 2026. The findings below are about the rule set's structure — which tests help and which hurt. The level of any single number belongs to the reconstruction, not to the live screen.

The original screen vs the S&P 500

The reconstructed screen edged the S&P 500 on raw return and lost on every risk-adjusted measure — a drawdown ten points deeper, and it was invested in only 26 of 42 quarters, usually in one to three names. It lost on risk-adjusted measures in both halves of the window as well: Sharpe 0.87 against the index's 1.05 in the first half, 0.52 against 0.90 in the second.

What else was tried

Three added tests were registered and run. Each one made the result worse, for the same reason: each shrank an already small portfolio. Quality persistence — requiring the ROA test to pass a year earlier as well, with positive free cash flow — returned a Sharpe of 0.51 and sat in cash for 20 quarters. Requiring the dip to sit inside a rising 200-day average returned a Sharpe of 0.44 and 27 cash quarters; a name below its average whose average is still rising is a rare, conflicted state. Requiring the P/E to also sit below the name's own five-year median returned a Sharpe of 0.51 and 23 cash quarters.

One further variant asked the question this research asks of everything: does insider buying agree? It required at least two insiders to have bought on the open market in the 90 days before formation, counted strictly from Form 4 filings as accepted by the SEC by that date. Across all 104 passer-quarters in the study, that condition was met once — and that single position lost 13.0% the following quarter. Eleven passer-quarters had exactly one buying insider; 92 had none. The screen's dip picks and insider accumulation are different animals, and inside this screen the confluence does not occur. As a filter it held cash in 41 of 42 quarters.

That honest-universe test has now reported, and the verdict was no Matty Ice 3.0 — 2.0 stands. The figures are in the “What might work going forward” section below.

One change · The pre-registered winner

Matty Ice 2.0

Matty Ice 2.0 is the original screen with one change: the price test — at or below its 200-day average — is removed. Every other test and the PEG ranking are unchanged: return on assets at least 10%, forward P/E at most 25, 2-year EPS growth at least 10% a year, one share class per company, ranked by PEG.

The dip test was the defect. It kept the portfolio out of the strongest compounders while they were working and admitted them mainly while they were falling. Deleting it was the only registered change that improved every measure at once, on identical data, against both the original screen and the index. Tonight, 2.0 passes 19 names — every name on the original list above, plus the names the dip test alone was holding out.

As of Nov 10, 2026 (committed snapshot) · Refreshed nightly — the same 2:45 AM ET run as the original screen.

Tonight's 2.0 list — 19 pass

The % vs 200-day column is printed for both lists because the number still describes each name — 2.0 simply no longer tests on it. Every name links to its record: a name with a full record page opens it, and the few names without one open the live stock lookup instead.

One more condition · The Test 5 winner · A candidate, graded from tonight

Matty Ice 2.0 + Trend Guard

Study dated November 10, 2026.

The trend-guard version is Matty Ice 2.0 with one condition added: a name that passes 2.0 is held only while its price sits above its own 200-day average. If it is a good company in a downtrend, it is not bought yet — it can enter the first night it passes with its price back above the line. Every other test and the PEG ranking are unchanged: return on assets at least 10%, forward P/E at most 25, 2-year EPS growth at least 10% a year, one share class per company, ranked by PEG.

The idea is the dip test with its sign flipped. The original screen bought only names priced below their averages — and two studies, on two different universes, found that leg was poison: on the honest pond it was the worst arm tested, at a Sharpe ratio of 0.52. Test 5 asked the opposite question on honest ponds — the S&P 500 and the Nasdaq-100 as they actually stood each quarter, 2016–2026 — and the flipped condition was the strongest of six candidate upgrades tried: one of three that beat the index under a winner rule written down before the test ran, on both ponds, and the one with the widest margins at the most gates. The full record of that test is printed in the test record below.

What it did on the honest ponds

S&P 500 pond — point-in-time membership, 42 quarterly formations, 2016–2026:

Nasdaq-100 pond — the same rules, the same window:

The halves check, which the winner rule required: on the S&P pond the trend guard cleared SPY's Sharpe ratio in the first half (1.06 against 1.01, formations dated 2016–2021) and in the second (0.95 against 0.93, 2022–2026), and beat SPY in 67% of quarters in each half. On the Nasdaq pond it cleared QQQ's Sharpe in both halves as well (1.27 against 1.20; 1.02 against 0.81). Its S&P drawdown of −24.7% was the shallowest of any screen variant in the study and within a point of the index's own; on the Nasdaq pond its drawdown was shallower than QQQ's. Two costs are printed with the result: the guard trades more — about half the list turns over each quarter, roughly half a percentage point a year at the standard cost assumption used across this research — and on the Nasdaq pond it holds only about eight names, so those figures carry small-portfolio noise.

How the winner was picked — the multiple-testing disclosure

Test 5 ran ten variants — the 2.0 anchor, three ablations and six candidate upgrades — against the same 2016–2026 window on two ponds: twenty variant-pond results, of which the six upgrades on two ponds were twelve genuine “can this beat the index?” trials. Three upgrades passed on both ponds. The list was closed and written down before any computation, the winner rule was fixed in advance, and every variant's result is published in the test record below — there was no silent search over hundreds of combinations, and the three survivors are different mechanisms (a trend condition, a profitability-tail exclusion, a concentration cap), not one lucky knob at three settings. Against that: the survivors are correlated — all three select from the same 2.0 passer set on the same decade — so their passes are not independent draws, and selection flatters the winners; some margins in this study (a rival candidate passed one gate by 0.001) sit well inside that bias. The 2016–2026 window, however fair its membership, is one path of history — a decade in which US large-cap quality and growth both worked. Every survivor of this test is a candidate, not a proven strategy. The only proof path is the one this research program has used throughout: freeze the rule, run it forward on the live nightly screen, and grade it against SPY from tonight onward under the write-once scorecard discipline.

Tonight's trend-guard list — 7 of the 19

19 names pass Matty Ice 2.0 tonight; 7 of them also pass the trend guard. The other 12 pass 2.0 but sit at or below their 200-day averages tonight, so this version does not hold them until the tape confirms — a name with too little price history to measure an average is not held either, because the guard cannot confirm it.

As of Nov 10, 2026 (committed snapshot) · Refreshed nightly — the same 2:45 AM ET run as the other screens.

The % vs 200-day column is the guard itself: every name on this list sits above zero on it tonight, by construction. Every name links to its record: a name with a full record page opens it, and the few names without one open the live stock lookup instead.

Built for the next regime · Frozen before it ran · Graded from tonight

The Forward Screen

Live from Nov 10, 2026 · Design study dated November 10, 2026.

Every other screen on this page was graded on the decade that just happened. This one was designed for the decade ahead. Its rules were written from mechanisms first — what should still be true in ten years — with the thresholds fixed in writing before any data was run, and the audit that followed was allowed only to veto a whole rule, never to re-tune one. Nothing was vetoed. It fishes three ponds at the weights of the regime thesis at the top of this page — 50% US value, 30% developed international, 20% emerging markets — and from tonight it is graded in public, forward. For the two international ponds, forward grading is not the second-best evidence. It is the only evidence they will ever have.

A screen is a lens, not a recommendation. Passing means a name cleared these six tests tonight, on its own filed numbers — it says nothing on its own about whether the company's insiders are buying or selling, and nothing here is a track record: the record starts tonight, at 100, in the scoreboard below.

The three ponds

The weights are a regime bet — the positioning argued at the top of this page — not the output of a backtest; no optimizer chose them, and a cap-weighted world portfolio (roughly 64% US, 25% developed, 11% emerging) is a different, larger bet on the regime that just ended. There is deliberately no regime switch: a timing overlay that moves to cash when the market falls was tested in this program and it de-risked after the falls and missed the recoveries, so the bet lives in the weights and the lists simply re-form every quarter. Companies incorporated offshore — Bermuda, the Cayman Islands and similar — are counted as a class of their own and never quietly assigned to a pond, and one share class per company survives, by the same CIK rule as the other screens.

The six tests, and what each one reads

• Profitability spine — return on assets at least 10%Trailing-twelve-month net income / latest total assets, from the company's own SEC filings (US GAAP and IFRS reporters read under one declared tag map). Cheap and profitable — never cheap alone
• Anti-trap tail cut — the least profitable quarter of passers is droppedThe bottom return-on-assets quartile within the pond is removed; the names nearest the profitability line are where value traps concentrate
• Relative valuation — trailing P/E at most 1.25 times the pond medianLatest close / trailing diluted earnings per share, against the median of its own pond — no absolute cap, so the test means the same thing in any rate regime or sector mix
• Not deteriorating — 2-year EPS growth at worst −10% a yearTrailing EPS now vs two years ago. A company that climbed out of a loss, or whose two-year figure cannot be computed, passes on positive earnings today — the recovery rule, printed rather than hidden
• Shareholder yield at least 2%Declared dividends per share / price, plus buyback yield (the split-adjusted share count now vs a year ago). A component that cannot be measured from the filings counts as zero — and is counted, in the coverage line under each list
• Trend gate — price above its 200-day averageLatest close vs the mean of the last 200 daily closes. The same flipped dip test as the trend-guard version above, doing risk work, not return work
• ConstructionEqual weight within each pond; ponds at 50/30/20; a pond with no passers tonight holds its weight in cash — it is not redistributed

What the design did on the one pond we could test

Only the US pond could be audited historically — the international ponds cannot be, for reasons printed in full below. On the point-in-time S&P 500, 2016–2026, the full six-test stack returned 17.08% a year against SPY's 15.13%, tied the index on risk-adjusted return (Sharpe 0.97 against 0.99), beat it on downside-adjusted return in both halves of the decade (Sortino 1.86 against 1.67), and carried the shallowest maximum drawdown on the table, −21.0% against the index's −23.9%. It is not the return-maximizer in our studies — the trend-guard version above holds that title. It is the drawdown-minimizer with comparable return, assembled from parts chosen because their mechanisms should survive a different decade.

Tonight's lists

As of Nov 10, 2026 (committed snapshot) · Refreshed nightly — the forward screen runs at 3:20 AM ET, after the other screens. Membership as of Nov 10, 2026.

US value — 37 pass

No flags tonight: the list is at least ten names wide and fully invested.

Pond of 513 companies; 513 could be measured tonight; median trailing P/E 23.95. Coverage: return on assets measurable for 503, earnings in US dollars for 494, a dividend figure for 358, a buyback figure for 464.

Developed international — 3 pass

Concentration flag: fewer than ten names pass tonight, so this pond's list is a concentrated portfolio, not a diversified one. The flag is part of the output — it stays on until the list widens.

Pond of 753 companies; 743 could be measured tonight; median trailing P/E 20.06. Coverage: return on assets measurable for 603, earnings in US dollars for 460, a dividend figure for 43, a buyback figure for 505.

Emerging markets — 0 pass

No names pass tonight — this pond's weight sits in cash until the next formation. The empty list is the honest output of the frozen rules, printed rather than filled.

Pond of 408 companies; 407 could be measured tonight; median trailing P/E 18.17. Coverage: return on assets measurable for 306, earnings in US dollars for 251, a dividend figure for 12, a buyback figure for 204. A dash in a yield column means that component could not be measured from the filings and counted as zero, exactly as the rule states; a dash in the growth column means the two-year comparison could not be computed and the name passed on positive earnings today.

The forward scoreboard

Grading began Nov 10, 2026 — no grades yet; first reads at 30 days. The portfolio starts at 100 tonight, beside the 50/30/20 blend of SPY, EFA and EEM and SPY itself, and is chained nightly from each night's frozen lists — write-once snapshots, never restated, never backfilled. What it shows at 30, 90 and 180 days, and over rolling quarters, is the whole point of this section.

Basis: net income, assets, earnings per share and share counts from SEC companyfacts — US GAAP and IFRS reporters under one tag map declared in writing before the job first ran; per-share figures are read only in US dollars per share, so a filer reporting earnings only in another currency is a printed gap, never a converted estimate. Prices: Yahoo daily closes; grading closes are dividend-adjusted. A name that cannot be measured on a test is a gap in the nightly feed — never estimated, never passed. The full nightly file: darts-forward-screen-latest.json.

Two species of cheap

GARP vs deep value

Matty's screen is a growth-at-a-reasonable-price screen — GARP. Its tests are a value investor's discipline pointed at growing companies: a quality floor (return on assets at least 10%, so the growth is profitable growth), a price cap (forward P/E at most 25, so the growth is not bought at any price), a growth floor (earnings compounding at least 10% a year), and a PEG ranking that orders what passes by how little each unit of growth costs. The 2.0 revision keeps all of that and drops only the dip test.

That combination is the bridge between the value temperament and the AI buildout. Tonight's list makes the point in the tape's own numbers: Micron passes at 4.2 times forward earnings and NVIDIA at 17.9 times — the buildout's central names, bought at multiples a value investor can read without blushing. The screen does not ask a value investor to believe in the buildout. It asks what the buildout's suppliers earn, and what those earnings cost.

Deep value is the other species, and it does not care about growth at all. The version studied here — in a 15-year study of the South Korean market — ranks a market on two measures only, price against book value and dividend yield, and holds a name while it stays in the cheapest 30% on that combined rank. A stability overlay runs first: the most heavily traded quartile and the least profitable quartile are excluded before the ranking starts. That overlay is the value-trap filter — the difference between cheap because mispriced and cheap because the business is going away.

The distinction is not academic. In that study, a shipbuilder once topped the pure value ranking at 0.53 times book value with a 2.7% dividend yield — the cheapest name in the market on the numbers. Within a year it was in its creditors' hands. A ranking that cannot tell that name from a mispriced one buys both, and the stability overlay exists because of exactly that row in the table.

A deep-value book looks nothing like the list at the top of this page. The study's current simulated portfolio sits near 0.34 times book value with a dividend yield near 5%, and everything in it trades below book. The two screens will almost never hold the same stocks — one buys profitable growth when its price is reasonable, the other buys assets when their price is distressed — and that is the point of running both. They are answers to two different questions, and each keeps the other honest about what “cheap” means.

Regimes, not predictions

What might work going forward

Every backtest samples one regime and reports it as if it were the weather. The 2016–2026 window behind this page's study was a growth regime: capital was cheap, then AI capital spending rewarded scale and momentum, and value disciplines spent a decade explaining themselves. The decade before it ran the other way — in the United States, 2000–2009 belonged to value. Neither window is a law. Each is a sample.

The Korean evidence adds the more useful lesson: a regime change is not the only way cheap stocks get paid. Through the very decade growth was winning elsewhere, deep value in South Korea compounded at double-digit rates — because that market kept de-rating, pushing more and more profitable companies below book, where the screen could buy earnings and assets at widening discounts. Cheapness joined to profitability paid without any rotation arriving to rescue it. What actually failed in expensive markets was the shallow version — “cheaper large caps” that were merely less expensive than the index and cheap in no absolute sense at all.

Across the themes this page's database tracks, the two disciplines divide the work unevenly, and tonight's lists show how. In Data and AI — the buildout itself — names cycle hard: capacity gluts and digestion quarters knock even the strongest suppliers down the multiple table, and that is where the GARP screen finds them; five of tonight's twelve original passes are buildout memory and hardware names. Technology carries both screens at once — it holds the profitable growers the GARP tests want and, in its out-of-favor corners, the asset-cheap names a deep-value ranking would surface. Robotics and Space runs mostly empty under both: the theme is young, few of its companies are profitable yet, and a screen whose first test is return on assets has little to measure. Tonight neither list holds a single Robotics & Space name — stated plainly because it is the honest output. Sports and consumer names produce genuine value passers in both disciplines: tonight the original list holds Deckers in footwear and Tapestry in apparel, at 11.5 and 16.5 times forward earnings respectively.

Two tests stood in marked slots on this page, their results to be added either way. Both have now reported:

Both verdicts were no. A third test then asked the narrower question — not a new screen, but whether one added condition could make 2.0 beat the index on the honest ponds — and it reported yes: the trend-guard version above now runs as a third nightly list, a candidate graded in public like the other two, with the full study printed in the test record below. The regime argument — growth's decade or value's, this pond or that one — is not settled by essays. The tape arbitrates it, one graded quarter at a time.

For review · Every honest-pond study, every variant, unedited

The full test record

Everything the honest-pond research measured, in one place: the three studies that followed the first backtest, with their full tables and verdicts as reported. Measured results are printed unedited — what this page believes, and what it is doing about it, are stated in the sections above.

Test 5 — could better filters beat the index on a fair pond?

Study dated November 10, 2026.

The question, as put to the research: plain 2.0 does not beat SPY on the honest S&P 500 pond — would it have, with better filters? Ten variants were registered in advance — the 2.0 anchor, three ablations (one or two of the screen's legs at a time) and six candidate upgrades — and every one was run and is reported here. Universes: point-in-time S&P 500 membership (the primary pond) and point-in-time Nasdaq-100 membership (the confirmatory pond), 42 quarterly formations, 2016–2026, on the same declared reconstruction as the earlier studies — trailing P/E and trailing two-year EPS growth standing in for the live screen's forward-estimate legs, as-filed fundamentals only. The Russell 1000 was not run: no point-in-time membership history for it exists in our data, and running today's list backward would rebuild exactly the survivorship this test exists to remove.

The winner rule, written down before the test ran: a variant beats the benchmark on a pond only if all of the following hold — its full-window Sharpe and Sortino exceed the benchmark's; its compound annual return exceeds the benchmark's; its maximum drawdown is no more than 5 percentage points deeper than the benchmark's; and in each half (formations dated 2016–2021, then 2022–2026) its Sharpe is not below the benchmark's half Sharpe by more than 0.05. A confirmed winner must pass on both ponds; a variant passing one pond only is flagged pond-specific.

S&P 500 pond — full window (benchmark: SPY):

Nasdaq-100 pond — full window (benchmark: QQQ):

The anchor gate held: V0 reproduced the 3.0 study's honest-pond result exactly (16.27%, Sharpe 0.9473) before any other number was read, and the no-filter basket reproduced its control exactly (12.55%, 0.79) — the comparisons stand on the earlier studies' own ground. The ablations carry their own lesson: the P/E (value) leg does the most work of the screen's three legs, the growth leg the least, and no single leg or pair beats the index — the legs are worth more together than apart, just not enough, unmodified, to clear the index on the primary pond.

Disclosures, printed with the result. V5's thinnest gate: its S&P first-half Sharpe gate passed by 0.001 (0.959 against a floor of 0.958) — the thinnest margin in the study — and its full-window S&P Sharpe clears SPY by +0.03; it is a pass, and it is also the pass most likely to be luck-adjacent. V7's inherited pass: the Nasdaq pond averages only 12.7 passers, so a top-25 cap never binds there — V7 is numerically identical to plain 2.0 on that pond, and its Nasdaq pass is the anchor's pass, not independent confirmation of the PEG idea; its full-window S&P Sharpe margin is +0.011. V4's failed gate: the best raw numbers on both ponds (S&P 19.25% at Sharpe 1.07; Nasdaq 26.26% at 1.20) failed the winner rule on exactly one check — its S&P drawdown, −30.5%, is 6.6 percentage points deeper than SPY's, past the registered 5-point limit — so under the rule as written in advance it is flagged pond-split, not a winner; the rule was not bent for it. V8's clear negative: a cash-flow-yield floor was slightly worse than 2.0 on the S&P pond and the worst variant on the Nasdaq pond (12.48%, Sharpe 0.66) — inside a growth index, that floor systematically selects the ex-growth names the index is leaving behind. And the multiple-testing disclosure in the trend-guard section above is part of this result, not boilerplate: three correlated survivors out of six candidates on one decade are candidates, and the forward record arbitrates.

The 3.0 study — 2.0 on the honest universe

Study dated November 10, 2026.

The first honest-pond rerun: the identical reconstructed screen on point-in-time S&P 500 membership — names entering and leaving as the index changed — with two pre-registered candidates for a “Matty Ice 3.0”: a regime switch (half in cash for any quarter where SPY closed below its own 200-day average at formation) and risk weighting (positions weighted in inverse proportion to their trailing volatility). A candidate earned the name only by beating both plain 2.0 and SPY on Sharpe over the full window and in both halves. Full window — 42 quarters, 2016–2026:

Verdicts, as reported. The filters are real: 2.0 beat the no-filter basket of the same universe clearly (Sharpe 0.95 against 0.79; 16.27% against 12.55% a year) — the opposite of the hindsight-pond result, where the basket won. But 2.0 does not beat the index: ahead of SPY on raw return and a hair ahead on Sortino (1.68 against 1.67), it lost on Sharpe (0.95 against 0.99) with a deeper drawdown (−27.0% against −23.9%) — a good concentrated quality/value screen that roughly matches the index risk-adjusted, with more return and more drawdown; not an index-beater. The dip leg is poison, confirmed twice: the original screen with its price-at-or-below-its-average leg was again the worst arm on the table (Sharpe 0.52; 0.68 on the hindsight pond). Both candidates failed: the regime switch failed at the first gate (Sharpe 0.82 against 2.0's 0.95) — a quarter-end trigger de-risks after the fall and stays half-out through the first recovery quarter, cutting drawdown by only about a point while costing about 4 points of annual return across its 7 regime quarters; risk weighting came close (0.93) but trailed plain 2.0 and SPY, beating 2.0 in the second half alone when the rule required all three windows. Verdict: no Matty Ice 3.0 — 2.0 stands.

The 4.0 study — the Korea deep-value screens on US indexes

Study dated November 10, 2026.

The two deep-value screens from a 15-year study of the South Korean market, transplanted exactly as registered to point-in-time US index membership with as-filed fundamentals. V-VALUE ranks every member on a composite of price-to-book (cheapest first) and trailing dividend yield (highest first) and holds the cheapest quintile, selling only when a holding falls below the top 30% of the ranking. V-STAB runs the same ranking after first excluding the most heavily traded quartile and the least profitable quartile — the study's value-trap filter. Financials and REITs are excluded from the ranking, as in the original study, and kept in the control and the benchmarks.

S&P 500 — full span, September 2016 – October 2026 (122 months; benchmark: SPY total return):

Nasdaq-100 — full span, September 2018 – October 2026 (98 months; benchmark: QQQ total return):

What the tables say. On the S&P 500 the plain value screen bought genuine value — average holdings at 1.86 times book against the universe's roughly 9, a 3.26% dividend yield against 1.43% — and still lost: to the index (Sharpe 0.72 against 0.99) and to its own universe's equal-weight control (0.79), meaning in the S&P the screen subtracted a little versus simply owning everything equally. Inside the Nasdaq-100 the same screen fought QQQ to a risk-adjusted near-tie (0.89 against 0.91 — a 0.004 gap in the 2020–2026 window) with about ten points less drawdown (−22.7% against −32.6%), and beat its own control clearly (0.80) — the best showing of the study, and still not a beat: about 1.4 percentage points a year less return, ahead in only 3 of 7 full calendar years.

The stability overlay inverted. In the Korean study the overlay was the main claim; in US data V-STAB underperformed plain value on risk-adjusted return in all five reported universe-and-window cells — it won 0 of 5. The diagnostic shows why. Bucketing every ranked name by the two tails and pooling its forward 12-month return: the bucket the Korean study says to avoid at all costs — most heavily traded and least profitable — returned +25.4% on the S&P universe and +63.2% on the Nasdaq universe, against Korea's −12%. Heavy trading in 2016–2026 US large caps marked the liquid winners of the AI buildout, not speculation, so the overlay excluded the returns and kept a smaller book at higher average price-to-book than plain value. One half of the overlay did point the right way in US data: the least-profitable-only bucket was the weakest cell in both universes (+11.0% S&P, +7.6% Nasdaq, pooled) — the profitability tail was genuinely toxic, a thread Test 5 later picked up as its V5 variant. The overlay as a whole is a Korea-specific regularity, not a portable one.

The value decade is NOT EVALUABLE — recorded, not hidden. The test required as-filed fundamentals — what a company had actually reported by each formation date. EDGAR's machine-readable data does not exist before the 2009–2011 filing mandate: price-to-book coverage is 0.000 for every year 2000–2008 in both universes, and under the study's pre-registered 70%-coverage bar the S&P test can only start in 2016 and the Nasdaq test in 2018. The 2000–2009 window — the one regime where the paper's logic should shine — cannot be replayed honestly from the data that exists; anyone showing a US value backtest “proving” that decade from this data is using numbers that were revised, restated or backfilled later. The regime question stays open, in both directions.

Verdict: no Matty Ice 4.0. Under the rules registered before the test ran, the stability variant had to beat plain value in a majority of the five reported cells — it won 0 of 5 — and plain value had to beat its benchmark's Sharpe in the full span (with the value-decade clause evaluable) — it beat it in 0 of 5 cells, and the value-decade clause is NOT EVALUABLE for either universe. Neither screen earned the name, and no US deep-value list was added. Matty's GARP screen remains the bridge.

Appendix · The first study, kept in full

2.0 in the first backtest — the hindsight pond

This is where the page’s first study now sits: at the bottom, kept in full, caveat and all. Its universe was chosen with the answer in hand — a list of AI names tagged in November 2026, held backward through the decade those names won — so its figures describe that pond, not the screen. The honest-pond studies above are the record that counts; this one is preserved because the page does not delete its own history.

2.0 in the backtest

2.0 was invested in all 42 quarters and was the only variant that met the winner rule written down in advance — ahead of both the original screen and the S&P 500 on risk-adjusted measures over the full window and in each half separately (Sharpe 1.38 in the first half, 1.03 in the second). That is a clean, unbiased comparison: same universe, same dates, one test removed.

The mechanics under the numbers

Why the backtest came out the way it did

Start with the dip test, because it explains most of the table. A test that admits a stock only when its price sits at or below its own 200-day average is a test that says no almost all the time — and says it loudest about the strongest companies, whose prices spend most of their lives above their averages. Across 42 quarterly formations the reconstructed screen held cash in 16 quarters and averaged just 2.5 names per formation. A portfolio that owns nothing most of the time does not have a stock-picking problem; it has an absence problem. Its return, good or bad, is the story of a handful of names in the quarters it happened to hold them.

That is also why every added test made the result worse. Quality persistence, the rising-average condition, the own-history value check — each was a sensible idea applied to a basket that was already down to two or three names. Every new gate can only remove names or add cash quarters, and the cash quarters rose to 20, 27 and 23. Filters multiply; baskets shrink. On a small basket, the marginal test is almost never measuring what it names — it is measuring the cost of owning even less.

Insider confluence failed for a different reason: the two patterns live in different weather. This screen buys prices that have already fallen below their averages — extended, unloved, often still falling. Insider clusters form when executives judge a price attractive against what they know about orders, pipelines and margins, which is usually a different moment in a stock's life, and frequently a rising one. In ten years of quarterly formations the two met once. A rule that fires once a decade is not a filter; it is an anecdote with a threshold.

None of this required believing the reconstruction was the screen itself. The literal screen reads analyst estimates as they stood on the day — a data set nobody in this building owns in point-in-time form — so the study declared its substitutions up front, ran every variant through the same protocol on the same dates, and wrote the variant list down before computing anything. The conclusions that survive are structural: which legs help, which legs hurt, which combinations cannot coexist. Those do not depend on whether the trailing stand-ins match the estimate-based original quarter for quarter.

And over all of it sits the control. Hand the same universe to a portfolio with no tests at all and it beat every filtered version on risk-adjusted terms. The pond made the money: a set of AI names, tagged after the fact, carried through the decade that rewarded exactly those names. That is why nothing on this page treats the backtest as a verdict on what the screen earns from here. The run at the top of this page is graded forward, nightly, on the classified universe as it stands on the day — no hindsight tags, no chosen pond — and that record, not this study, is the evidence that counts.