The Phase-3 rotation basket, tested against 2018-19 and 2022-23: every expression pays from the hindsight date and none pays from the date our own rule fires

Illustrative, not a recommendation. Every ticker below is named as an instance of a phase-defined universe rule, to make the rule checkable. None is a buy, sell, or short recommendation, and nothing here authorizes a paper trade or a live order. The founder's standing frame applies: goal is alpha, horizon is position-length, capital cycle is the first thesis and not the frame.

The question

Verbatim: *"What does a Phase-3 (post-peak memory cycle) rotation basket actually look like — in the 2018-19 and 2022-23 DRAM downturns, which trade expressions (memory-consumer margin relief, semicap/equipment de-rating, outright short, exit-to-cash) actually paid, over what holding period, and what universe rules would have selected them in advance rather than in hindsight?"

Context: [[2026-09-25-hbm-dram-phase-2-3-transition-q3-2026]] found the rotation unexecutable because the Phase-3 side of the trade has never been specified. This brief is about what you buy once it turns, not about what triggers the turn (that is [[2026-09-26-financing-led-capital-cycle-phase3-precedent]]).

What we already know (from the vault)

• No Phase-3 basket exists anywhere in the vault. [[2026-09-25-hbm-dram-phase-2-3-transition-q3-2026]] states it directly after a full-vault search, and names the four candidate shapes this brief tests. Its only directional sketch is flagged as non-recommendation: rotate toward memory buyers whose input costs fall and away from the producer and wafer-fab-equipment complex.
• Every named basket in 01-projects/investing/ is the Phase-2 side. memory-cycle-v1.1 holds MU, SMH, SNDK, INTC; the backtest universe adds WDC, STX, MRVL. These are the positions a rotation would be selling.
• The vault already has a mechanical down-cycle rule, in anchors/memory-cycle-v1.1/phase-history-notes.md: down-cycle entry fires when the TrendForce DRAM-spot monthly aggregate shows price-down or price-neutral signals dominating for 3+ consecutive months AND at least one tier-1 vendor cuts capital-expenditure guidance. Both legs must hold. That conjunction is the whole story of this brief.
• The labeled phase dates in phase-history.csv are hindsight dates, not firing dates. The file labels down-cycle entry at 2018-09 and 2022-03. The rule's second leg (a tier-1 capex cut) did not exist on either date.
• Prior vault work measured the drawdowns but never the rotation. [[2026-09-09-hbm-plateau-hypothesis-backtest-variant]] carries base rates (roughly 56% equity drawdown in 2018, roughly 50% in 2022); [[2026-09-07-memory-supplier-inventory-days-dio-anchor]] builds the Micron days-inventory series through both downturns from filings; [[2026-09-09-2018-capacity-announce-asp-rollover-lag]] dates the 2018 rollover. None of the three asks what you should have owned instead.

What the web says

• The 2018 capex-cut leg is dateable to 2018-12-18. On its fiscal-Q1-2019 earnings release Micron cut fiscal-2019 capital expenditure to roughly $9.0-9.5B from roughly $10.5B, extended a 5% cut to DRAM wafer starts, and lowered its 2019 DRAM bit-demand outlook to about 16% from about 20% (SEC 8-K exhibit 99.1; figures via EE Times). So the vault's down-cycle rule completes both legs on 2018-12-18, roughly three months after the label date.
• The 2022 capex-cut leg is dateable to 2022-09-29. Micron cut fiscal-2023 capex roughly 40% and wafer-fab equipment roughly 50%, calling the oversupply unprecedented. The vault already carries this citation inside phase-history.csv at the 2022-09 row. That is roughly six months after the 2022-03 label date.
• The margin-relief mechanism is real at the profit-and-loss level but gets offset. Dell's fiscal-2021 Form 10-K management discussion states component costs were deflationary in aggregate for both the Infrastructure Solutions Group and the Client Solutions Group, and in the same breath that the benefit was offset by higher supply-chain expediting costs. Memory is reported industry-wide as roughly 10-70% of bill-of-materials depending on the system, so the input-cost swing is large, but it is competed into price rather than retained.
• Web coverage of both downturns is dominated by 2025-26 shortage commentary, which is the opposite regime. I did not find a reputable published study that measures downturn rotation returns with a stated ex-ante universe rule. That absence is itself a finding: the rotation basket is folk knowledge, not a documented strategy.

The test I actually ran

Because no published study answers this, I built the test. Universe rules first, then prices, so nothing is curated backwards.

Ex-ante universe rules (each applicable on the signal date with only information available then):

R2 is the only rule that is fully mechanical with no judgment (SIC code plus a market-cap screen). R1 is close to mechanical. R3 is not. "Purchased memory is a named cost input" cannot be reduced to a filter I can defend as ex-ante; I hand-assembled it and am saying so plainly rather than back-fitting a rule. Treat every R3 number below as weaker evidence than R1, R2 or R4.

Signal dates. Two per cycle, deliberately. The label date is what phase-history.csv records. The ex-ante date is the first month-end after both legs of the vault's own down-cycle rule were satisfiable in real time.

Method. Equal-weight, buy-and-hold from the signal date, split- and dividend-adjusted closes from Yahoo Finance end-of-day. A name is excluded from a basket if it was not listed on the signal date. Horizons 3, 6, 12, 18, 24 months. Benchmarks SPY and SMH.

Result: equal-weight total return (%) from signal date

What each expression actually did

Outright short the producers. Paid, and only, from the hindsight date, and only over 3-6 months: -33.3% over three months in 2018 and -35.0% over six months in 2022 (a short captures roughly the negative of those, before borrow and slippage). From the ex-ante date the same short loses instantly: producers returned +30.9% over the next three months in 2018. The expression is real and the entry is not reachable by our rule.

Semicap de-rating as a short. Failed in both cycles at every horizon. From the 2018 label date semicap fell 17.8% over three months, which is only 3.6 percentage points better than shorting SPY. From both ex-ante dates semicap beat SPY at every horizon, by 58.7 points at twelve months in 2018 and 27.4 points in 2022. The reason is visible in the pre-signal window: the semicap basket had already fallen 26.0% (2018) and 28.2% (2022) in the twelve months into the ex-ante date. Equipment de-rates ahead of the phase call, then leads the recovery. Its 24-month max drawdowns (-38.3% in 2018-19, -47.3% in 2022-23) are memory-sized, but they are spent before the signal.

Memory-consumer margin relief as a long. The weakest of the four on evidence, and the one whose headline number is most misleading. From the 2018 label date it returned -12.8% at twelve months, underperforming SPY by 16.4 points and SMH by 25.6 points; every member except Apple was negative at twelve months. The 2022 +344.5% at 24 months is one stock: Super Micro returned +2553% over that window on an artificial-intelligence server story that has nothing to do with memory input costs. Median member return over the same window was +23.4%. Excluding Super Micro, the basket returned +46.5% at twelve months from the ex-ante date against SMH's +58.4% — it underperformed the complex it was supposedly rotating out of. The margin-relief trade did not pay in either cycle once you remove the one name that was trading on something else.

Exit to cash. The only expression that paid from a hindsight date at short horizons without a stock-selection call: +14.8 points over SPY at three months in 2018 and +16.2 points in 2022, +20.8 points at six months in 2022. It then loses decisively: -52.8 points versus SPY at 24 months from the 2018 ex-ante date. Cash is a 3-6 month expression with a hard expiry, not a rotation.

Convergences and contradictions

• Convergence. The vault's own statement in [[2026-09-25-hbm-dram-phase-2-3-transition-q3-2026]] that the Phase-3 side "was never built" is confirmed and sharpened. It was never built because the three non-cash expressions do not survive an honest ex-ante entry date, so there was nothing to write down.
• Contradiction with the vault's sketch. [[2026-09-25-hbm-dram-phase-2-3-transition-q3-2026]] points at memory buyers and away from producers and wafer-fab equipment. Measured, that is the wrong direction in both cycles: consumers underperformed the memory complex, and semicap outperformed from every ex-ante date. The sketch is intuitive and does not survive the test. It should be retired as a default rather than carried forward.
• Contradiction with [[2026-09-26-financing-led-capital-cycle-phase3-precedent]], or rather a joining with it. That brief concluded financing distress marks memory bottoms, not tops, and that the only implied buy-side action is at the Phase-4 trough. This test independently lands in the same place from price data: in both cycles the equity trough arrived within days of our down-cycle rule's confirmation. Micron's closing low was 2018-12-24 against a rule confirmation of 2018-12-18, and 2022-09-26 against 2022-09-29. A tier-1 capex cut is a capitulation print, and capitulation is a bottom marker. n=2, so this is a mechanism claim supported by two observations, not a calibrated probability.

Synthesis for RDCO

The honest answer to the question as posed is that a Phase-3 rotation basket, triggered by a Phase-3 signal, has no support in either downturn. All three non-cash expressions are hindsight artifacts. They pay when measured from the date a later labeler writes into a phase file, and they fail or invert when measured from the first date the vault's own mechanical rule could have fired. The gap is three months in 2018-19 and six months in 2022-23, and that gap contains most of the drawdown. Micron peaked on 2018-05-29 and on 2022-01-14, four and two months respectively before even the label date. The equity market prices the phase change before the fundamental series confirms it, which is exactly the property that makes a fundamentally-triggered rotation unbuildable.

This reframes what the vault should build. The follow-up from [[2026-09-25-hbm-dram-phase-2-3-transition-q3-2026]] reads "specify the Phase-3 rotation basket." On this evidence that ticket is aimed at a trade that does not exist. The specifiable thing is narrower and less satisfying: a pre-committed trim schedule on the Phase-2 book, keyed to gross-margin and inventory readings rather than to a phase trigger, executed while the thesis still looks right. Both cycles offer the same tell — Micron's gross margin peaked in the fiscal quarter ended 2018-08-30 and again in the quarter ended 2022-03-03, and in both cases days-inventory turned within two quarters of that peak, per [[2026-09-07-memory-supplier-inventory-days-dio-anchor]]. A margin-peak trim rule fires roughly a quarter before the label date and two quarters before the ex-ante rule. That is the only ex-ante-reachable edge this test surfaces, and it is a risk-reduction rule, not an alpha rule. It should be evaluated honestly as such rather than dressed up as a rotation.

There is a second, cleaner conclusion for the four-expression menu. Three of the four should be struck. Shorting semicap failed in both cycles and would have been the single most expensive expression from either ex-ante date. Memory-consumer margin relief failed in 2018-19 outright and in 2022-23 once the one unrelated name is removed, and it carries the additional problem that its universe rule is the only one I could not make mechanical. Outright shorting the producers is the one expression with real magnitude, and it is unreachable by any trigger the vault currently has. That leaves exit-to-cash, which is not a basket and needs no universe rule, and whose entire measured edge lives in a 3-6 month window off a date we cannot identify in real time. A menu that reduces to "hold less" is a legitimate answer, and it is a materially different answer from "build a basket."

The structural lesson generalizes past memory. The vault labels phases from outcomes and then reasons as if the label were available at its own timestamp. Backtests built on phase-history.csv inherit that. This is the same defect [[2026-09-09-hbm-plateau-hypothesis-backtest-variant]] found when it judged the plateau variant "an unconditional hold with decorative exits," and it is why that brief's honest rerun showed buy-and-hold beating the rules in four of six windows. Any future use of the phase file for position logic should carry a second column: the date each transition was confirmable, not just the month it is assigned to. That is a small, concrete, and checkable change to an existing artifact.

Confidence and limitations

• Verified against exchange data. Every return figure in the tables comes from split- and dividend-adjusted end-of-day closes (Yahoo Finance). These are exchange-derived and I consider them near-primary. The 2018 capex-cut date and amounts trace to Micron's SEC 8-K.
• Directional, not verified. The margin-relief mechanism at the profit-and-loss level. I have Dell's fiscal-2021 statement that component costs were deflationary and the benefit was offset, but I could not extract segment gross-margin series for fiscal 2019, 2020 and 2023 from the filings within this brief's fetch budget. The claim "margin relief happens but is competed away" is directional.
• Directional, not verified. The direction of survivorship bias in R2 (below).
• Survivorship. My price set contains only names that still trade under a resolvable ticker. That is a real limitation and it cuts in a specific direction. Several members an ex-ante 2018 semicap screen would have included were acquired mid-window at premiums (Electro Scientific, Xcerra, Versum, Cabot Microelectronics are the ones I believe apply; I did not verify the dates against filings). Removing acquisitions-at-a-premium understates R2's return, which strengthens rather than weakens the "shorting semicap failed" conclusion. R1 is unaffected: no US-listed memory producer failed in either window, because the industry's bankruptcies (Qimonda 2009, Elpida 2012) predate both. R3 lost two members to data problems (Pure Storage returned no data; SMART Global / Penguin Solutions was not resolvable at the 2018 signal), and the second of those is the purest memory-module name in the rule, so R3's coverage gap is worse than its member count suggests.
• n=2. Two cycles, one rule, one asset class. The "capex cut lands at the equity trough" result is the most striking finding here and also the one most likely to be coincidence. It should be treated as a hypothesis to pre-register, not a calibrated edge.
• Overlapping regimes. The 18- and 24-month horizons from the 2022 ex-ante date run into the 2023-24 artificial-intelligence up-cycle. Those columns measure the next cycle as much as this one.
• In-sample risk. Both cycles are already in the vault's phase file and were already used to design memory-cycle-v1.1, per [[2026-09-09-hbm-plateau-hypothesis-backtest-variant]]. I did not construct any rule to fit them, but I knew their shape before I started.

Why this is in the vault

It closes the open follow-up "specify the Phase-3 rotation basket" from [[2026-09-25-hbm-dram-phase-2-3-transition-q3-2026]] with a negative result, which means the Notion ticket for that build should be re-scoped rather than executed: three of the four candidate expressions are struck on evidence, and the surviving work is a trim rule on the existing Phase-2 book plus a confirmable-date column on anchors/memory-cycle-v1.1/phase-history.csv. It also gives /investing:backtest-thesis a concrete defect to guard against, since every backtest built on the phase file currently enters on a date that was not knowable.

Open follow-ups

• Does a margin-peak trim rule beat the phase-trigger rule on the same two cycles? Researchable: Micron's quarterly gross margin and days-inventory are in filings and partly already in [[2026-09-07-memory-supplier-inventory-days-dio-anchor]]; the rule fires on reported numbers with a known report date, and the price test is the one already built here.
• What is the confirmable date for every transition in phase-history.csv? Researchable: each row already carries an evidence citation; the work is dating when that evidence was published, not judging it. It produces a second column, not a new artifact.
• Did PC and server original-equipment-manufacturer gross margins actually expand in fiscal 2019-20 and fiscal 2023-24? Researchable from Dell, HP Inc., HPE and NetApp segment disclosures. This separates "the margin relief did not happen" from "it happened and the market had already priced it," which are different diagnoses of the same failed trade.
• Does the capex-cut-marks-the-trough pattern hold in the 2001 and 2008-09 memory cycles? Researchable: the vault's phase file already dates both, and Micron's capex guidance history runs back through both in its filings. It takes n from 2 to 4 without any live-account exposure.
• Which names would a strict SIC-3559-plus-market-cap screen have returned on each signal date, including those since acquired or delisted? Researchable against EDGAR's point-in-time filer index, which is public and historical. It converts the survivorship caveat in this brief into a measured number.
• Is there an ex-ante-mechanical definition of "memory consumer" at all? Researchable via EDGAR full-text search over risk-factor language. If no defensible filter exists, R3 should be deleted from the menu rather than carried with a hand-built list.

Related

• [[2026-09-25-hbm-dram-phase-2-3-transition-q3-2026]]
• [[2026-09-26-financing-led-capital-cycle-phase3-precedent]]
• [[2026-09-09-hbm-plateau-hypothesis-backtest-variant]]
• [[2026-09-07-memory-supplier-inventory-days-dio-anchor]]
• [[2026-09-09-2018-capacity-announce-asp-rollover-lag]]
• [[2026-06-28-chip-memory-cycle-phase2-phase3-indicators]]
• [[2026-07-07-dram-hbm-phase2-phase3-early-signals]]
• [[project_investing_markov_capital_cycle]]

Sources

Vault

• ~/rdco-vault/06-reference/research/2026-09-25-hbm-dram-phase-2-3-transition-q3-2026.md
• ~/rdco-vault/06-reference/research/2026-09-26-financing-led-capital-cycle-phase3-precedent.md
• ~/rdco-vault/06-reference/research/2026-09-09-hbm-plateau-hypothesis-backtest-variant.md
• ~/rdco-vault/06-reference/research/2026-09-07-memory-supplier-inventory-days-dio-anchor.md
• ~/rdco-vault/06-reference/research/2026-09-09-2018-capacity-announce-asp-rollover-lag.md
• ~/rdco-vault/01-projects/investing/anchors/memory-cycle-v1.1/phase-history.csv
• ~/rdco-vault/01-projects/investing/anchors/memory-cycle-v1.1/phase-history-notes.md
• ~/rdco-vault/01-projects/investing/backtests/2026-05-18-memory-cycle-v1.1-v2-honest-rerun.md
• ~/rdco-vault/01-projects/investing/theses/2026-05-18-memory-cycle-v1.1.md

Primary / near-primary

• Micron Technology, Form 8-K exhibit 99.1, fiscal Q1 2019 results, filed 2018-12-18 — https://www.sec.gov/Archives/edgar/data/0000723125/000072312518000152/a2019q1exhibit991-pressrel.htm
• Dell Technologies, Form 10-K fiscal 2021 — https://www.sec.gov/Archives/edgar/data/1571996/000157199621000007/dell-20210129.htm (component-cost language; segment margin series not extracted)
• Yahoo Finance end-of-day adjusted closes, retrieved 2026-09-29, for all tickers listed in the universe-rule table. Computation scripts and raw output: /tmp/p3px/run.py, /tmp/p3px/detail.py, /tmp/p3px/raw.csv (ephemeral; re-runnable from the rules above).

Secondary

• EE Times, "Micron Cuts Capex, Wafer Starts" — https://www.eetimes.com/micron-cuts-capex-wafer-starts/