Memory Cycle v1.1 — executable thesis (post-backtest revision)
Supersedes [[2026-05-17-memory-cycle-v1]]. Backtest evidence (1 + 2 below) showed buy-and-hold beat all 3 mechanical-exit variants by 19-91pp on the 2021-2025 window, and the original 90d-return-proxy strategy underperformed by 135-192pp on 2019-2025. The thesis was right; the execution rules were wrong. v1.1 keeps the thesis + the tranche-entry logic, KILLS all mechanical price-based exits, ADDS anchor-based position-add logic, and expands the ticker basket per 2026 Q1 smart-money signal (SANDISK + INTEL).
Why v1.1 (the empirical follow-through)
Two backtest reports converged on the same finding:
- [[01-projects/investing/backtests/2026-05-17-memory-cycle-v1-walk-forward.md]] — walk-forward 2019-2025, 90d-return proxy as exit. Strategy returned +109% MU / +110% SMH vs buy-and-hold +302% MU / +245% SMH. -192pp / -135pp underperformance.
- [[01-projects/investing/backtests/2026-05-17-memory-cycle-v1-real-anchor-rerun.md]] — same window with real DRAM-spot anchor, 3 variants tested. Buy-and-hold beat all 3 by 19-91pp. The "real anchor" rescue did NOT materialize — exits at the cycle bottom hurt regardless of signal source.
Diagnosis from both reports converges: the tranche-accumulation logic is value-additive; the mechanical exit logic is the entire source of underperformance. Exiting a multi-year cyclical at a -15% / -20% / 2-consecutive-negative-quarter signal IS selling the bottom by construction.
Cross-corroboration: [[01-projects/investing/anchors/smart-money/2026-05-17-aggregate-2yr-backfill-summary]] — 2026 Q1 saw Druckenmiller + Tepper + Tiger Global all open SANDISK + INTEL positions in the same quarter as the 3-manager Amazon add. Three independent positive-sum managers crossed a structural-memory line at the same time the thesis says we're mid-cycle. That's the kind of corroborating signal v1's mechanical exits would have steamrolled.
The thesis (unchanged from v1)
HBM (high-bandwidth memory) supply is structurally short through 2027 because hyperscaler AI training and inference demand has outrun the 18-24 month fab + packaging cycle that gates new memory capacity. Every HBM wafer also crowds out ~3 commodity DRAM wafers (Micron's stated conversion ratio), so the tightness cascades from HBM into DDR5, DDR4, and eventually NAND. The memory-maker oligopoly (Micron, SK Hynix, Samsung) has pricing power for the first time since the 2017-2018 supercycle — except this time the demand driver is multi-year capex from hyperscalers with $750B+ of committed 2026 spend, not a cyclical consumer-electronics pulse. Buy MU (US-listed pure-play) tranche-accumulated, use SMH as the semis-basket proxy, hold through the cycle without mechanical exits, exit only on fundamental anchor break.
See v1 for the full bull case, demand-side anchor table, and earnings-inflection writeup. v1.1 inherits all of that — the structural argument did not change. What changed is how we trade against it.
Spec changes vs v1 (load-bearing)
1. KILL all mechanical price-based exits
| Removed from v1 | Why killed |
|---|---|
| Phase-marker 90d return < -15% exit | Backtest 1: fired at cycle bottoms in 2020 + 2022 + 2024; locked in losses 1-3 months before major rallies. Sensitivity cliff at -13.5% threshold. |
| DRAM spot trend ≤ 0 for 2Q exit | Backtest 2 Variant B: fired 2023-01-03 at the DRAM cycle bottom, missed +52% SMH rally over next 6 months. -91pp vs buy-and-hold. |
| DFP -15% from peak exit | Backtest 2 Variant C: 31 trades across both tickers, churned through positive-trend pullbacks. -19pp vs buy-and-hold even in the best parameterization. |
| Time-based stops | Never specified in v1 but ruling out for clarity. No "exit after N quarters" logic. |
| Any "trim to 1R on single-anchor flip" mechanical | v1 had this as a de-risking step. Replaced by founder-review-trigger (see below). |
There is no price-based or single-quarter-anchor-based exit rule in v1.1. Positions ride through drawdowns. The thesis itself has a stop (the fundamental anchor-break conditions below) — when the thesis stops, the position stops.
2. Keep tranche-entry logic (unchanged from v1)
Backtest trade logs showed T2/T3 entries firing at price points that subsequently produced 50%+ rallies (2020 COVID lows, 2025 April dip). The dollar-cost-average-into-pullbacks portion of the strategy is value-additive and stays.
| Tranche | Trigger | Add size |
|---|---|---|
| 1 (initial) | Founder APPROVE on /decisions/ page | 0.5R |
| 2 (scale-in) | -5% from tranche-1 avg price OR thesis-confirming catalyst (HBM capex announcement, hyperscaler capex guide-up, smart-money manager opens position) | +0.5R |
| 3 (full) | -10% from tranche-1 avg OR earnings catalyst with thesis-positive data | +1.0R (brings name to 2R) |
3. NEW: anchor-based position-ADD logic (the inversion)
This is the opposite shape from a stop loss. When fundamental anchors STRENGTHEN beyond the thesis baseline, ADD to the position up to a higher cap.
Add triggers (any one):
| Trigger | Add size | Confirmation |
|---|---|---|
| Hyperscaler capex guide revised UP >5% in any quarter | +0.5R per name (up to 3R cap) | Confirmed via earnings transcripts of 2+ of GOOG/AMZN/META/MSFT |
| DRAM spot contract +10% QoQ for 2+ consecutive quarters | +0.5R per name (up to 3R cap) | TrendForce monthly summaries |
| HBM cadence accelerating (vendor announces capacity bring-forward, OR new offtake at fixed price >2 years) | +0.5R per name (up to 3R cap) | MU / Hynix / Samsung earnings calls or press releases |
| Smart-money manager opens NEW memory position (2+ tracked managers same quarter) | +0.5R per affected name (up to 3R cap) | EDGAR 13F via /investing:smart-money-watch |
| MU/SMH down >20% from recent 90d peak AND ALL fundamental anchors still bullish | +0.5R per name (up to 3R cap) | Founder + Ray confirm anchors via quarterly review |
Drawdowns become BUYING signals when fundamentals confirm. This is the Druckenmiller-doctrine inversion: instead of "stop loss on a 15% drawdown," we read the drawdown as a discount on a structurally-bullish position and add.
4. Anchor-based EXIT logic (the only exits)
Required exit triggers — any ONE fires a founder-review gate; ANY TWO confirmed across 1-2 quarters = full position exit:
| Trigger | Severity | Action on first signal | Action on confirmation |
|---|---|---|---|
| Hyperscaler combined capex revised DOWN >10% across 2+ consecutive quarters | High | Founder review | Close ENTIRE bucket |
| Memory oversupply explicitly announced by 2+ vendors | High | Founder review | Close ENTIRE bucket |
| DRAM contract prices declining for 4+ consecutive months (sustained fundamental shift) | High | Founder review | Close ENTIRE bucket |
| HBM capacity online matching or exceeding demand projections (cycle-peak signal) | High | Founder review | Close ENTIRE bucket |
| 2+ tracked smart-money managers exit the same memory name in the same quarter | Medium | Founder review | Trim affected name to 1R, await second signal |
| Major memory maker announces capacity expansion that closes bottleneck in <18 months | Medium | Founder review | Trim to 1R |
| CXMT or YMTC achieves volume HBM3E-equivalent production | Medium | Founder review | Trim to 1R |
| Single hyperscaler signals 2027 capex cut >15% on earnings call | Medium | Founder review | Trim to 1R |
| Founder kill-switch (channel: "halt memory" / "pause investing") | n/a | Close immediately | n/a |
Single signal of any severity = Ray surfaces to founder via /decisions/ page within 7 days. Founder picks action. No autonomous exit on a single anchor flip.
5. Position size cap raised from 2R → 3R per name
To support the add-on-strength logic, max position per name lifts from 2R ($10k) to 3R ($15k). Memory bucket cap stays at 4R ($20k) unless founder explicitly approves a bucket expansion (which would trigger a fresh /decisions/ page).
Practical implication: if both MU and SMH ride to 2R each (full v1 sizing), they consume the entire bucket. Under v1.1, a single name CAN reach 3R via anchor-add logic — but the bucket cap of 4R means doing so reduces room for other names. The 4R cap is the binding constraint; per-name 3R is the headroom.
6. Ticker universe expansion: add SNDK + INTC
Per [[01-projects/investing/anchors/smart-money/2026-05-17-aggregate-2yr-backfill-summary]]:
- 2026 Q1: Druckenmiller + Tepper opened SANDISK positions (new) — first time SANDISK shows up on the smart-money-corroboration board for memory thesis
- 2026 Q1: Druckenmiller + Tiger Global opened INTEL positions (new) — broader semi/foundry exposure with memory adjacency
Updated memory bucket basket:
| Ticker | Role | Initial position | Max | Rationale |
|---|---|---|---|---|
| MU | Primary US-listed DRAM pure-play | 1R | 3R | HBM3E shipping NVIDIA; HBM4 ramps Q2 2026; 2026 supply pre-sold; capex $20B |
| SMH | Semis basket / ETF hedge | 1R | 3R | Captures HBM tailwind via Hynix + MU + NVDA without single-name risk |
| SNDK | NAND / memory tier-2 smart-money add | 1R | 3R | NEW v1.1 — 2026 Q1 Druckenmiller + Tepper entry (positive-sum capital-allocator signal); NAND tightness cascade from HBM crowd-out |
| INTC | Broad semi / foundry exposure (smaller) | 0.5R | 2R | NEW v1.1 — 2026 Q1 Druckenmiller + Tiger Global entry; smaller weight because Intel is broader-play than pure-memory |
Bucket math: initial deploy at 0.5R per ticker (consistent with tranche-1 schedule) = 2.0R = $10,000 initial. Full bucket at max per name + 4R cap = $20k total (cannot fully maximize all 4 names simultaneously — bucket cap binds).
Korean primaries (000660.KS Hynix, 005930.KS Samsung): still excluded due to Alpaca tradeability. MRVL / WDC / ON: defer to v2.
7. Quarterly review cadence (Ray drives, founder reviews 4x/yr)
Ray runs each calendar quarter end (March, June, September, December) and pulls the four-anchor state:
- Hyperscaler combined capex direction (GOOG/AMZN/META/MSFT earnings + 10-Q deltas)
- DRAM spot + contract price trend (TrendForce monthly summaries)
- HBM capacity-online cadence (vendor earnings + CES announcements)
- Smart-money 13F deltas for memory names (
/investing:smart-money-watchoutput)
Output filed to ~/rdco-vault/01-projects/investing/theses/reviews/<YYYY-Q[N]>-memory-cycle-v1.1.md. Surfaced to founder via channel + /decisions/ page IF any single anchor flips toward exit (per section 4 above). Otherwise filed silently to vault — no per-month nag, no founder interrupt.
Cross-domain translation (refresh)
The Snowflake credit-cycle 2020-2022 analog from v1 remains the load-bearing translation. Adding the Druckenmiller doctrine refinement.
Stanley Druckenmiller, per the Berens distillation we file in 06-reference/: "the way you make real money is to find the great opportunity and bet very heavily on it. Then let your winners run." His handling of NVDA 2023-2024 is the structural blueprint v1.1 adopts:
- Built position early in the cycle (Q4 2022 - Q1 2023, when conviction was high and consensus was bearish)
- Let it run through 2023 +200% rally without mechanical trims
- Trimmed only when capex direction softened (Q1 2024)
- Re-entered when fundamentals re-confirmed
That shape — early conviction sizing, ride the multi-year cycle, exit only on fundamental anchor break — is structurally identical to v1.1. v1 tried to be tactical inside the cycle. v1.1 sizes early, holds through volatility, and only adjusts on capex-direction signals or smart-money confirmation/disconfirmation. The Snowflake analog gives us the shape of the cycle; the Druckenmiller doctrine gives us the trading discipline against it.
The single fact that would force capitulation to bear case (unchanged from v1, restated): any single hyperscaler signing a 3+ year HBM offtake at fixed price WOULD validate structural demand. Conversely, ANY tier-1 hyperscaler cutting 2027 capex >15% on an earnings call = anchor-break signal that triggers founder review.
Position sizing (R-unit, v1.1)
Per parent README: 1R = $5,000. Memory bucket cap = 4R ($20k). Max per name = 3R ($15k) [raised from 2R in v1].
Initial deploy (this thesis, founder click-through gates)
| Ticker | Tranche 1 (initial) | Trigger |
|---|---|---|
| MU | 0.5R = $2,500 | Founder APPROVE on /decisions/ page |
| SMH | 0.5R = $2,500 | Founder APPROVE on /decisions/ page |
| SNDK | 0.5R = $2,500 | Founder APPROVE on /decisions/ page |
| INTC | 0.5R = $2,500 | Founder APPROVE on /decisions/ page |
Total initial deploy: 2.0R = $10,000 (paper).
Subsequent tranches (per name)
Per section 2 (tranche entry) and section 3 (anchor-based add). No mechanical exits. Profit "trims" from v1 are RETIRED — Druckenmiller doctrine says let winners run; trimming at +50%/+100% is the v1 mistake we explicitly walk back.
Disqualifying conditions (thesis-archive triggers)
Two-or-more of section 4's HIGH severity triggers confirmed across 1-2 quarters = thesis archived, bucket closed, post-mortem filed. Same logic as v1 but with much sharper anchor definitions and no single-anchor noise-firing.
Phase markers (informational only — NOT exit triggers in v1.1)
| Phase | Definition | Position posture |
|---|---|---|
| Phase 1 — Tightness (current) | All 4 anchors bullish: HBM sold out, capex flat-or-up, DRAM spot flat-or-up, smart-money adding | Tranche-accumulate; consider anchor-strength adds |
| Phase 2 — Plateau | 1 anchor flips, other 3 hold | Hold; surface to founder via review |
| Phase 3 — Rollover | 2+ HIGH severity anchors flip for 1-2 consecutive quarters | Close per disqualifying-condition logic |
These are descriptive labels for the quarterly review, NOT mechanical triggers.
Caveats / open risks
- Backtest sample is thin. Real-anchor rerun covers 2021-2025 (one full DRAM cycle). v1.1's hold-through-cycle conclusion rests on this 1-cycle sample. Walk-forward across 2017-2020 (DRAM 2018 crash) not testable until we backfill earlier DRAM spot data.
- Buy-and-hold beat tranche-with-exits, but tranche-only (entry, no exit) was NOT directly tested. v1.1's implicit assumption is that tranche-entry + buy-and-hold combines the entry edge with the hold edge. Untested — should be added to backtest queue.
- 2026 Q1 smart-money signal on SNDK + INTC is fresh (13F filings <60 days old). Reading the signal as structural may be premature — re-check next quarter's 13F to confirm positions held or added.
- Anchor-add logic untested. Section 3's "ADD on strength" pattern was not in v1 and has no backtest yet. Risk: over-concentration into a peaking cycle if anchor strength is confused with cycle top.
- 3R per name cap is bigger than 2R cap. Single-name drawdown risk goes up. Acceptable in paper; revisit before any live-capital migration.
Related
- [[2026-05-17-memory-cycle-v1]] — superseded; v1 frontmatter updated to status: superseded-by-v1.1
- [[2026-05-12-innermost-loop-ai-infrastructure]] — parent thesis (memory = chips-layer bottleneck sub-layer)
- [[01-projects/investing/backtests/2026-05-17-memory-cycle-v1-walk-forward]] — empirical evidence #1
- [[01-projects/investing/backtests/2026-05-17-memory-cycle-v1-real-anchor-rerun]] — empirical evidence #2
- [[01-projects/investing/anchors/smart-money/2026-05-17-aggregate-2yr-backfill-summary]] — corroboration source for SNDK + INTC adds
- [[01-projects/investing/README]] — operating-model boundary
- [[~/.claude/skills/investing-build-thesis/SKILL.md]] — SOP this thesis follows
- [[feedback_calibrate_overconfidence]] — applies: walk back tactical-cleverness in favor of conviction-sized hold
Open companion deliverables
/decisions/2026-05-18-paper-trade-memory-v1.1-go.html— paper-trade-go decision page (4-option click-back: APPROVE / REVISE / ARCHIVE / DEFER)_decisions.jsonentry — surfaces on HQ decisions index- Prior decision
2026-05-17-paper-trade-memory-v1-go→ archived with cross-reference to v1.1 - Quarterly review skill instrumentation — defer to v2 (manual quarterly review by Ray in interim)
- Backtest: tranche-entry-only variant (no exits) vs buy-and-hold — queue for next backtest cycle
Changelog
- 2026-05-18 (v1.1 post-backtest) — Supersedes v1. KILLED all mechanical price-based exits per backtest evidence (buy-and-hold beat all 3 mechanical variants 19-91pp). KEPT tranche-entry logic (value-additive per trade logs). ADDED anchor-based position-add logic (Druckenmiller doctrine — drawdowns are buying signals when fundamentals confirm). RAISED per-name cap from 2R to 3R. EXPANDED ticker basket to MU + SMH + SNDK + INTC per 2026 Q1 smart-money entries by Druckenmiller / Tepper / Tiger Global. ADDED smart-money 13F deltas as 4th anchor signal. Quarterly review cadence locked. Pending: founder deploy gate via /decisions/ page.