01-projects/investing/anchors/memory-cycle-v1.1

phase history notes

2026-05-18·investing-anchor·! medium
investinganchorphase-historymemory-cycle-v1.1

Memory cycle v1.1 — phase history notes

Companion to [[phase-history.csv]]. First full run, 2026-05-18. This file is the methodology + honest-limits log for the labeled transitions, written so a future re-labeling pass (or /investing:backtest-thesis v2) knows what was strong evidence vs what was inference.

Methodology used

Mechanical rules applied (where data permits)

Qualitative rules used where mechanical rules fail

Decision rules from thesis applied mechanically

Per phase_definitions in v1.1:

Coverage achieved

Cycle Transitions labeled High-confidence count Medium Low Notes
1997 5 0 4 1 Pre-data era; relying on industry retrospectives + DDR launch + consolidation events
2001 5 0 2 3 Worst-covered cycle; DOJ price-fixing dates are best mechanical anchor we have
2008 7 5 1 0 Best pre-TrendForce cycle (Qimonda bankruptcy = sharp, dateable signal; Micron 10-K data available)
2018 5 4 1 0 Strong vendor financial data (Micron FY18 peak GM) + SK Hynix fab dates available
2022 6 6 0 0 Best-covered cycle; TrendForce monthly + Micron capex-cut + Samsung production-cut all dateable to month
2024-current 5 4 1 0 Live cycle; TrendForce monthly data + smart-money 13F + vendor earnings all available
Total 33 19 (58%) 9 (27%) 5 (15%)

Aggregate: 33 transitions labeled across 6 cycles. 58% high-confidence; 85% high-or-medium confidence. Meets the skill's self-check threshold (>33% high-or-medium, >2/3 not-low).

Top 3 most-confident phase calls

  1. 2022 down-cycle entry (Mar 2022) — TrendForce monthly data shows the regime shift cleanly across Q1 2022; Micron capex cut Sep 2022 confirms; Samsung 50% production cut confirms. Multi-source convergence to within ~1 month.
  2. 2008 down-cycle confirmation (Sep 2008 + Jan 2009) — Qimonda's Jan 2009 bankruptcy is a sharp, dateable consolidation event; cumulative losses Q3 2007-Q4 2008 well-documented in industry retrospectives. The bottom is unambiguous.
  3. 2018 capacity-online + cycle-peak (Jun-Sep 2018) — Micron FY2018 GM peaks at 58.9% (extreme oligopoly pricing) immediately preceded by SK Hynix M15 Cheongju production start; Micron stock peaks ~$64 then falls 56% within months. Classic capacity-online → down-cycle transition with clear vendor-financial trace.

Known gaps + caveats for backtest-v2

Sparse pre-2017 data — handle with skepticism

The 1997 and 2001 cycles have only qualitative evidence (industry retrospectives + DOJ records + consolidation events). A backtest run against these labels should expect ~3-month uncertainty on transition dates and should not place strategy decisions on tight phase boundaries pre-2010. The 2008 cycle is better because Qimonda + Micron 10-K + Elpida bailout dates are sharp, but even there, the recovery-to-demand-recognition transition (~2010-06) is medium-confidence at best.

Hyperscaler-capex anchor unusable pre-2014

The v1.1 thesis uses hyperscaler capex direction as anchor #1 — but hyperscalers (AWS, Azure, GCP, Meta) were not the dominant DRAM demand driver until ~2014-2015. Pre-2014 cycles cannot be re-labeled using this anchor; the labels rely on PC/enterprise demand commentary instead. Implication: cycles 1997, 2001, 2008, and the early phase of 2018 are structurally different from current cycles (different demand driver = different dynamics). A backtest spanning 1997-2026 is implicitly comparing two regimes.

Regime change risk — 2024-current cycle may NOT be cyclical

Per founder's v1.1 framing and the Castellano "HBM Has Broken the Memory Cycle" hypothesis, the AI-driven HBM tightness may be structurally different from prior cycles. Specifically: vendors are pursuing "minimize oversupply risk" capex discipline (Samsung explicit statement, Q1 2026 earnings) rather than the classic build-flood-crash cycle. This means the 2024-current cycle may NOT have a cleanly dateable down-cycle entry coming — if the founder's thesis is right, capacity discipline holds and the next transition is plateau, not crash. The backtest harness should treat the current cycle as a "live, may-not-complete" cycle and not force a synthetic crash date.

TrendForce data starts Feb 2021 — earlier gap

The DRAM-spot CSV nominally claims 2017-2026 coverage in its filename, but the actual data starts Feb 2021. Pre-2021 DRAM-spot labeling relied on retrospective industry sources, not the structured monthly scraper output. If a future re-labeling pass wants tighter 2017-2020 transition dates, scraping TrendForce news archives 2017-2020 or pulling DRAMeXchange historical data would help.

Multiple-raters disagreement: 1-2 months expected, larger for pre-2010

Within 2017-2026 (mechanical rules apply): two researchers using the v1.1 phase_definitions and TrendForce data would agree on transitions to within ~1 month. Within 2008-2016 (vendor 10-K + retrospective sources): ~3-month agreement. Pre-2008: ~6-month agreement is realistic, possibly worse. Backtest sensitivity analysis should test phase-window edges at ±1 month (recent cycles) and ±3 months (older cycles).

Smart-money anchor only goes back to 2024 Q2

The smart-money-watch v0 backfill covers 2024 Q2 - 2026 Q1 (8 quarters). For the 2018 cycle and earlier, smart-money 13F evidence was NOT used as a phase anchor — would require a separate historical 13F backfill (3-5 manager 5-year backfill is the natural next step). For the 2024-current cycle, the 2026 Q1 multi-manager SANDISK + INTEL entries support the demand-recognition phase but are not load-bearing alone.

Phase boundaries are not sharp

This is the deepest honest caveat. Phases are labels imposed on continuous underlying state. The "down-cycle" of 2022 began with first price-down signals Dec 2021, dominated by Mar 2022, was deepest by Q1 2023 (Micron GM -32.7%), and "ended" sometime Jul-Nov 2023. Picking any single transition date is a choice. The CSV picks transitions where the founder's defined mechanical rules from phase_definitions first fire, which is the most defensible default.

What the backtest-v2 should do with this

  1. Use the high-confidence transitions as primary phase boundaries. That's ~19 of the 33 entries. The 5 low-confidence pre-2001 entries are inputs the backtest can choose to fold into adjacent transitions if they're not load-bearing.
  2. Sensitivity-test against ±1 month (recent) / ±3 months (older) windows. If thesis P&L is sensitive to a 1-month phase-boundary shift, the strategy is over-fit to the labels and should be reconsidered.
  3. Treat 1997, 2001 cycles as supplementary, not primary. The 2008, 2018, 2022, and 2024-current cycles are the four with usable mechanical phase boundaries. Backtest performance on those four cycles is the meaningful signal.
  4. Treat 2024-current as a live, may-not-complete cycle. Don't assume a down-cycle transition is coming on a fixed clock. The Castellano/Samsung-discipline hypothesis is that this cycle may stay in Phase 2 plateau structurally longer than prior cycles.

Open follow-ups

Sources used in this labeling pass

Changelog