06-reference/research

hbm plateau hypothesis backtest variant

2026-09-09·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
memory-cyclebacktesthbminvestingfalsifiability

The plateau variant is specifiable, but as written its exit rule cannot fire — and it has n=1 testable cycle

The question

Verbatim: "Test the 'HBM has broken the cycle' plateau hypothesis as an explicit backtest variant (current cycle does not crash; exit only on China-entrant disconfirmation or smart-money exit)."

Context: this is open follow-up #5 of [[2026-06-28-chip-memory-cycle-phase2-phase3-indicators]], itself inheriting a follow-up from the memory-cycle v1.1 phase-history notes. The lead RDCO investing thesis is the chip-fab/memory capital cycle at position/capital-cycle horizon, currently judged Phase 2.

What we already know (from the vault)

What the web says

Convergences and contradictions

Synthesis for RDCO

The variant is specifiable, and here is the spec. Universe: point-in-time US-listed memory names per the v1.1 construction rule (MU, SMH, SNDK, INTC as the live basket; WDC, STX, MRVL in the backtest universe). Entry: unchanged from v1.1 tranches (0.5R initial, scale-in at -5%, full at -10%, 1R = $5,000, 4R bucket cap). Exit, as literally worded in the question: (a) China-entrant disconfirmation, or (b) smart-money exit. Sizing, benchmarks and reporting inherit the [[investing-backtest-thesis]] v2 contract — dual benchmarks (buy-and-hold universe and SMH/SPY), mandatory leave-one-out, trade-count-aware precision, no 2-decimal Sharpe below 50 trades. Mode A, fixed-rule, no parameter fitting.

But as literally worded, neither exit leg can function. Leg (a) is arithmetically unreachable on DRAM: the September 8 derivation says no CXMT bit share triggers Phase 3 against strong demand, so scoped to DRAM the leg never fires. Scoped to NAND it has already fired, because YMTC crossed 14% in August. A rule that is either never-true or already-true depending on an unstated scope is not a rule; it is a placeholder. Leg (b) is weaker still: in v1.1, "2+ tracked smart-money managers exit the same name in the same quarter" is a MEDIUM-severity signal that trims to 1R and awaits a second signal, and 13F filings arrive roughly 45 days after quarter end. Against an asset class that has repeatedly fallen 50-80% inside 12-18 months, a lagged trim-to-1R is not an exit. So the honest characterisation of the variant as specified is: an unconditional hold with decorative exits. That may still be the right posture. It should be named as such rather than dressed as a rule set.

The statistical ceiling is the finding, and it is hard. This variant cannot be backtested in any meaningful sense. Its claim is about one cycle, the live one, and n=1. Every prior cycle in phase-history.csv terminated in a dateable down-cycle by construction, so running the plateau variant over 1997, 2001, 2008, 2018 and 2022 does not test the hypothesis; it measures the cost of being wrong about it. That counterfactual is worth computing and is already latent in the honest rerun: the buy-and-hold-universe column is the never-exit case, and it returned +86.7% / +38.0% / +90.4% / +286.6% / +14.5% / +1127.1% across the six windows, beating the v1.1 rules in 4 of 6. That is real evidence for the posture (holding beat trading) and none at all for the hypothesis (that this cycle does not crash). Two caveats keep it from being stronger than that: the windows are phase-aligned and end after recovery, which flatters a holder who never had to mark to market at the trough, and the honest rerun does not report per-cycle maximum drawdown for the buy-and-hold benchmark, so the survivability of the hold is unmeasured. Per [[2026-06-01-cpcv-deflated-sharpe-autoinv-validation]], the correct output for a book at this horizon is "here are the cycles, one by one, and here is the leave-one-out" — not a deflated Sharpe. Wiring any Sharpe-family statistic to a one-cycle hypothesis would be exactly the false rigor the v2 harness was built to forbid.

What this changes. The marginal question is smaller than the follow-up implied. Path A of the honest rerun (buy-and-hold plus HIGH-severity anchor kill switch) is already most of the plateau variant. The delta between Path A and the plateau variant is only whether the two named legs replace the four HIGH-severity anchor triggers, and with n=1 that comparison cannot be resolved empirically. The defensible move is to convert the variant from a backtest into a pre-registered forward test: write the exit rule down now, with the scope ambiguity resolved and a price-based backstop added, timestamp it before the end-September 2026 quarterly review, and let the live cycle produce the seventh data point. That does not require and does not imply any paper-trade deployment; it is a documentation act. Three spec defects must be inherited deliberately if a run happens anyway: the days-inventory-outstanding unit is incompatible with the "5-6 weeks" tripwire row and will read as permanently triggered if dropped in; the China row must split into 5a (DRAM/CXMT, joint condition) and 5b (NAND/YMTC, already crossed); and the LTA leg must be marked unresolved rather than asserted.

Why this is in the vault

It closes open follow-up #5 of [[2026-06-28-chip-memory-cycle-phase2-phase3-indicators]] and the matching follow-up in the memory-cycle v1.1 phase-history notes, and it determines what the end-September 2026 quarterly anchor review is allowed to conclude: specifically, that the choice between holding through and exiting on schedule cannot be settled by any run of the [[investing-backtest-thesis]] harness, so it stays a founder judgment call rather than becoming a harness output.

Open follow-ups

  1. Ingest Castellano's primary piece. The hypothesis has driven four vault documents on the strength of its title. Read "HBM Has Broken the Memory Cycle: Why This Shortage Is Different," extract the actual mechanisms and any thresholds the author names, and check whether he has revised it since.
  2. Resolve the LTA contradiction. Does a 3+ year fixed-price hyperscaler HBM offtake exist as of September 2026? Primary source (10-Q/10-K commitment disclosures, earnings-call transcripts), not trade press. This is the single fact that most moves the plateau prior.
  3. Compute per-cycle maximum drawdown for the buy-and-hold-universe benchmark in the honest rerun. Without it, "holding beat trading in 4 of 6 cycles" is a return statement with no survivability statement attached.
  4. Write the pre-registered forward-test rule with the China leg scope-resolved (DRAM vs NAND), a price-based backstop, and a hard timestamp, so the live cycle yields a clean seventh observation instead of a retrospectively-rationalised one.
  5. Decide whether the plateau variant should replace or supplement the four HIGH-severity anchor triggers. Framed as a decision with named consequences, since it cannot be framed as a test.
  6. Refresh the pricing anchor. The freshest DRAM/HBM pricing in any vault memory document is July 2026, leaving a two-month blind spot as of today.

Related

Sources

Vault

Web