01-projects/investing/backtests

memory cycle v1.1 v2 honest rerun

2026-05-18·investing-backtest
investingbacktestv2memory-cycle-v1.1honest-rerunmulti-cyclesurvivorship-free

Memory cycle v1.1 — v2 honest multi-cycle backtest

First execution of the /investing:backtest-thesis v2 SOP. Replaces v1's calendar-fold methodology with multi-cycle phase-aligned windows, per-window universe construction, anchor-data-driven phase markers, and leave-one-out reporting. The conclusion is uncomfortable: the v1.1 strategy DOES NOT beat buy-and-hold of its own universe. It does beat broad-market benchmarks (SMH, SPY) on aggregate, but only because the underlying memory basket beats them — not because the rules add alpha.

Headline (honest)

Mean strategy return across 6 cycles: +246.8% (95% bootstrap CI: +15.9% to +649.5%) Mean BH-universe-equal-weight: +273.9% (95% bootstrap CI: +50.3% to +632.1%) Strategy UNDERPERFORMS BH-universe by ~27pp on the all-cycles mean.

The strategy beat BH-universe in only 2 of 6 cycles (2008 +17pp, 2024-current +96pp). It lost to BH-universe in 4 of 6 (1997 -103pp, 2001 -13pp, 2018 -149pp, 2022 -11pp).

Leave-one-out collapses the headline. Drop the 2024-current AI-supercycle window and aggregate strategy return drops from +246.8% to +51.6%. Drop the same window from BH-universe and it drops from +273.9% to +103.2%. Without the AI supercycle, both strategy and benchmark are unimpressive — and strategy is HALF of BH-universe.

The strategy DOES beat the broad-market benchmarks on aggregate: vs BH-SMH +149pp on mean, vs BH-SPY +192pp on mean. But this is the "you were just along for the AI ride" critique made visible — beating SMH is mostly a side effect of memory beating broad semis, not a v1.1 rules win. Removing the 2024-current cycle, strategy mean drops to +51.6% vs SMH +74.6% (also unavailable) — and the alpha vs broad benchmarks disappears too.

Verdict: this is a buy-the-AI-supercycle bet with extra trades. The v1.1 rules do not add alpha vs buy-and-hold of the same universe.

Per-cycle results

Cycle Window Universe Tradable Strategy BH-univ BH-SMH BH-SPY Strat vs BH-univ Trades
1997 1996-01 → 2000-12 MU, TXN, HXSCY 2/3 -16.8% +86.7% -50.3% +3.0% -103.4pp 14
2001 2001-01 → 2006-12 MU, IFNNY 2/2 +25.4% +38.0% +15.3% +55.5% -12.6pp 13
2008 2007-01 → 2013-12 MU, IFNNY, SPSN 3/3 +107.4% +90.4% +63.4% +85.7% +17.0pp 15
2018 2016-01 → 2020-12 MU, WDC, STX, MRVL 4/4 +138.0% +286.6% +307.2% +94.8% -148.6pp 27
2022 2021-01 → 2023-12 MU, WDC, STX, MRVL 4/4 +3.8% +14.5% +51.9% +32.1% -10.6pp 26
2024-current 2024-01 → 2026-05 MU, WDC, SNDK, STX, MRVL, INTC 6/6 +1222.7% +1127.1% +198.8% +55.3% +95.6pp 37

Strategy wins vs BH-universe: 2/6. Strategy wins vs BH-SMH: 4/5 (SMH inception ~2000). Strategy wins vs BH-SPY: 3/5.

Per-ticker color (2024-current cycle = entire AI-boom outlier)

The 2024-current cycle dominates the entire backtest. Per-ticker:

The SNDK number especially is structurally suspect — 15 months of price data captured in a once-per-generation AI-storage repricing event. The 1222.7% strategy figure on this cycle should be read as "the AI supercycle happened" not "the rules captured it."

Aggregate metrics with CI

Metric Value 95% Bootstrap CI
Mean strategy return +246.8% (+15.9%, +649.5%)
Mean BH-universe +273.9% (+50.3%, +632.1%)
Mean BH-SMH +97.7% (not computed - missing 1997 cycle)
Mean BH-SPY +54.4% (not computed - missing 1997 cycle)

Confidence interval interpretation: the CI on strategy mean spans +16% to +650%. That's not a precise measurement — that's "we cannot statistically distinguish this from a moderate positive return OR a spectacular one." With N=6 cycles and 2024-current being a massive outlier, no further precision is honest.

Per the v2 SOP "no 2-decimal Sharpe on <10 trades" rule: Sharpe is not reported. Per-cycle trade counts: 13-37, but each "trade" is a tranche entry/exit within a single position. Effective trade count is ~24 (4 per cycle × 6 cycles) — well below the 50 threshold for 2-decimal precision.

Leave-one-out check

This is the diagnostic the founder mandated. Drop each cycle in turn and recompute the aggregate:

Dropped cycle Strategy mean (N=5) Delta vs all-6 BH-univ mean (N=5) Note
1997 +299.5% +52.7pp +311.3% Drop = better (1997 was strategy's worst loss)
2001 +291.0% +44.3pp +321.1% Drop = modestly better
2008 +274.6% +27.9pp +310.6% Drop = strategy loses a win
2018 +268.5% +21.8pp +271.3% Drop = strategy loses big underperformance
2022 +295.4% +48.6pp +325.8% Drop = modestly better
2024-current +51.6% -195.2pp +103.2% Drop = aggregate COLLAPSES

Headline-changing finding: dropping the 2024-current cycle takes strategy mean from +247% to +52%, and BH-universe from +274% to +103%. Without the AI supercycle, both strategy AND benchmark drop to "ordinary" returns, AND the strategy is HALF of BH-universe (52% vs 103%). The entire "strategy is comparable to BH" framing is propped up by one extreme outlier.

This matches v1's critique: a single dominant cycle is doing the heavy lifting, and the strategy's mechanical rules are not what made it work.

Bootstrap confidence intervals

Bootstrap (5000 iterations, with replacement) on per-cycle returns:

Both CIs span an order of magnitude. The honest reading: with N=6 cycles and one extreme outlier, no statistical statement about "strategy > or < BH-univ" can be made with confidence. Distribution is too wide.

Sensitivity sweep

Drawdown-add threshold (key v1.1 parameter — "add 0.5R on >20% drawdown with bullish anchors"):

Threshold Mean strategy return
-16% +237.8%
-20% (locked default) +246.8%
-24% +241.6%

Strategy is NOT sensitive to this parameter (±5pp on mean across the sweep). Good news: not overfit to this specific number. Bad news: this is because the parameter rarely fires meaningfully — most of the action is in the entry tranches and the no-mechanical-exit hold-through, not the drawdown-add logic.

Regime-change diagnostic

Pre-2014 vs post-2018 cycles, tested separately:

Era Cycles Strategy mean BH-univ mean Strategy vs BH
Pre-2014 1997, 2001, 2008 +38.7% +71.7% -33.0pp
Post-2018 2018, 2022, 2024-current +454.9% +476.1% -21.2pp

Verdict: same conclusion in both regimes. Strategy underperforms BH-universe by similar margin (~25-33pp) in BOTH the pre-hyperscaler era (1997-2008) and the post-2018 era. The "regime change" hypothesis (pre/post-hyperscaler memory cycles are structurally different) is supported by the level of returns being radically different (38% vs 455%) — but the strategy-vs-BH-univ DIFFERENTIAL is consistent across regimes. The rules don't add alpha in either era.

Founder's "Castellano broke the cycle" hypothesis — that 2024-current is structurally different from prior cycles — is supported by the magnitude difference (1222% vs 107% for the next-best 2008 cycle). But it doesn't rescue the v1.1 rules.

Thesis vs strategy separation

Did the anchor data behave as the thesis predicts? (THESIS validation)

Yes — and this is the key finding the founder should preserve. The phase-history.csv shows 6 cycles each following the demand → capacity-announce → capacity-online → down-cycle → recovery sequence in the order the v1.1 thesis predicts. The 2018 cycle (Micron FY18 GM 58.9% peak → 56% stock crash within months) and 2022 cycle (TrendForce monthly + Micron capex-cut + Samsung production-cut all dateable to month) are particularly clean confirmations. The thesis structure IS empirically supported by 6 cycles of memory-industry history.

The 2024-current cycle is the one that may NOT follow the prior pattern (Samsung "minimize oversupply risk" stance, no clear down-cycle entry yet, HBM offtakes 2+ years out). That's the Castellano "HBM broke the cycle" question and it remains live.

Did the rules capture the predicted moves? (STRATEGY validation)

No. The v1.1 rules (tranche-in at phase-1 demand-recognition, hold through cycle, exit only on anchor break) DO NOT outperform buy-and-hold of the same universe at the same entry date. The tranche-add logic creates a slightly worse average entry price (waiting for -5%/-10% drawdowns means missing some immediate rallies), and the hold-through logic is equivalent to BH by construction.

The v1.1 thesis is right. The v1.1 RULES don't add value beyond just buying the universe at the same date.

Survivorship-bias confession

What gaps exist in the universe data, by cycle:

Cycle Intended universe Available via yfinance Excluded (gap)
1997 MU, TI memory unit, Hyundai ADR, pre-merger Samsung, Toshiba ADR MU, TXN HXSCY (Hyundai Electronics ADR — yfinance shows symbol delisted/missing); Samsung/Toshiba not US-listed at the time
2001 MU, Infineon, Hynix ADR, ProMOS, Mosel Vitelic MU, IFNNY Hynix ADR (different ticker pre-2012); ProMOS/Mosel never US-listed
2008 MU, Qimonda, SK Hynix ADR (post-restructuring), Elpida, Powerchip, Spansion MU, IFNNY (Qimonda parent), SPSN Qimonda (bankrupt Jan 2009; delisted before yfinance coverage); Elpida (6665.T, not US-listed); Powerchip (TW); Hynix ADR pre-2012
2018 MU, SK Hynix ADR, WDC, STX, MRVL MU, WDC, STX, MRVL Hynix ADR (HXSCL — yfinance unreliable)
2022 Same as 2018 + SNDK post-spin MU, WDC, STX, MRVL SNDK didn't exist yet (spun Feb 2025)
2024-current + SNDK + INTC All 6 available None

What this likely does to results:

Honest gap acknowledgment: the surviving-name bias is structural and not fixable without paid data sources. Reported numbers should be read as "this is what the strategy did on the names that have continuous US-listed price data" — which is itself a real-world tradeable constraint and not entirely a methodology defect.

"Is this overfit?" diagnostic

Three checks:

  1. Parameter sensitivity (sensitivity sweep above): strategy varies ±5pp across ±20% sweep on drawdown-add threshold. NOT overfit to this parameter. Low signal here.
  2. Cycle-by-cycle consistency: strategy beats BH-univ in 2/6 cycles. NOT consistent — the rules do not reliably add value cycle-after-cycle. If the rules captured a structural truth, we'd expect 4/6 or 5/6 wins. 2/6 is barely better than coin-flip and is fully explained by one dominant cycle.
  3. Out-of-sample validation: the v1.1 rules were designed AFTER seeing v1's 2019-2025 results. The 2018-2022 cycles in this backtest are essentially in-sample. The pre-2014 cycles (1997, 2001, 2008) are the closest to true out-of-sample — and the strategy loses to BH-univ in 2 of 3 of those (1997 -103pp, 2001 -13pp, 2008 +17pp = mean -33pp). Out-of-sample evidence supports the "no alpha" verdict.

Verdict: the strategy is not overfit in the parameter-sensitivity sense, but it doesn't have a real edge either. It's not memorizing noise; it's just not doing better than buying and holding the same names.

Caveats

  1. Hyperscaler-capex regime change — pre-2014 cycles cannot use the v1.1 thesis's #1 anchor (hyperscaler capex direction). The 1997, 2001, and 2008 cycles use PC OEM and consumer demand as substitute drivers. Per phase-history-notes: this means the strategy under test in those cycles is not quite the same strategy as in the 2018+ era. Honest read: the strategy is being tested against a different (worse for it) demand regime.

  2. Pre-2017 DRAM spot opacity — TrendForce monthly DRAM spot data starts Feb 2021. Pre-2021 phase markers come from earnings retrospectives + bankruptcy dates + DOJ records. Phase boundary uncertainty is ±3 months pre-2010 vs ±1 month post-2017. Strategy entries snapped to phase-1 demand-recognition dates with this uncertainty baked in.

  3. Castellano "HBM broke the cycle" hypothesis — the 2024-current cycle may NOT have a down-cycle transition coming. Samsung's "minimize oversupply risk" Q1 2026 stance, HBM4 logic-die prices +40-50%, multi-year offtakes — these are evidence the cycle is structurally different from prior. If true, the strategy under test (which assumes cyclical exits) may be testing a no-longer-applicable framework. Backtest result reflects 2024-2026 mark-to-market only; live cycle outcome unknown.

  4. SNDK is a 15-month dataset in a once-per-generation event — the +4151% per-ticker number is real but should not be treated as a sustainable repeating signal. Removing SNDK from the 2024-current cycle would meaningfully reduce the 1222% cycle return.

  5. Survivorship gaps — Qimonda (bankrupt 2009), Elpida (bankrupt 2012), Hyundai Electronics ADR (delisted) are NOT in the backtest. v2 SOP requires including bankruptcies as "forced exits" — the harness does this for names with continuous yfinance data, but cannot include names yfinance never had. Cycle results may be modestly inflated as a result.

  6. N=6 cycles is small. Bootstrap CIs span an order of magnitude. No strong statistical conclusion is possible.

  7. Mode A, not Mode B. This is fixed-rule out-of-sample replicated across 6 cycles — NOT walk-forward parameter re-estimation. The v1.1 rules were locked at thesis-spec time (2026-05-18). No parameter optimization happens within the backtest. Per v2 SOP this is honest about what walk-forward actually means.

Recommendation for paper deployment

DO NOT DEPLOY v1.1 AS A MECHANICAL STRATEGY. The honest answer is: buy-and-hold the same universe wins.

Specifically:

  1. The v1.1 rules do not add alpha vs BH of the same universe on the all-cycles mean (-27pp), on 4 of 6 individual cycles, and especially without the 2024-current outlier (-52pp).

  2. The thesis itself remains supported. 6 cycles of phase-history are consistent with the v1.1 thesis structure. The error is in trying to be tactical within the cycle.

  3. Two paths forward, founder picks:

    Path A: Deploy thesis as buy-and-hold + thesis-archival kill switch. This is the base case the founder mandated and v1.1 should have collapsed to. Buy MU + WDC + SNDK + STX + MRVL + INTC equal-weight at thesis-confirmation date. Hold. Exit only on the anchor-break HIGH-severity triggers from section 4 of the thesis (2 of 4 confirmed). This captures the thesis-is-right finding without paying the tranche-timing cost.

    Path B: Iterate to v1.2 with explicit edge-source hypothesis. Before another mechanical-rule backtest, the v1.2 spec must answer: WHAT specific edge are the rules trying to capture beyond buying and holding? Tranche-timing on -5%/-10% drawdowns demonstrably does not add value. Anchor-strength adds rarely fire and don't help. If there's no testable edge hypothesis, there are no rules to backtest.

    My recommendation (Ray): Path A. The honest finding is that the thesis is right, the strategy rules are not. Drukenmiller doctrine taken to its logical conclusion IS buy-and-hold + thesis-archival kill switch. The v1 backtest already showed this; the v2 multi-cycle test confirms it across regimes. Continuing to refine mechanical rules is the gold-plating trap.

  4. If founder still wants to deploy v1.1 paper-trade, do so for INFORMATION VALUE only. Live execution will produce a 7th cycle's worth of data and reveal whether the live cycle behaves like 2024-current (AI supercycle continues) or like 1997/2018 (mean reversion to "strategy loses to BH"). But do not deploy with capital expectations that v1.1 will beat BH-universe — the evidence says it won't.

Cost / scope summary

Related

Changelog