01-projects/investing/backtests

memory cycle v1 real anchor rerun

2026-05-17·investing-backtest·status: completed·! medium
investingbacktestmemory-cyclereal-anchordram-spotdfpmusmh

Memory v1 Re-Test with Real Anchors — Comparison Run

Headline

Buy-and-hold STILL wins. Real DRAM-spot anchor did not rescue the strategy. The 192pp underperformance in the original 90d-return-proxy backtest was NOT primarily a proxy artifact — the core problem is "exit a multi-year cyclical at the bottom" regardless of what signal you use to define the bottom.

Variant Portfolio TR Sharpe Max DD Exits vs A delta
A. Buy-and-hold (control) +228.79% 0.81 -57.6% 0 —
B. DRAM-spot-driven exit +137.85% 0.64 -57.6% 1 (each) -90.94pp
C. DFP-from-peak + anchor re-entry +209.79% 0.81 -52.5% 23/8 -19.00pp

Per-variant results

Variant A — Buy-and-hold (control)

Ticker Start End Return Sharpe Max DD
MU $85.94 (2021-02-25) $292.51 (2025-12-30) +240.35% 0.76 -57.6%
SMH $114.53 $363.31 +217.22% 0.86 -45.3%
Equal-weight portfolio +228.79% 0.81 -57.6%

No exits. Drawdowns are severe (MU -57.6% in 2022, SMH -45.3%) but the recovery in 2023-2025 dwarfs the drawdown loss.

Variant B — DRAM-spot-driven exit (real anchor)

Exit when DRAM spot quarterly trend ≤ 0 for 2 consecutive quarters AND price ≥10% off recent peak. Re-enter when DRAM spot trend > 0 for 1 quarter.

Ticker TR Sharpe Max DD Exits Trade dates
MU +167.75% 0.67 -57.6% 1 exit 2023-01-03 @ $49.63 / reentry 2023-07-03 @ $63.08
SMH +107.95% 0.62 -45.3% 1 exit 2023-01-03 @ $99.34 / reentry 2023-07-03 @ $151.53
Portfolio +137.85% 0.64 -57.6% 2 total

Critical failure pattern. The 2-consecutive-quarter persistence filter (2022Q2 and 2022Q4 both -1.0) finally triggered an exit on 2023-01-03 — exactly at the DRAM cycle bottom. The strategy then waited for the DRAM trend to flip positive (which it did in 2023Q2 at +0.333), but the actual quarter-end reading wasn't available until early July 2023, by which time SMH had already rallied +52% off the bottom ($99 → $151). The strategy locked in a real loss and missed the entire recovery leg.

The Max DD doesn't improve (-57.6% MU / -45.3% SMH) because the exit fired AT the bottom — it captured the full drawdown then locked it in by exiting.

Variant C — DFP-from-peak with thesis-anchor re-entry

Exit at -15% from running peak. Re-enter when price recovers +10% from trough OR DRAM trend turns positive.

Ticker TR Sharpe Max DD Exits Notes
MU +195.18% 0.71 -52.5% 23 Hyperactive — DRAM trend was positive in 13/19 quarters, so "OR DRAM positive" re-entry fires almost every day-after exit
SMH +224.41% 0.91 -44.0% 8 Marginally beats buy-hold on SMH (+224% vs +217%)
Portfolio +209.79% 0.81 -52.5% 31 total

Hyperactivity bug. The "OR DRAM trend positive" re-entry condition uses the prior quarter's average direction_numeric. Because DRAM was structurally positive in 13 of 19 quarters in the test window, the re-entry condition is almost always satisfied. Consequence: most exits get reversed on the very next bar. This effectively turns Variant C into "buy-and-hold with a tax on every 15%+ drawdown" — and the marginal improvement on SMH (+7pp) is offset by an underperformance on MU (-45pp) because MU has more 15%+ pullbacks within positive-trend quarters than SMH does.

If you flip C's re-entry to AND (instead of OR), exits become sticky and the variant would behave like B (sell at bottom, wait for re-entry). The OR formulation is the structural reason Variant C only loses 19pp instead of 91pp.

Sensitivity sweep

Variant B sweep

Parameter Sweep Portfolio TR
persistence (consecutive negative quarters) 1 +141.74%
persistence 2 (base) +137.85%
persistence 3 +228.79% (equals buy-and-hold — rule never triggers)
reentry_persistence 1 (base) +137.85%
reentry_persistence 2 +134.85%
reentry_persistence 3 +97.34%
price_drop 0.05 to 0.20 flat at +137.85% (insensitive)

B is a knife-edge. persistence=3 reverts to buy-and-hold (the 2 negative quarters in 2022 weren't enough), but persistence=1 or 2 always fires the same exit on 2023-01-03 because by then the cycle was deep. price_drop is irrelevant because the price had already declined more than any reasonable threshold by the time the persistence condition was met. There is no parameter combination of B in the swept range that BEATS buy-and-hold — the best you can do is set persistence high enough that the rule never fires.

Variant C sweep

Parameter -20% -10% 0% (base) +10% +20%
dd_trigger +204.14% +219.14% +209.79% +193.48% +185.79%
reentry_recovery +217.09% +213.46% +209.79% +199.16% +189.38%

C is reasonably robust — returns range from +186% to +219% across ±20% on either parameter, no cliff. But the BEST sensitivity point (+219% with dd_trigger=-13.5%) still loses to buy-and-hold (+229%). The robustness verdict is "stable but suboptimal." No parameter combination in the swept range matches or beats Variant A.

"Did the real anchor change the conclusion?"

No — it confirmed the original backtest's structural finding while ruling out the proxy as the primary culprit.

Comparison Original (90d-return proxy) Real-anchor rerun
Period 2019-2025 (7 folds) 2021-2025 (constrained by DRAM data start)
Active-strategy TR MU +109% / SMH +110% B: +138% portfolio; C: +210% portfolio
Buy-and-hold TR MU +302% / SMH +245% A: +229% portfolio
Underperformance -192pp MU / -135pp SMH B: -91pp / C: -19pp

The proxy backtest fired multiple times per year (25-26 total trades) and was discontinuous — the prior report identified a knife-edge at the phase-marker threshold. The real-anchor backtest fires more conservatively (B = 1 exit per ticker; C = 8-23 exits) and is mostly smooth, but still cannot beat buy-and-hold. The original report's diagnosis was correct in spirit but understated the structural problem: even with a FAITHFUL fundamental anchor (real DRAM spot trend), exiting a multi-year cyclical at the bottom is a losing trade. The 91pp underperformance of Variant B is not a proxy artifact — it's the inherent cost of having ANY rules-based exit on a thesis with this drawdown profile.

The narrower gap (-91pp vs -192pp originally, scaled for the shorter window) is partly because the test window starts in 2021 (post-COVID-bottom shake-out is excluded) and partly because the real anchor genuinely fires LESS often than the price proxy did. But the sign and shape of the conclusion is identical: exits hurt; the tranche-entry logic is the only value-add.

Recommendation for paper deployment

Same recommendation as the original report: archive the rules-based exit logic and paper-deploy a tranche-entry-then-hold variant.

Specifically:

  1. Drop ALL mechanical exits. Both B and C lose to A. Even with the real anchor, no mechanical exit rule we tested adds value over the 2021-2025 horizon.
  2. Keep the tranche-accumulation logic from the v1 framing doc. The buy-the-dip entries at -5%/-10%/-15% remain value-additive (per original backtest's trade logs).
  3. Re-cast the "phase-marker" concept as a manual review trigger, not a mechanical exit. When DRAM spot trend goes negative for 2+ quarters AND smart-money managers are reducing memory exposure AND hyperscaler capex guides are being cut — that's a FOUNDER-REVIEW alarm, not an autonomous-exit signal. The mechanical version of this alarm consistently sells the bottom.
  4. Use the real anchors for entry timing, not exit timing. A new long can require "DRAM trend ≤ 0 for ≥2 quarters" as a confirmation that we're near a cycle trough (i.e. buy when the cycle is bottoming). Then hold through the next up-leg without a mechanical exit.
  5. 2026 Q1 smart-money signal is a corroborating data point — Druckenmiller + Tepper + Tiger Global all opened SANDISK + INTEL positions in the same quarter, which is the closest thing to a "smart money confirms memory cycle is structurally positioned" signal we get without insider trading. Reinforces the THESIS validity even as it dings the STRATEGY validity.

Paper-deploy spec for v2: tranche-accumulate on a fresh thesis-confirmed long, hold to manual founder review. No mechanical exit. Re-evaluate the spec after 12 months of paper trading + real anchor data accumulation.

Caveats

  1. TrendForce data starts Feb 2021 — we miss the 2018-2020 cycle entirely. The 2018 DRAM crash and 2020 COVID rebound are precisely the kind of conditions where a real anchor might add the most value. The 2021-2025 window contains only ONE full cycle (peak ~2021Q4, trough ~2022Q4, recovery 2023-2025). One cycle is statistically thin; conclusions about strategy edge from a 1-cycle sample are weak.
  2. Only 1 of 3 real anchors actually wired in. Hyperscaler capex direction and HBM cadence are still stubs. The original thesis specified 2-of-3 confirmation; this test uses 1-of-1. A real 2-of-3 implementation might add filtering value that we cannot test today.
  3. 2026 Q1 smart-money signal is too fresh to weight heavily. SANDISK and INTEL adds appear in 13F filings that are <60 days old. Real position-sizing intentions vs trader-speculation is unknown until next quarter's filings reveal whether positions were added or trimmed.
  4. No transaction costs or slippage modeled. Variant C has 31 trades (vs 0 for buy-hold); at ~5bps slippage per side, C's gross-to-net erosion is ~3pp, which would push C's underperformance vs A from -19pp to ~-22pp. Material but doesn't change the ranking.
  5. DRAM spot direction_numeric is a coarse aggregator. TrendForce news items are not all of equal weight or scope (DDR4 retail vs DDR5 contract vs HBM3e have very different cycle dynamics). We're collapsing all of them into a single +1/-1 signal averaged per quarter. A more granular implementation (DRAM contract vs spot, DDR vs HBM, etc.) might surface different trends.
  6. Variant C's "OR" re-entry is structurally a bug. With AND logic the re-entry would be much stickier and the variant would behave like B. The OR formulation is the only reason C only loses 19pp; in a more rigorous formulation, C and B converge.
  7. Equal-weight portfolio understates name-selection skill. A real deployment would size MU and SMH differently based on conviction (MU is single-name memory exposure; SMH is broader semis ETF with memory weight as a fraction). Not modeled.
  8. No 2-of-3 confirmation logic from the original thesis. The spec says "single phase marker flips bearish but other 2 hold → trim to 1R." We can't test that without real capex + HBM data. The current Variant B is effectively 1-of-1 (DRAM-only).

Files

Related