A memory-chip index has to decide how much of Samsung Electronics belongs in it. Samsung is a memory maker, but it also sells phones, displays and appliances. Three common answers give very different weights:
Card
Rule
Samsung in the index
A
Full market cap (index-provider style)
the whole company
B
Pure-play by sales: market cap × memory-segment share of sales
about two fifths of the company
C
Sum-of-parts by operating profit: market cap × memory-segment share of operating profit
almost all of the company in the latest quarter
The other members (SKHX, MU, SNDK) are pure memory companies and are weighted by full market cap under every card. Every card caps one member at 50% and rebalances weekly.
What "optimal" means (pre-stated, before any result)
The selection criteria are fixed now so that no one can pick a winner after seeing the returns:
Exposure to the memory cycle. This is measured as correlation with the memory-only members' equal-weight basket.
Cost. This is turnover plus funding, measured at real costs.
Drawdown. This is the maximum peak-to-trough fall.
Return is reported but not used to choose. The backtest window, from 2026-02-20 (when SNDK started trading) to the latest quarter-end, covers only three quarters. That is too short to rank by return, so the three cards also run forward as paper books and are judged out-of-sample.
Selection criteria
memory-cycle exposure (correlation with the equal-weight basket of the memory-only members)
Samsung market cap = common + preferred shares valued at their own quarter-end prices; no segment adjustment
Other members
full market cap (SKHX, MU, SNDK are pure memory companies)
Cap per member
50%
Leverage
1.5x
Rebalance
weekly
Facts as of
latest common quarter-end of all four members (2Q 2026 at registration)
Interpretation notes
Index-provider style. Treats Samsung as a whole company. Segment share is clipped to 0..1; a quarter with negative total operating profit falls back to card A for that quarter (recorded, not hidden).
Samsung market cap (as in A) × memory-segment share of quarterly sales (about two fifths at registration)
Other members
full market cap (SKHX, MU, SNDK are pure memory companies)
Cap per member
50%
Leverage
1.5x
Rebalance
weekly
Facts as of
latest common quarter-end of all four members (2Q 2026 at registration)
Interpretation notes
Counts only the part of Samsung that sells memory, measured by revenue. Segment share is clipped to 0..1; a quarter with negative total operating profit falls back to card A for that quarter (recorded, not hidden).
Samsung market cap = (common + preferred × quarter-end price ratio) × memory-segment share of quarterly operating profit (almost all at registration)
Other members
full market cap (SKHX, MU, SNDK are pure memory companies)
Cap per member
50%
Leverage
1.5x
Rebalance
weekly
Facts as of
latest common quarter-end of all four members (2Q 2026 at registration)
Interpretation notes
Counts the part of Samsung that earns the profit; this is the rule the Forge pilot uses today. Segment share is clipped to 0..1; a quarter with negative total operating profit falls back to card A for that quarter (recorded, not hidden).
Registered
2026-10-06T00:00:00Z
Card hash
sha256:602d9c7e…6bf316
Evidence
backtestAs of 2026-09-30Funding: applied
Window 2026-02-20 .. 2026-09-30. Return is shown but is not the ranking key.
Card A
Card B
Card C
1. Memory-cycle correlation
0.92
invested days 0.96
0.94
invested days 0.98
0.93
invested days 0.97
2. Cost
161% turnover
funding net -$3,831
fees $56
167% turnover
funding net -$3,984
fees $62
162% turnover
funding net -$3,864
fees $57
3. Max drawdown
-60.6%
2026-06-22 → 2026-07-29
-62.3%
2026-06-22 → 2026-07-29
-61.1%
2026-06-22 → 2026-07-29
Total return (not the ranking key)
+79.2%
+95.1%
+83.5%
NAV race
Paper NAV from $10,000, daily. Shaded area under each line is its drawdown from the running peak. Ticks on the axis are weekly rebalances; flags mark the day a new fact set applies. Hover or touch to read a date.
A Full market capB Pure-play by salesC Sum-of-parts by operating profitMemory-only basket (SKHX, MU, SNDK)
Weight ribbon
Each member's weight at every weekly rebalance, held until the next one. Hover or touch to see the fact quarter behind a weight.
Card AFull market capSMSNSKHXMUSNDK
Card BPure-play by salesSMSNSKHXMUSNDK
Card CSum-of-parts by operating profitSMSNSKHXMUSNDK
The rule cards are hashed and never edited. Every deviation between a card as registered and the way a run actually executed it is recorded here, with its date, so readers can judge the evidence.
2026-10-06: backtest run (Forge, window 2026-02-20 to 2026-09-30)
Deviation 1. Preferred-share price ratio definition (cards A and C).
The cards say Samsung's preferred shares are valued with the "quarter-end price ratio". The engine used the last-month average price ratio from the business-report price table. The two measure the same thing in spirit, but the definitions differ. The results therefore test a near variant of A and C as registered. A re-run with the exact quarter-end ratio is needed before replication can mark A or C "replicated" on this point.
Deviation 2. 4Q 2025 segment figures are derived.
Samsung's 4Q 2025 memory-segment sales and operating profit were computed as full year minus nine months (2025 business report), not taken from a printed three-month figure. This affects cards B and C for the rebalances that used the 4Q 2025 fact set (22 of them).
As-known-at rule used.
A fact counts from the day after its publication date (the filing date). The 4Q 2025 set applied from 2026-03-18, 1Q 2026 from 2026-08-15 (the preferred ratio appears only in the half-year report), and 2Q 2026 from 2026-08-18.
Run facts.
No fallback, clip, 50% cap hit or safety trim occurred.
The deficit rule (card C falls back to A in a negative-profit quarter) was available but never triggered.
Variants tried before these three: 0.
Costs: fees, a 10 bp builder fee, a fixed 1 bp per side slippage assumption, and funding applied.
Authors
author pending
How to contribute
Propose a variant card with its own rule and its own hash.
Verify a fact against the filing it cites.
Replicate the backtest and report whether you get the same numbers.
The thesis and its cards are markdown files in the repository under docs/thesisbook/samsung-memory-index-weighting/. Each card is hashed: card_hash is the sha256 of the canonical JSON of its fields except card_hash.