UMPI-KR DRAM Methodology#

Status: approved, version 1.0.0, effective 22 September 2026. This document defines the Urdais Memory Price Index V1 as a family of two monthly official-data DRAM indexes built entirely from Korean government statistics.

The two series are published independently and are never combined. There is no UMPI headline number and no composite between them: one is a quality-adjusted price index produced by a statistical agency, the other a unit-value index Urdais computes from customs value and weight, and a blend of the two would be a number with no referent.

The demo memory values still visible in the product are not published under this or any methodology version, and the production read path cannot serve them.

This document follows the principles of the Urdais methodology framework.

This supersedes the launch role of UMPI-DRAM Spot, which is deferred rather than withdrawn. That document defined six DRAM chip instruments in USD per chip from a commercial spot board. It remains a complete and reviewable specification, and it remains unpublishable: the source that carries those six instruments does not permit public commercial display of its prices. Nothing here revives it, and nothing here deletes it. See Relationship to the deferred spot architecture.

What changed, in one sentence. Urdais stopped trying to reproduce a commercial spot board it cannot license, and started measuring DRAM prices with statistics it can obtain, store, recompute, retain and publish on its own authority.

Output Identity#

There is no composite UMPI headline, and this version does not create one. The two series measure different economic objects by different methods, and a weighted blend of them would be a number with no referent. They are never averaged, never summed, never chained, and neither is ever used to fill a gap in the other.

The two series are also not the six Phase 2A chip instruments under new names. They do not inherit those identifiers, and those identifiers are not reused here.

Output Definition#

Series A — UMPI-KR DRAM PPI#

UMPI-KR DRAM PPI is the Bank of Korea's official Producer Price Index for DRAM, published by Urdais as a cited monthly series with a Urdais-calculated month-over-month change.

It measures the price of DRAM at the domestic producer shipment stage in Korea, as surveyed and quality-adjusted by the Bank of Korea under its own statistical methodology. Urdais does not compute the index level; the Bank of Korea does.

What it is not, stated because each of these is a plausible misreading:

• Not a spot-market price, and not a transaction tape.
• Not DDR5-specific, DDR4-specific or DDR3-specific. BOK surveys one DRAM commodity; no generation, density, organization, speed bin or grade split exists (verified — see Source verification).
• Not branded versus eTT. That distinction is a commercial spot-board grade and appears in no official statistic.
• Not a module price, not a wafer price, and not a price per chip in any currency.
• Not an export price. Producer-stage and export-stage are different measurements, and BOK publishes both; V1 takes the producer one.

Series B — UMPI-KR DRAM Export UV#

UMPI-KR DRAM Export UV is a monthly unit-value index Urdais calculates from Korea Customs official export statistics for DRAM chips under HSK 8542321010, as declared export value in USD divided by declared export weight in kilograms, rebased to an index.

Urdais computes this one. The inputs are official; the index is Urdais's derivation, and is labelled as such wherever it appears.

A unit value is not a price. Its movement combines genuine price change with change in what was exported — generation mix, density mix, vendor mix, product mix inside the code, and specification change. It is a trade unit-value index, and it is never described as a DRAM price, a DDR5 price, a chip price, a quality-adjusted price, or a market close.

This warning is not a caveat appended for safety. It is the definition: the series is an honest measurement of export unit value, and it would be a dishonest measurement of price.

The PPI / UV distinction, frozen#

The two series answer different questions and are constructed by different parties under different rules. This version freezes the separation.

They are never averaged, blended, chained, spliced, or used to impute one another. A month missing from one is not filled from the other. A reader is always told which one they are looking at.

Canonical Series#

Series A — UMPI-KR DRAM PPI#

Series B — UMPI-KR DRAM Export UV#

The published Series B value is an index. The underlying USD/kg is retained internally for reproducibility and may be exposed as metadata. It is never presented as a price per chip, and the two must never appear in a way that invites the reader to treat kilograms as chips.

Codes deliberately excluded#

Because MCP sits in a sibling code rather than inside the DRAM chip code, Series B is not knowingly contaminated by HBM-class multichip packages. That is a structural property of the classification, not a claim about every declaration made under it.

Source verification#

Every identifier above was confirmed live against the official systems on 22 September 2026, not inherited from a community wrapper. The prior research mapping was correct in one place and incomplete in another, and both are recorded.

Two corrections to the prior research#

1. A DRAM export price index does exist, and the prior mapping missed it. Earlier research recorded the export-price item as unconfirmed, and a community wrapper mapped only the semiconductor aggregate 30911AA on table 402Y014. Verification found that 402Y014 (수출물가지수 기본분류) indeed has no DRAM item — its finest semiconductor split is 개별소자 and 집적회로 — but that the by-commodity export table 402Y016 carries 30911201AA DRAM, 2020=100, monthly, range 197101–202608, basket weight 103.4. It is real, it is long, and it is not in V1; see Deferred and future candidates.
2. An item code alone does not identify a series. 30911201AA means DRAM in both 404Y016 (producer prices) and 402Y016 (export prices), and the two tables return different values for the same month. A series is identified by the triple (stat_code, item_code, cycle), plus any group dimension the table carries. An implementation that keys on the item code alone will silently publish the wrong statistic.

Methodology#

Series A — nothing is recomputed#

The published value for a reference month is the value BOK publishes for that month, in BOK's own base, unchanged. Urdais applies no arithmetic to the level, no rebasing, no smoothing, no seasonal adjustment and no currency conversion. The only Urdais calculation on Series A is the MoM change.

Urdais publishes BOK's number as a cited BOK series. It does not present the level as a Urdais construction, and it does not claim ownership of the source fact.

Series B — the unit value and its rebasing#

For a reference month t, from the official export figures for HSK 8542321010:

uv_usd_per_kg_t = expDlr_t / expWgt_t

Both terms are export-side and declared on the same basis. Import figures are never used: Korean import values are reported CIF while exports are FOB declared, so a ratio mixing them would measure freight and insurance as well as price.

The published index rebases that quantity to a fixed historical base:

UV_index_t = 100 × uv_usd_per_kg_t / uv_usd_per_kg_base

where

uv_usd_per_kg_base = ( Σ expDlr_m for the 12 months of 2020 ) / ( Σ expWgt_m for the 12 months of 2020 )

The base is the calendar-year 2020 aggregate unit value: total 2020 export USD divided by total 2020 export kilograms. Base = 100.

Why this base, and not the alternatives. Urdais has no general base-100 index convention to inherit — UTVI deliberately publishes a level rather than a rebased index, and UBWI publishes a percentage — so the choice is made here and justified rather than assumed.

• 2020 matches the base period of the Bank of Korea series beside it. The two series then sit on comparable scales, which makes the chart honest about their shapes without implying they measure the same thing. Their bases agree; their methods do not, and the surface says so.
• It is fixed and historical, so it never moves. A base of "the first live month" would silently rebase the whole history every time the product relaunched, and would make a series with thirty years of available history appear to begin at launch.
• An annual aggregate, not a single month. A single base month would import that month's mix and noise into every value in the series forever. Summing value and weight across twelve months before dividing gives a weight-weighted annual unit value rather than an average of twelve ratios, which is the arithmetic that matches what a unit value is.
• It is reproducible by a third party from the same public API without any Urdais input.

The base is frozen by this methodology version. Changing it is a new version, it is disclosed, and it does not silently restate published history.

A BOK rebase does not rebase Series B. If the Bank of Korea moves its PPI to a later base, Series A's unit changes because BOK's unit changed — that is a source event, recorded as such — while Series B stays on its own frozen 2020 base until a Urdais methodology version says otherwise. The two bases may then differ, and the surface must state each series' base rather than implying a shared one.

Change calculation, both series#

The canonical change is MoM: the value for a reference month against the value for the immediately preceding reference month, within the same series, the same source lineage, the same methodology version and the same base regime.

MoM_t = (value_t − value_{t−1}) / value_{t−1}

No other change is canonical in V1. YoY may be derived later under a version that defines it; it is not defined here.

Cadence#

Both series are monthly. There is no intraday value, no daily value, no weekly value, and no interpolation of any kind.

• The observation identity is the reference month, not the retrieval date.
• BOK PPI is released roughly three weeks after month end, preliminary and then revised.
• Korea Customs figures update around the 15th of the following month. V1 uses final monthly figures and does not publish 10-day or 20-day provisional trade counts as if they were the month.
• A month with no official figure has no point. It is not zero, not carried forward, not interpolated, and not estimated from the other series.
• The surface shows the latest reference month and the vintage it came from, never a synthetic "as of today".

Vintage and revisions#

Official statistics are revised, and a product that overwrites them loses the ability to explain what it published. This follows the settlement discipline already established for UTVI.

For every observation, Urdais stores:

Four rules:

1. Every retrieval is recorded append-only. Nothing is overwritten.
2. The published value for a reference month is the value from its most recent retrieval.
3. A later change is a supersession, never an edit. The prior value stays retrievable, and a previously published figure remains explainable.
4. Provisional and final are distinguished and labelled, so a reader can tell a first print from a settled one.

A revision changes the current published history under the revision rules above; it does not retroactively rewrite what Urdais stated at an earlier time, and the lineage to explain the earlier statement is retained.

Historical backfill#

Backfill is permitted and bounded, per series. No interpolation, in history or in the present.

Series A. Backfill from the earliest official DRAM observation available under the currently linked BOK series — 199501 at verification. BOK publishes linked indexes across its base regimes; Urdais preserves that official linking and does not splice incompatible base regimes on its own initiative. If BOK does not supply a link across a regime boundary, the boundary becomes a published methodology boundary rather than a silent join.

Series B. Backfill only where all hold: HSK 8542321010 is definitionally comparable over the span, both expDlr and expWgt exist for the month, and any classification change is documented. The base requires the twelve months of 2020, so the series cannot be published before that base is computable.

If the HSK classification changes materially, a methodology boundary is created. Unlike classifications are never silently spliced into one series.

Mix shift is not a defect to be corrected out of Series B. Over a long span the migration from DDR3 to DDR4 to DDR5 and the rise of higher densities is part of what the series records, and the methodology says so rather than pretending to a purity it does not have.

Rights and attribution#

Both sources are official public data, obtained and used under published government terms. Urdais owns its methodology, its calculations and its normalization. It does not own the source facts, and it does not claim to.

Bank of Korea#

Two obligations follow and are binding on the surface: every display of Series A names the Bank of Korea, and the MoM change is disclosed as a Urdais calculation rather than presented as a BOK figure.

Korea Customs Service#

A further constraint applies to both, and is a rule about product shape rather than about citation: Urdais publishes a processed, documented index product. It does not resell an unmodified official table as a paid replica, and it does not present official figures in a way that distorts them.

Prohibited sources#

The following are barred from production ingestion for V1, and none may be added without an explicit new methodology and source-rights version. Several are barred for reasons that survive any amount of technical convenience.

A restriction on a figure attaches to the figure, not to the page it is read from. Ingesting a government footnote, a news article or an aggregator that reprints a commercial survey does not cure the survey's terms.

Ingestion Contract#

This section specifies the contract every retrieval obeys. It is not itself an authorization to collect: a source may be collected from only while its own rights gate holds, and approval of this methodology does not grant, widen or cure any source's terms.

Stored Outputs#

For as long as any published value depends on them: every admitted observation with its vintage and provenance; the computed uv_usd_per_kg and the base aggregate for Series B; every published value with its methodology version and reference month; every rejection with its reason; and the attribution metadata each source requires.

Validation / Quality Rules#

Published Surface#

Recorded as requirements for a later phase. No frontend change is made by this document, and the current demo surface is untouched.

Relationship to the deferred spot architecture#

UMPI-DRAM Spot is deferred, not deleted and not superseded in substance. Its six instruments, its USD-per-chip unit, its branded/eTT separation and its 1D close-to-close change remain a correct specification of a different product that Urdais cannot currently license.

• It is not UMPI V1, not production-ready, and not the default public methodology for UMPI.
• Its six instrument identifiers stay attached to it and are not reused by the series defined here.
• It becomes relevant again only if Urdais obtains written rights permitting public commercial display of a spot board's prices, from any source.
• Nothing in this document weakens its gates, and nothing in it revives them.

The two are different economic objects. A monthly official index is not a slower version of a daily spot price, and this family does not pretend to be one.

Deferred and future candidates#

Recorded so that a later phase does not rediscover them, and so that none is mistaken for V1 content.

Lineage and Reproducibility#

Published value → methodology version → stored observation with its vintage → agency response with its identifiers → official source.

Both series are independently reproducible by a third party: the identifiers, formula, base and rules in this document are sufficient to recompute every published figure from the same public APIs. That property is the reason this architecture exists, and it is the substantive gain over the deferred spot product.

Publication and Revision Behavior#

• Publication requires an approved methodology version and a working implementation whose values reproduce. Neither substitutes for the other.
• Production is fail-closed. Where requirements are unmet, no value is published and no demo, placeholder or estimated value is shown in its place.
• Attribution is published with the value, in the form each agency's terms require.
• Revisions supersede and are recorded; they never silently overwrite.

Methodology Version#

1.0.0. Status approved, effective 22 September 2026.

Approval permits publication of the two series defined here, and nothing else. It does not approve the deferred six-instrument spot architecture, DDR5/DDR4/eTT chip quotes, USD per chip, HBM prices, a composite between the two series, the Bank of Korea export price index, or any proprietary vendor data. Each of those remains outside V1 and would require its own version.

Approval changes the methodology's standing, not the sources' rights. The Bank of Korea determination stays ambiguous_requires_legal_review with its founder-accepted-risk marker: an approved methodology says Urdais may publish this measurement, not that a question about a source credential has been answered.

0.1.0-draft is retained as the superseded predecessor rather than deleted. Values calculated under it were internal-only and were never published; they are regenerated under 1.0.0 rather than relabelled, because the methodology version is part of a calculation's identity.

Why approval did not relabel the draft's calculations. A published value must be able to say which public methodology governs it. The draft's records carried the draft's identity, so promoting them by changing a column would have made the lineage assert something that was not true when the number was produced. They are superseded by deterministic recalculation under 1.0.0 from the same inputs, and the draft rows remain for audit.

Why a new lineage rather than a version bump of the spot methodology. This product changes the economic object, the sources, the unit, the cadence, the instrument universe, the change calculation and the publication semantics. Nothing of the spot specification survives into it except the family name. Bumping umpi.md to 0.2.0-draft would assert a continuity that does not exist and would destroy a specification that is still correct for its own product. The spot document therefore keeps its version and its content and is marked deferred; this document starts its own lineage at 0.1.0-draft.

An approved version is never edited in place. A change is made by superseding this version with a successor, so that published values stay attributable to the rules that produced them.

Version History#

1.0.0, 22 September 2026 — approved, effective 22 September 2026. First approved version, and the first that permits publication. The economic objects, units, cadence, change definition, rights gates and refusals are exactly those frozen at 0.1.0-draft: nothing about what is measured changed at approval, which is why the version step is a promotion rather than a redefinition. What it adds is standing — the two series may be published, independently and with their attribution, and Series B carries its trade-unit-value warning wherever it appears. Records calculated under the draft are superseded by recalculation under this version rather than relabelled. Supersedes 0.1.0-draft.

0.1.0-draft, 22 September 2026. First definition of the UMPI-KR DRAM family. Pivots UMPI V1 from six proprietary-source daily USD-per-chip spot instruments to two monthly official-data indexes, and defers rather than deletes the spot specification. Defines Series A as the Bank of Korea DRAM Producer Price Index published as a cited series, and Series B as a Urdais-calculated export unit-value index from Korea Customs value and weight under HSK 8542321010, rebased to the 2020 calendar-year aggregate = 100. Freezes the price-index versus unit-value distinction and forbids averaging, blending or mutual imputation; freezes monthly cadence, the MoM change and the ban on interpolation; freezes the append-only vintage and supersession rules and the per-series backfill boundaries; records the attribution and modification-disclosure obligations of both agencies; and bars every proprietary and commercial-lineage source from V1. Records the live verification of all identifiers on 22 September 2026, including the discovery that a DRAM export price index exists on 402Y016 and that an item code alone does not identify a BOK series. No source is ingested, no observation is created, and no production effective date is established.