Methodology · retail-v3-mainstream
How the DDR5 retail index works
A transparent, mainstream current-product basket is more useful than a large opaque one. These rules define what enters each weekly point, what stays out, and how the basket is kept relevant as products change.
1. Two capacity anchors
The index follows 32GB (2×16GB) and 64GB (2×32GB) desktop UDIMM kits. Both capacities are common enough to support multiple brands and large enough that tiny low-capacity products do not dominate the price-per-GB signal.
2. Product eligibility
| Rule | Included | Excluded |
|---|---|---|
| Brand universe | Kingston, Corsair, G.Skill, TEAMGROUP | Every other brand, regardless of apparent bargain price |
| Product family | One representative per memory family and capacity | RGB, colour, and closely related cosmetic variants counted as separate observations |
| Market evidence | Rating ≥4.2 and either ≥25 ratings or ≥50 bought in the past month | Missing/low ratings and thin demand evidence |
| Fulfilment evidence | Prime, meaningful recent purchase activity, or an explicit in-stock signal | A returned price with none of those availability signals |
| Stock depth | Available with no explicit thin-stock warning | Out of stock or “only 10 left” and below |
| Compatibility | Normal platform/profile qualifications | Explicit “not compatible,” discontinued, or end-of-life warnings |
| Generation | Explicit DDR5 | DDR4, DDR3, ambiguous titles |
| Form factor | Desktop UDIMM / 288-pin | Laptop SODIMM / 262-pin |
| Class | Mainstream unbuffered desktop kit | ECC, RDIMM, LRDIMM, server or workstation memory |
| Condition | New | Used, renewed, refurbished, open-box |
| Capacity | 32GB or 64GB total, two modules | Single modules, four-module kits, ambiguous totals |
| Price health | $0.50–$40.00 per GB broad parser guardrail | Values outside the guardrail |
The broad guardrail is designed to catch parsing errors and accessory/bundle mismatches, not suppress a real market move. Excluded observations keep a reason code in the private database.
3. Basket construction
The target is six active products: three 32GB kits and three 64GB kits. Ranking prioritizes recent purchase activity, rating count, star rating, Prime availability, and explicit technical identity—not the lowest price. Selection takes one product per brand before allowing a second product from the same brand.
Before ranking, titles are mapped to product families such as Kingston FURY Beast, G.Skill Flare X5, and G.Skill Trident Z5 Neo. Only one member of a family can represent a capacity. When both pass every gate, the non-RGB version is preferred over its RGB or colour variant.
An active product is removed immediately when a current search exposes a hard quality failure. Otherwise it is paused after six missed searches—about two weeks—and a qualified current replacement can enter the basket. Historical observations remain intact.
4. Weekly aggregation, mean, and median
Retail searches run Monday, Wednesday, and Friday. Each eligible product first becomes a comparable nominal-dollar value:
A product’s weekly value is the median of its valid checks during the Monday-through-Sunday UTC week. For n eligible active products, the published mean is the sum of those product-week USD/GB values divided by n. The published median is their middle sorted value—or the average of the two middle values when n is even. We publish both because the mean reacts more to a broad move while the median describes the center more robustly.
5. The rebased index
The first valid live observation is 100. On later weeks, the collector calculates current/previous price relatives for products present in both successful observations. The previous index is multiplied by the mean or median overlapping-product relative. If no products overlap, the previous index level is carried forward and the overlap count exposes the break.
Dollar-per-GB values remain the primary result. The rebased index is a continuity aid, not a way to conceal basket composition.
6. Data-health labels
Good means at least four eligible active products from at least three brands. Thin means a valid point exists with less breadth. Initializing means no valid point exists. A timestamp always identifies the latest successful run.
7. Collection and quality checks
Scheduled observations run Monday, Wednesday, and Friday. Each run uses a broad and a mainstream-brand search for both capacities. Four healthy requests per run refresh prices, demand evidence, and replacement candidates—about 52 requests per average month against the primary provider’s 100-request tier.
Large price changes, new basket members, and routine sample checks receive BrowserFetcher verification through rotating private sticky sessions. The verifier blocks expensive resources, records bytes, stops on challenges, and has a 75 MB monthly ceiling. Proxy results never turn an unknown offer into a guessed price.
8. One-year historical backcast
The default trend begins in August 2025 with four established desktop DDR5 kits: Corsair Vengeance and Crucial Pro at 32GB, plus Kingston FURY Beast and G.Skill Trident Z5 Neo RGB at 64GB. The reconstruction uses dated public listed-price histories stored as a fixed archive.
When multiple values exist on one date, the last published value for that date is retained while every same-day value remains in the audit note. A price is carried forward for at most 45 days. Each product-month must cover at least 15 days, except the partial August 2026 bridge month, and becomes the median of its applicable daily prices divided by kit capacity.
The monthly backcast uses the same overlapping-product price-relative calculation as the weekly live index. The historical series is rebased to 100 at the August 10, 2026 handoff; the live series starts at 100 there. A vertical marker identifies the change from the four-product historical basket to the six-product live MPC basket. This makes the trend comparable while keeping the basket change explicit.
The historical basket measures trend, not an exact dollar estimate for the live six-kit basket. The older Stanford DAM cheapest-retail series remains stored as external research context but is no longer the default chart line.
9. SSD retail index
The SSD chart is a separate series and never changes the DDR5 basket or index. Its scope is new consumer internal M.2 2280 NVMe PCIe 4.0 drives at nominal 1TB and 2TB capacities. SATA, external and portable drives, M.2 2230, PCIe 3.0, PCIe 5.0, enterprise hardware, used or renewed products, heatsink-only variants, bundles, and explicit stock of ten units or fewer are excluded.
The launch selector prioritizes matched Kingston NV3, Crucial P310, and WD_BLACK SN7100 families across both capacities. Eligible alternatives may replace a family after repeated availability failure, but a basket change never rewrites a historical point. Each listing must have a rating of at least 4.2 and either at least 25 reviews or at least 50 reported purchases in the past month, plus credible fulfilment evidence.
Monday, Wednesday, and Friday observations are collapsed to one product-week median. Separate 1TB and 2TB series and a combined series publish basket mean, median, sample count, brand count, and continuity index.
SSD historical reconstruction
The default trend begins in August 2025 using public PriceHistory.app histories for the same six exact launch products. Multiple values on one date retain the last published value while every same-day value remains recorded in the audit note. A listing can carry forward for at most 45 days; each product-month normally needs at least 15 covered days and becomes the median applicable listed price normalized per terabyte.
The monthly historical segment and weekly live segment use the same overlapping-product price-relative calculation. Both are rebased to 100 at the August 10, 2026 handoff, which is visibly marked on the chart. Dollar values on either side remain independently observed; the dashed bridge discloses the source/cadence change instead of treating the handoff as an ordinary price movement.
Advanced chart calculations
Advanced view combines the published historical and live median points, then calculates 5-, 20-, and 200-calendar-day simple moving averages. Between sparse monthly or weekly publications, the latest published value is carried forward; the pre-August 2025 lookback is initialized flat at the first published value. The volatility band is the 20-day simple moving average plus or minus two population standard deviations. These indicators are descriptive overlays on this retail series, not forecasts or trading signals.
This measures finished consumer SSD listings, not NAND flash spot, contract, wholesale, or transaction prices. Controllers, firmware, endurance, warranties, channel inventory, and retailer behavior can all affect retail SSD prices independently of raw NAND.
10. High-VRAM GPU retail index
The GPU tracker is a separate observed-retail series and does not reuse the broad Semiconductor PPI or alter DDR5 and SSD history. Its scope is new standalone GeForce RTX 5090 32GB, RTX 5080 16GB, and RTX 5070 Ti 16GB desktop add-in cards with explicit GDDR7 memory, observed in Amazon US listings.
One exact search per model runs weekly. Eligible listings must identify the supported GPU and correct VRAM capacity, come from ASUS, Gigabyte, MSI, PNY, Zotac, or NVIDIA, and have new condition. A displayed rating below 4.0 fails; otherwise the listing needs at least three ratings, five reported purchases in the past month, Prime, or an explicit in-stock signal. This allows a newly listed high-cost board with no displayed rating to qualify only when Amazon exposes credible fulfilment evidence. Laptops, prebuilt systems, workstation servers, water blocks, coolers, brackets, bundles, used or renewed cards, unsupported Super variants, and implausible prices are excluded. An explicit “only one” or “only two” stock signal stays in the observed set as thin_stock because scarcity is relevant to this series, but it receives a ranking penalty so deeper-stock boards are preferred.
The active target is three distinct board families per GPU model, with at least two required to publish. Selection prioritizes market evidence, rating, Prime or explicit in-stock evidence, and board-family diversity—not the lowest price. Hard eligibility failures pause immediately; ordinary search absence uses the same six-run grace policy as other retail baskets.
Each model publishes its weekly listed-price mean and median plus a chain-linked trend index using products present in adjacent successful weeks. The combined GPU index exists only as a normalized proportional overview; the page never presents one combined dollar price across the three tiers.
GPU historical reconstruction
The historical segment begins in March 2025 and uses fixed public PriceHistory.app Amazon US archives for eight exact ASINs: two RTX 5090 boards and three boards in each 16GB model. The third live RTX 5090 was omitted because the discovered archive belonged to a different Amazon market. Source values remain preserved in the immutable fixture and history table; four values outside the live collector’s reviewed model bounds are excluded from calculation rather than silently corrected.
Multiple values on one date retain the last published value while every same-day value remains in the audit note. A valid listed price can carry forward for at most 45 days. Each product-month needs at least 15 covered days, and each published model-month requires at least two exact cards. Product-month medians feed the same overlapping-product price-relative calculation used by the live model index.
The monthly historical segment is rebased to 100 at the August 24, 2026 live handoff. The weekly live segment starts independently at 100, and a vertical marker discloses the source, cadence, and basket transition. Historical and live dollar medians remain independently observed; the connecting bridge is a disclosure, not an ordinary weekly price movement.
GDDR7 and HBM are both DRAM products but are not interchangeable. HBM uses stacked dies and advanced packaging for data-center accelerators, while these cards use discrete GDDR7. AI demand may affect shared capital, wafer, engineering, packaging, and GPU-allocation constraints, but this retail series cannot prove that HBM demand caused a card-price change.
11. Known limitations
- Listed prices are not confirmed transactions.
- Sales tax is excluded; shipping is excluded unless a future source exposes it consistently.
- Retail availability and product selection can change between observations.
- Timing, speed, endurance, warranty, seller, cooling, power, and performance differences are not quality-adjusted.
- Each index represents its defined US retail basket, not every retailer or the entire worldwide market.
- Weekly data is not an intraday quote and does not predict the next move.
12. Revisions
gpu-backcast-v1-eight-card adds the March 2025–August 2026 reconstruction from eight exact Amazon US ASIN archives, deterministic outlier exclusions, and an explicit live-calculation handoff. retail-v3-mainstream-gpu1 adds separate weekly RTX 5090, RTX 5080, and RTX 5070 Ti Amazon US listing series without duplicating a broad macro proxy. ssd-backcast-v1-six-drive adds the August 2025–August 2026 SSD reconstruction from the six exact launch products and an explicit live-calculation handoff. retail-v3-mainstream-ssd1 adds the separate 1TB and 2TB M.2 NVMe PCIe 4.0 SSD index without altering DDR5 history. backcast-v1-four-kit adds the August 2025–August 2026 DDR5 historical reconstruction and explicit calculation-handoff marker. retail-v2-weekly established the DDR5 launch baseline on August 12, 2026 with three observations per week and weekly product medians. retail-v3-mainstream replaced that same baseline week’s DDR5 basket with a strict four-brand universe, private market-evidence and stock gates, compatibility exclusions, brand-first diversification, and product-family deduplication.
DRAM chip spot benchmarks
The DRAM spot-price chart publishes separately labelled DDR5, DDR4 and DDR3 chip benchmarks captured from the TrendForce / DRAMeXchange spot table. Session averages, lows and highs are nominal USD per chip. Each exact specification, including eTT, remains separate.
Collection began in August 2026 and attempts one daily capture. Source sessions use Taipei time (UTC+8). We preserve immutable revisions and display the latest stored revision for a chip and source timestamp. We neither fill missing dates nor reconstruct older history. Captured change equals the latest session average divided by the previous captured average, minus one, multiplied by 100; it is distinct from the source-reported session change. Spot benchmarks are separate from contract prices, assembled modules and retail kits.