Methodology

Article 1 — Source tiers and as-of dates

1. Every data point carries a source tier: T1 primary analyst or industry press with its own data collection (TrendForce/DRAMeXchange, earnings calls, company newsrooms), T2 major press with secondary reporting, T3 derived or aggregated (news wires, aggregator boards).
2. Every data point carries an as-of date and a confidence label (high/medium/low). Missing as-of dates are marked as such; no date is invented.
3. Conflicting sources stay side by side. We do not average and do not pick after the fact; a conflict is a finding, not noise.

Article 2 — Forecast pipeline (Q+1 to Q+3)

1. Forward series run through a Holt damped exponential smoothing pipeline (damping factor φ = 0.8; smoothing parameter α selected per series by backtest on enforcement of the 80% band).
2. Bands are built constructively from the backtest residuals: 80% and 95% bands, symmetric on the QoQ scale. Bands are not adjusted to outcomes after publication.
3. Where a primary analyst (TrendForce) publishes its own quarterly estimate, that is taken as the consensus point; the model runs alongside it, not through it.
4. Output is machine-readable: forecast/forecast-output.json, generated by forecast/forecast.py (public in source).

Article 3 — Horizon: no levels beyond Q+3

1. Beyond Q+3 this editorial team publishes no numeric levels. The reason is measured, not principled: WSTS revised its own 2026 growth from $975 bn (Nov 2025 edition) to $1.51 tn (Jun 2026 edition) — a more-than-doubling revision within seven months.
2. For 2027–2029 we publish named, anchor-bound scenarios (Outlook): per year a direction (tight / easing / loose) forced by dated, capital-committed events. A scenario without an anchor per year is not shown.

Article 4 — Regime flags

1. A regime break is marked when the consensus itself reclassifies its outcome by more than 50% within two publication cycles, or when a primary source changes a structural premise (export restrictions, EOL announcements, capex-guidance changes).
2. A regime flag does not suppress existing data; it marks that the validity of existing bands has widened.

Article 5 — Publication cycle and revisions

1. The edition appears monthly (next: 2026-10-10) and is revised in between on a material event (earnings, contract rounds, cleanroom openings).
2. Revisions are additive: a revised value does not replace the previous one in the public series but gets a revision step. TrendForce’s revision ladder (+90–98% → +58–63% → +13–18% → +8–13%) is visible on these pages as an example of the method.

Article 6 — Revision table of this document