CommodityWatch / Methodology
This is the long-form version of the methodology notes that appear inline throughout the product. If a metric is not obvious from its card-level explanation, the fuller treatment is here.
01
CommodityWatch prioritises public, official, or institutionally disclosed sources. Source classifications are operational product metadata, not legal advice. In practice that means four source families.
Deliberate exclusions matter just as much as inclusions. CommodityWatch does not mirror article bodies or reproduce paid assessment values. That leaves known coverage gaps in some specialty markets, but the trade-off is intentional: every number on the site should carry a visible source and caveat trail.
02
Ingestion runs on schedules calibrated to each source cadence. Weekly series such as the EIA Weekly Petroleum Status Report and EIA natural gas storage are refreshed around their release windows. Monthly series such as WASDE, FRED macro indicators, and related institutional releases are pulled on their own schedules. Daily series such as benchmark prices and AGSI+ storage are refreshed on a fixed cycle.
Where the underlying source publishes revisions, CommodityWatch preserves release timestamps, vintage metadata, and revision markers instead of silently overwriting the prior observation. Selected derived calculations also store deterministic hashes of their inputs so changes to the numerator, denominator, or source vintage can be traced.
DemandWatch is served through the root app's DemandWatch API path, with route JavaScript owning presentation mapping.
03
Absolute values are often misleading in isolation. A crude inventory reading can look comfortable in one month and tight in another. CommodityWatch therefore pairs many inventory and demand series with seasonal context drawn from the same week of year over the prior five years.
Seasonal bands are expressed through percentile ranges rather than a single average. The displayed bands use the 10th, 25th, 50th, 75th, and 90th percentiles, which makes it easier to identify whether the latest print sits inside the normal range, toward an edge, or outside it entirely.
04
Days of supply is one of the clearest examples of a metric that looks simple but depends on denominator discipline. CommodityWatch documents the convention used for each series instead of presenting a single generic formula.
CommodityWatch currently publishes computed days-of-supply for crude, gasoline, and distillate using EIA inventory and demand inputs. Other petroleum-product days-of-supply series are added only when the public store supports them.
Natural gas storage is currently shown as reported inventory and seasonal context. Natural gas days-of-supply is not currently published in the root app.
AGSI+ working gas is shown as reported storage at aggregate or country level. European gas days-of-supply is not currently published in the root app.
For grains, the standard metric is stocks-to-use ratio rather than days of supply. CommodityWatch keeps the naming distinct to avoid mixing weekly storage coverage with marketing-year balance-sheet math.
CommodityWatch does not force days-of-supply across markets where the denominator would be weak or misleading. Base metals, many softs, precious metals, and ETF-linked holdings only receive the metric when a defensible consumption basis exists.
05
Demand data is not homogeneous. Some indicators are direct measurements of physical offtake; others are proxies or contextual signals. CommodityWatch makes that distinction explicit through a seven-tier reliability taxonomy.
Measured physical offtake such as product supplied, electricity demand, or domestic use.
Processing volumes that track underlying demand closely, such as refinery inputs or crush data.
Import and export measures used where direct domestic demand coverage is weak or delayed.
Sector-level activity such as construction, auto output, or vehicle miles traveled.
Forward-looking measures such as new orders, permits, or similar pre-consumption signals.
Broad economic backdrop rather than commodity-specific demand measurement.
Heating, cooling, and precipitation-linked signals whose reliability varies by market and season.
06
HeadlineWatch standardises source attribution, timestamps, commodity mapping, and sentiment metadata where that metadata is available from the ingestion workflow. The enrichment layer combines deterministic taxonomy rules with model-assisted classification used during processing.
The current shipped product does not publish AI-written summaries or trading calls. PriceWatch TimesFM projections are the only shipped model-generated market-context overlay, and they are labelled as projections rather than CommodityWatch market views. If LLM-generated fields are added later, they will carry explicit disclosure rather than being presented as silent editorial judgement.
07
PriceWatch can display an optional TimesFM projection on the detail chart. The projection is generated only when the optional projection step runs after a PriceWatch data refresh. The actual published observations remain the solid line. The model-implied future path is shown as a dashed line only after the user requests it.
The pilot uses TimesFM 2.5 with checkpoint google/timesfm-2.5-200m-pytorch. Inputs are
published PriceWatch price histories regularized to the source cadence: daily series use a business-day
grid, weekly series keep weekly cadence, and monthly series keep monthly cadence. The system uses raw
price levels, model input normalization, positive-output enforcement, quantile crossing fixes, and
stores 80% interval bounds when quantiles are returned.
One 12M projection is stored per successful series run and sliced into the supported 1M, 3M, 6M, and 12M presets. Custom horizons are not supported in this pilot. Successful runs are preserved as audit history, but the UI exposes only the latest successful run.
TimesFM projections are model-implied market context from published PriceWatch history. They are not investment advice, trading recommendations, or guarantees of future prices.
08
Source families currently used in the product or supported by the repo-visible ingestion workflows include the following.
EIA, DOE, FERC, AGSI+, Ember, FRED, and related official or public reference sources.
USGS, World Bank reference series, FRED, IMF, and selected exchange warehouse disclosures where source terms support public display.
USDA WASDE, PSD Online, NASS, FAS GAIN, CBOT, ICE, and other official agricultural releases.
FRED, Eurostat, and other public economic releases that help frame cross-market demand conditions.
CPC, NOAA, NWS, and the publication calendars of the agencies and institutions being tracked.
Each module still carries its own local attribution and metric notes; this page is the shared policy layer.
09
TimesFM projections are labelled model-implied context overlays, not CommodityWatch market views.
Indicators are surfaced as context for interpretation, not buy-or-sell instructions.
Paid vendor datasets and inferred reproductions of them sit outside scope by design.
If a source is paywalled, CommodityWatch can link to it but does not reproduce its body text.
10
Material changes to formulas, source coverage, taxonomy rules, or page-level disclosure will be logged in the public methodology changelog.
Methodology is treated as a versioned asset, not a launch-day checkbox. If a convention changes, the change should be visible.