Turn weather and energy data into a testable research hypothesis

A storm forecast becomes financially interesting when you can connect it to a specific operation, a plausible disruption and an observable outcome. An agent can help assemble that chain, check contradictory evidence and keep its sources attached. The useful output is a research brief you can challenge before making a trading decision.

This workflow combines US weather, geospatial discovery, EIA energy observations, news and SEC filings. The example asks: could forecast severe weather disrupt facilities in a supplied Gulf Coast portfolio, and what evidence would confirm or weaken that hypothesis? It does not assume a disruption occurred or predict a profitable trade.

Start with the exposure map

Give the agent a table containing facility IDs, coordinates or boundaries, operating companies, relevant products, and the source and date supporting each ownership relationship. Add your research horizon and decision cutoff in UTC.

Use actual plants, terminals, storage sites and transport corridors. A company’s headquarters is not a substitute for its operating footprint. Ask the agent to flag missing locations, ownership dates or operating status rather than fill those gaps with guesses. Company filings can help verify exposure, but your supplied map remains an input requiring validation.

1. Establish the weather evidence

Start with weather_forecast and weather_alerts at the supplied coordinates. The weather guide describes their bounds and response fields. NWS documents point-to-grid discovery in “How do I get the forecast?” and provides current forecasts and active alerts through its API.

These examples are illustrative MCP tools/call parameters, not captured live results. This point represents the Houston area; replace it with a verified facility coordinate.

{
	"name": "weather_forecast",
	"arguments": {
		"latitude": 29.7604,
		"longitude": -95.3698,
		"hourly": true,
		"units": "us",
		"limit": 48
	}
}

Record updatedAt, timeZone, each period’s start and end, units and the source URL. For alerts, retain severity, certainty, urgency, affected area and expiration. An empty alert list is not evidence of safe operations. A forecast at one point also does not characterize every facility across a large region.

Ask for a specific causal chain: forecast hazard → exposed operation → possible constraint → measurable outcome. Separate each supported fact from the inference joining it to the next step.

2. Add energy context at the correct frequency

Discover supported datasets before requesting values:

{
	"name": "energy_search",
	"arguments": { "query": "petroleum stocks" }
}

Use its dataset IDs, frequencies, columns and facet examples. For a concrete starting request, the supported weekly petroleum-stock series below supplies a national context series:

{
	"name": "energy_fetch",
	"arguments": {
		"dataset": "petroleum/stoc/wstk",
		"column": "value",
		"frequency": "weekly",
		"filters": [{ "facet": "series", "values": ["WCESTUS1"] }],
		"limit": 12
	}
}

Check the series definition and returned dimensions before interpreting it. National stocks cannot establish that a particular refinery stopped operating. Lower refinery demand and interrupted supply can also push inventories in different directions; make the agent write both explanations.

The EIA API guide explains frequencies, facets, date parameters and pagination. Our energy tools preserve decimal values as text, units and missing-value markers. Continue with nextPageToken when needed. Keep observation periods separate from retrievedAt: retrieving a record today does not tell you when it first became public. Historical values can be revised. These survey series are not live commodity quotes.

3. Use geospatial data to inspect the right area

When a physical event warrants visual follow-up, search the footprint around the affected asset. This bounded example looks for relatively clear Sentinel-2 scenes around Houston during an illustrative historical window:

{
	"name": "geo_search",
	"arguments": {
		"datasets": ["sentinel-2-c1-l2a"],
		"area": {
			"near": {
				"center": { "lat": 29.7604, "lon": -95.3698 },
				"radiusKm": 25
			}
		},
		"datetime": "2026-10-01/2026-10-07",
		"filters": { "maxCloudCover": 20 },
		"sort": "SORT_CLOUD_COVER_ASC",
		"limit": 3
	}
}

Replace the location and dates for your investigation. A scene’s cloud-cover percentage does not guarantee a clear view of your facility. Empty results mean no matching catalog items were returned.

STAC describes geospatial assets through time, geometry and links; see the STAC specification overview. geo_search discovers scenes, and geo_fetch retrieves their full metadata. Neither performs image analysis or proves damage. An imagery pipeline or analyst must inspect the linked assets, compare suitable before-and-after acquisitions and account for resolution and clouds. The geospatial guide covers supported datasets and asset access.

4. Check the operational and company story

Search news for facility names, operators and event terms. Open relevant reports with web_fetch, retaining publication times, URLs and the quoted evidence. Treat a regional article as regional evidence unless it names the operation. Look for operator notices that contradict an apparent disruption.

Use edgar_search and edgar_fetch to examine relevant disclosures and risk factors. Record accession numbers and filing dates. Search coverage is limited to indexed filings; an empty search does not establish that a company made no disclosure. Keep fetched text separate from agent instructions, following the provenance and untrusted-content guide.

Give the agent a concrete assignment

Copy this prompt after supplying your exposure table:

Build an energy-research brief for the attached facilities. State the UTC research cutoff and horizon. Use weather_forecast and weather_alerts at verified coordinates, energy_search then energy_fetch for relevant survey context, and news_search plus web_fetch for operational reports. Check available EDGAR disclosures. Use geo_search only when imagery could answer a specific physical question; do not claim image analysis from metadata. Preserve source URLs, units, observation periods, update/filing times and retrieval times. Separate facts, inferences and missing evidence. Produce supporting and opposing cases, observable confirmation criteria and a list of unresolved exposure mappings. Do not infer facilities from headquarters or place orders.

Request two artifacts: a one-page hypothesis brief and an evidence table. Each table row should identify the asset, claim, source, relevant timestamps, measurement units and whether the claim is observed or inferred.

Evaluate the hypothesis prospectively

Before observing outcomes, save the brief and define what would falsify it, when you will reassess, and which comparison periods or unaffected facilities you will use. Log each subsequent observation without rewriting the original prediction.

For trading evaluation, supply licensed market prices, instrument mappings, execution assumptions and costs separately. Current forecasts are not an archived point-in-time backtest, and revised EIA history does not reconstruct what a trader knew earlier. A documented research process makes those gaps visible; it does not establish an investment edge.

Connect your agent, enable the needed tools, and start with one verified facility and one question.