Satellite Imagery Based Signal Research

When to Use

Use this skill when you already have a processed computer-vision metric out of an Earth-observation pipeline — a vehicle count for a retailer's lots, an estimated fill fraction for a cluster of external floating-roof crude tanks, an NDVI composite over a growing region — and you need to turn it into a research signal that a backtest can consume without cheating.

The attraction is real: these metrics observe physical activity directly, ahead of the official statistics the market trades on. Concretely, the two headline benchmarks publish on fixed, verifiable schedules:

• EIA Weekly Petroleum Status Report — released Wednesdays after 10:30 a.m. ET (summary, overview, and Tables 1–14 in CSV/XLS; the remaining PDF/HTML after 1:00 p.m. ET), reporting stocks as of the previous Friday. Holiday weeks slip a day.
• USDA NASS Crop Progress — released 4:00 p.m. ET on the first business day of each week, April 1 through November 30, with reporters responding as of Sunday. WASDE lands at 12:00 p.m. ET between the 8th and 12th of each month.

So a Monday tank read genuinely precedes the Wednesday print. That head start is the entire economic case for the dataset, and it is also precisely what evaporates the moment a backtest is sloppy about when the imagery-derived number actually became usable. This engine exists to make that boundary explicit rather than assumed.

The engine emits tradeable_from_iso (acquisition + vendor pipeline lag), rejects a baseline window that reaches past the observation being scored, and gates on the fraction of the footprint that was actually visible.

When NOT to Use

• Not a computer-vision pipeline. It does not read imagery, detect vehicles, segment tank shadows, or compute NDVI from bands. It consumes the scalar those steps produce. Everything about detection quality is upstream and out of scope.
• Not a forecast, and not aware of expectations. trading_signal_direction is a supply/demand sign convention: a high reading is bullish for the retailer, bearish for crude, bearish for the crop. Equities and futures trade against consensus, not against a metric's own 52-week history. A record parking count in a quarter where the street already modelled a record is not a bullish surprise. Pair with a consensus-relative construction — see credit-card-transaction-data-signal-construction for that pattern.
• Not a substitute for a point-in-time backtest harness. tradeable_from_iso is one timestamp on one observation. Enforcing it across a whole dataset, handling vendor restatements, and replaying as-delivered snapshots belong to backtesting-alt-data-strategies-with-realistic-availability-lag.
• Not a licensing, contract, or MNPI clearance. Whether you may trade on a dataset is a separate question from whether the arithmetic is sound. The SEC's first enforcement action against an alternative-data provider (In re App Annie Inc. and Bertrand Schmitt, Sept. 14, 2021, §10(b)/Rule 10b-5, $10M corporate penalty) turned on how the data was derived and represented, not on how a subscriber normalised it. Diligence the provenance — see alternative-data-vendor-due-diligence-checklist and insider-trading-controls-for-alternative-data-usage.
• Not a proxy for the official statistic. Shadow-based tank estimates see external floating-roof tanks only. EIA commercial crude stocks additionally include fixed-roof tanks, pipeline fill, and volumes in transit. The satellite series measures a subset and will diverge from the print it is supposed to front-run.
• Not usable on a single cloudy scene. Optical sensors return nothing through cloud. One obscured acquisition is not a low reading; it is an absent one. Set min_usable_pixel_fraction rather than treating a masked scene as data.
• Not a signal on its own for thin baselines. A Z-score against three prior observations is arithmetic, not normalisation.

Prerequisites

• A processed metric per observation: vehicle count, tank fill fraction, or NDVI. NDVI is (NIR − Red) / (NIR + Red) and is bounded to [−1, +1] (Landsat 8/9: (B5 − B4)/(B5 + B4); Landsat 4–7: (B4 − B3)/(B4 + B3)). Values outside that range indicate an upstream scaling error, not a strong signal.
• A point-in-time rolling baseline (baseline_historical_mean, baseline_historical_std) built only from observations knowable at the acquisition time. baseline_historical_std must be strictly positive.
• A timezone-aware ISO-8601 image acquisition timestamp. Naive timestamps are rejected — with a delivery lag added on top, an unzoned time is ambiguous by up to a day in exactly the direction that flatters a backtest.
• availability_lag_days taken from your vendor contract. The 2-day default is a placeholder, not a measured constant.
• Optionally usable_pixel_fraction (1.0 − cloud/shadow/snow/no-data). Sentinel-2 L2A ships a Scene Classification Layer flagging cloud shadow, low/medium/high-probability cloud, thin cirrus, and saturated or defective pixels — derive the fraction from it rather than eyeballing the scene.
• Optionally baseline_window_end_iso and baseline_observation_count, so the look-ahead check and the baseline-depth audit can actually run.

Workflow

1. Ingest the processed metric. One SatelliteObservation per asset per acquisition. Sensor cadence bounds how often this can update: Sentinel-2 revisits every 5 days with two satellites (2–3 days at mid-latitudes); Landsat 8/9 repeat every 16 days per satellite. A "daily" signal on a 5-day sensor is interpolation, and interpolation across a gap is a forecast you did not intend to make.
2. Declare the quality gate before looking at results. Set min_usable_pixel_fraction on the engine. It defaults to 0.0 (disabled) because no threshold is universally correct — it depends on footprint size and on how much masking the vendor already did. Choosing it after seeing which scenes it would drop is fitting the gate to the answer.
3. Normalise against the point-in-time baseline. The engine computes Z = (X − μ) / σ unrounded. If σ ≤ 0 it raises rather than substituting 1.0 — a degenerate baseline would otherwise rescale the raw deviation into a full-conviction signal (2,000 extra cars becoming "Z = 2000").
4. Let the look-ahead check run. Supply baseline_window_end_iso. A window ending after the acquisition time raises; a window ending exactly at it logs a warning, because the observation is then inside its own baseline and its Z-score is pulled toward zero.
5. Map the deviation to a direction. |Z| ≥ z_score_threshold takes a side, using the sign convention in HIGH_READING_DIRECTION; anything inside the band is 0.0. Thresholding uses the unrounded Z — a 1.4951 must not round into a trade. A signal type absent from that table raises rather than inheriting another domain's economics.
6. Stamp the availability boundary. tradeable_from_iso = acquisition + availability_lag_days, normalised to UTC. This is the value the backtest must honour — not timestamp_iso. Feeding the acquisition time to the backtest is the failure this whole skill is built to prevent.
7. Read confidence_pct as a rank, not a probability. It is min(|Z| / strength_saturation_z, 1) × 100, an uncalibrated monotone ordering. Never size a position from it.

Full procedure: see references/workflows.md. Standards reference: see references/standards.md. Printable pre-flight checklist: see assets/checklist.md.

Common Pitfalls

• Backtesting from the acquisition timestamp. The image exists at capture; the number does not. Detection, mosaicking, QA, and delivery sit in between. Use tradeable_from_iso, and take the lag from the contract rather than from this skill's default.
• A degenerate baseline silently becoming conviction. If σ is 0 (a constant or unpopulated baseline) and the code substitutes 1.0, the "Z-score" is just the raw deviation in raw units — always far past any threshold, always maximum strength, always wrong. This engine raises instead.
• Rounding the Z-score before thresholding it. Displaying Z to two decimals is fine. Deciding on the rounded value promotes every observation in [1.495, 1.5) into a full-conviction trade.
• Treating cloud as a low reading. A 90%-obscured scene is a missing observation. Scored as data it manufactures a spurious deviation, and because cloud is seasonally and geographically correlated, the resulting bias is systematic rather than random.
• Correcting floating-roof shadows the wrong way. The exterior shadow (on the tank's shaded side) scales with total tank height; the interior shadow scales with how far the roof has sunk below the rim. Both stretch and shrink together as solar elevation changes through the year. The published approach therefore takes the ratio — fill ≈ 1 − (interior shadow area / exterior shadow area) — which cancels the sun-angle dependence. Subtracting a seasonal adjustment from a bare shadow length leaves the confound in.
• Assuming NDVI maps linearly to yield. NDVI saturates over dense canopy: past closure, more biomass barely moves the index, which is why EVI exists. It is also phenology-dependent — the same 0.75 means different things at emergence and at grain fill — and cannot distinguish a large low-quality harvest from a small high-quality one.
• Comparing a satellite series against an official statistic with different coverage. Shadow methods see external floating-roof tanks; EIA counts far more than that. A persistent gap between the two is usually coverage, not alpha.
• Confusing panel growth with activity growth. A vendor adding lots, tanks, or fields mid-history raises the aggregate metric with no change in the underlying economy. Recompute the baseline whenever panel composition shifts.
• Trading the metric instead of the surprise. Retail equities respond to revenue versus consensus. Traffic versus its own 52-week mean is only half the comparison.
• Reading confidence_pct as a win probability. It saturates at an arbitrary |Z| = 3 by default. It ranks observations; it does not price them.

Verification

• Process a retail observation at Z = +2.0 and assert trading_signal_direction == +1.0; at Z = −2.5 assert −1.0.
• Process an oil-storage observation at Z = +2.5 (inventory build) and assert −1.0; at Z = −2.0 (draw) assert +1.0. Repeat both signs for NDVI and assert the crop-price convention (+Z → −1.0).
• Assert Z exactly at the threshold fires, and that observed_metric giving Z = 1.4951 stays neutral even though it displays as 1.50.
• Assert baseline_historical_std = 0.0 raises rather than returning a direction, and that a NaN or infinite metric raises rather than producing direction = 0.0 with a NaN strength.
• Assert a naive timestamp_iso raises, and that 2026-08-05T09:00:00-03:00 and 2026-08-05T12:00:00Z produce the identical tradeable_from_iso.
• Assert baseline_window_end_iso after timestamp_iso raises with a look-ahead message, and that a window ending exactly at acquisition warns.
• Assert that with min_usable_pixel_fraction=0.7, a scene at 0.3 returns direction 0.0 and quality_gate_passed False while still reporting its Z-score for audit.
• Assert every ImagerySignalType member has a HIGH_READING_DIRECTION entry.
• Run python -m unittest discover -s skills/satellite-imagery-based-signal-research/scripts and confirm a 100% pass rate.

Related Skills

• backtesting-alt-data-strategies-with-realistic-availability-lag
• credit-card-transaction-data-signal-construction
• weather-data-signal-research-for-commodity-strategies
• alternative-data-vendor-due-diligence-checklist
• insider-trading-controls-for-alternative-data-usage
• lookahead-bias-elimination
• web-scraped-sentiment-data-pipeline