Observable evidence. Explicit uncertainty. Verifiable outcomes.
The methodology separates real-world measurement from economic forecasting and maps forecasts to potential market relevance only after the economic claim is formed.
Observations
Detection
Economic factors
Forecast distribution
Market relevance
Research question
Can observable changes in physical production, movement, and utilization improve estimates of economic variables before official publication? Metrivane defines a target variable and forecast horizon before selecting sources or features.
Data sources
Sources may include satellite imagery, AIS vessel positions, mobility aggregates, gridded weather, road traffic, port records, and public economic statistics. Each source receives a provenance record, expected latency, geographic scope, and stability assessment.
Feature construction
Raw observations are transformed into economically interpretable factors such as terminal occupancy, vessel waiting time, freight corridor intensity, weather stress, and throughput pressure. Factors are normalized using only information available at each historical date.
Object and event detection
Computer vision and rules-based systems detect objects and events. Detection uncertainty is retained and aggregated; a detected object is evidence, not an economic conclusion. AIS matching and temporal consistency checks reduce duplicate observations.
Economic factor model
Correlated observations are combined into standardized latent factors. The factor layer is designed to separate measurement from forecasting and allows analysts to inspect which physical processes changed the estimate.
Forecasting model
Regularized econometric and time-series models produce a distribution over the target variable. Ensembles may combine autoregressive baselines, dynamic factors, and nonlinear models. Large language models are not the primary forecasting engine.
Validation
Evaluation uses walk-forward splits, vintage-aware data, simple baselines, calibration diagnostics, and proper scoring rules. Metrics cover all eligible forecasts rather than selected successes.
Revisions
Every material forecast revision is timestamped with its changed factors and source availability. The final pre-release snapshot is evaluated against the first stable official data vintage, with later revisions shown separately.
Limitations
Spurious correlations, data revisions, source instability, geographic bias, incomplete coverage, model drift, uncertainty, and alternative explanations can all impair a forecast. Relationships may fail in new regimes, and confidence intervals do not cover every form of uncertainty.
Data ethics and provenance
Metrivane prioritizes lawful, documented, and appropriately aggregated sources. It avoids individual-level profiling and presents activity as economic measurement rather than surveillance. Source licensing, retention, and geographic bias require continuing review.