general-time-series-forecasting

name	general-time-series-forecasting
description	Plan and review general time-series forecasting with classical models, ML, foundation models, backtesting, leakage checks, and uncertainty. Use for non-omics forecasting, sensor streams, operations, finance, clinical vitals, demand, or irregular temporal data.
user-invocable	true
license	Apache-2.0

General Time-Series Forecasting

Source adaptation: BioRouter-original synthesis inspired by Scientific Agent Skills time-series and forecasting patterns, rewritten under Apache-2.0.

When to Use

Use this skill for general forecasting or temporal prediction outside BioRouter's omics-specific temporal-genomics BioSkills. Use it for sensors, operations, finance, clinical vitals, weather, demand, events, or multivariate temporal panels.

Intake

Collect:

• Forecast target, horizon, cadence, and decision use.
• Entity structure: single series, panel, hierarchy, spatial-temporal, or event stream.
• Exogenous covariates and when they are known.
• Missingness, irregular sampling, seasonality, interventions, and regime changes.
• Train/validation/test dates and backtesting expectations.
• Required uncertainty: intervals, quantiles, scenarios, or calibrated probabilities.

Model Routing

• Baseline first: seasonal naive, moving average, ETS, ARIMA/SARIMA, or Prophet-style decompositions.
• Use ML when feature history and panel size justify it: gradient boosting, random forests, neural nets.
• Use deep/foundation models only after checking horizon, frequency, scaling, covariate availability, and license.
• Use causal methods when the task asks about interventions, not just forecasts.
• Use financial-accounting-econometrics for market/event-study tasks and clinical-decision-support-review for patient-impacting forecasts.

Backtesting Gates

1. Use time-ordered splits only.
2. Prevent leakage from future covariates, normalization, target encoding, or imputation.
3. Compare against simple baselines.
4. Report horizon-specific metrics.
5. Check calibration if intervals or probabilities are reported.
6. Preserve timestamp timezone, calendar, and missingness decisions.

Output Shape

Return:

• Forecast framing.
• Baseline and candidate models.
• Backtest plan.
• Leakage checklist.
• Metrics and uncertainty plan.
• Minimal reproducible code or commands if requested.