vera-data-timeseries-reviewing

name: vera-data-timeseries-reviewing description: >- Runs time series diagnostics and primary ARIMA modeling for temporal data. Produces time plot, ACF/PACF, seasonal decomposition, stationarity tests (ADF + KPSS), and one fully interpreted ARIMA model with forecast and prediction intervals. Identifies trend, seasonality, and non-stationarity; fits ARIMA/SARIMA as baseline. Advanced methods (Prophet, state-space, VAR) delegated to vera-data-timeseries-generating. Outputs .R and .py scripts with 2 publication-quality plots. Triggered when user has temporal/time series data and says "time series," "temporal data," "forecast," "ARIMA," "seasonal," "trend," "autocorrelation," "monthly data," "daily data," "quarterly," "stationarity," or names a time-indexed variable like sales over time, temperature, stock price, monthly passengers. Does not handle panel/longitudinal data (redirect to vera-data-repeated-reviewing, or vera-data-repeated-generating for growth-curve / mixed-model extensions), cross-sectional data, or spatial time series. allowed-tools: Read, Bash, Write, Edit

Time Series — Diagnostics & Primary Modeling

Scope Boundary
Workflow
Decision Tree
Required Inputs
Code Structure
Reporting Standards
Hypothesis Tests
Example Dataset
Method Status
Minimal Smoke Test
Cross-Skill Interface

Open-source skill.

Scope Boundary

Use this skill when:

The data are a univariate series indexed by calendar time and the first need is a classical ARIMA-style baseline.
Trend, seasonality, stationarity, and short-horizon forecasting are the immediate priorities.

Do not use this skill when:

The data are panel / longitudinal across subjects with few waves; use vera-data-repeated-reviewing or vera-data-repeated-generating.
The task is multivariate cross-series modeling, regime switching, or advanced forecasting from the start.

Workflow

Read each step file in workflow/ before executing that step.

Step	Responsibility	Executor	Document	Input	Output
Collect	Collect Inputs	Main Agent	`workflow/step01-collect-inputs.md`	User input	Structured input summary
Diagnose	Check Distribution	Main Agent	`workflow/step02-check-distribution.md`	Prior step output	PART 1 code block
Model	Run Primary Test	Main Agent	`workflow/step03-run-primary-test.md`	Prior step output	PART 2-3 code blocks

Decision Tree

1. STATIONARITY CHECK
   ├── Stationary (ADF p < .05 AND KPSS p ≥ .05) → fit ARIMA(p,0,q)
   ├── Non-stationary → difference, retest, fit ARIMA(p,d,q)
   └── Seasonal pattern detected → recommend SARIMA

2. MODEL SELECTION
   ├── ACF/PACF inspection → candidate ARIMA(p,d,q) orders
   ├── Auto-ARIMA for automated selection (AIC minimization)
   └── Ljung-Box residual diagnostics → confirm adequacy

Required Inputs

Role	What to collect
Time series (Y)	Variable name, units, what it measures
Frequency	Daily, weekly, monthly, quarterly, annual
Date/time index	Column name or implicit ordering
Exogenous (optional)	Any external predictors (for recommendation)

Code Structure

PART 0: Setup & Data Loading
PART 1: Time Series Diagnostics   → plot_01_ts_diagnostics.png
PART 2: Primary ARIMA Model       → plot_02_forecast.png
PART 3: Recommendation Block      → text listing additional analyses available

Reporting Standards

p-values: "< .001" not "0.000"; exact to 3 decimals otherwise
ADF test: report test statistic + p-value + number of lags used
KPSS test: report test statistic + p-value + bandwidth
Model notation: ARIMA(p,d,q) or ARIMA(p,d,q)(P,D,Q)[s]
Information criteria: AIC and BIC
Ljung-Box: Q statistic + lag + p-value. Use adaptive lag = min(⌊T/5⌋, 20) where T is series length — the fixed (5, 10, 15) scheme can underdetect misspecification in short series and overtest in long ones. Report the chosen lag explicitly.
Forecast intervals: 80% and 95% prediction intervals
Accuracy metrics: RMSE, MAE, MAPE on training data
Decimal places: 2 for coefficients, 3 for p-values
Non-significance: "not statistically significant at alpha = .05" — never "no effect"

Hypothesis Tests

Test	Null Hypothesis	Stationary if
ADF (Augmented Dickey-Fuller)	Unit root present (non-stationary)	p < .05 (reject null)
KPSS	Series is stationary	p >= .05 (fail to reject null)

Conflicting results → note ambiguity, proceed with differencing as conservative choice.

Example Dataset

R built-in AirPassengers: monthly airline passengers 1949-1960. Python: sm.datasets.get_rdataset("AirPassengers") (with offline fallback to bundled examples/airpassengers.csv).

Method Status

Status	Methods
Implemented in this skill	ADF/KPSS diagnostics, ACF/PACF-based ARIMA triage, baseline ARIMA/SARIMA fit, forecast intervals, residual checks
Implemented downstream in `vera-data-timeseries-generating`	ETS, GARCH, VAR, spectral analysis, regression with ARIMA errors, ML forecasting on lagged features
Out of scope in this open-source baseline	Panel time series, spatial time series, and forecasting families not explicitly named above

Minimal Smoke Test

Smoke-test prompt: "Run vera-data-timeseries-reviewing on AirPassengers, using the passenger count as the series and monthly frequency. Produce the standard baseline artifacts."

Cross-Skill Interface

Output:
├── code_r      → .R script
├── code_python → .py script
├── figures/    → 2 PNGs (diagnostics + forecast)
└── recommendations → text block (additional analyses available)

Next step: Invoke vera-data-timeseries-generating from this skillset to run the full pipeline (additional tests, subgroup analysis, modeling, manuscript generation). See ../../CROSS-SKILL-INTERFACE.md for the shared handoff contract.