# AI Time Series Forecasting

> Design and evaluate time-series forecasting — ARIMA, Prophet, NeuralProphet, LSTM, Transformer — with stationarity analysis, seasonality, exogenous variables, and backtest.

- Skill: `quantumquirkxyz/ai-time-series-forecasting` (Agent Skill)
- Install (CLI): `npx skillmds@latest add quantumquirkxyz/ai-time-series-forecasting`
- Raw SKILL.md: https://api.skillmd.com/api/skills/quantumquirkxyz/ai-time-series-forecasting/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: quantumquirkxyz (https://skillmd.com/u/quantumquirkxyz)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/quantumquirkxyz/ai-time-series-forecasting

---


## Contract

- **Input:** time-series data (date + target; optional exogenous variables).
- **Output:** forecast plot + comparison + recommendation.
- **Side effects:** none.
- **Dependencies:** time-series data source.
- **Stop condition:** backtest complete; recommendation made.
- **Risk:** medium — forecast affects planning; requires validation.
- **Boundary:** designs forecasting pipeline; does not make business decisions.

# Time-Series Forecasting

Build a **time-series forecast** — with stationarity analysis, seasonality, exogenous variables, and backtest — and recommend the best model.

## Process

### 1. Data inspection
Plot series; check for missing values, outliers, structural breaks. Check stationarity: ADF test, KPSS test. Decompose: trend + seasonal + residual.

**Completion criterion:** stationarity state stated; seasonality identified.

### 2. Feature engineering
Lag features (autoregressive terms); rolling statistics; exogenous variables (available at forecast time); categorical time features.

**Completion criterion:** feature set documented; no leakage.

### 3. Model selection
Classical (ARIMA / SARIMA / Prophet), gradient-boosted (XGBoost / LightGBM), deep learning (LSTM / Transformer / N-BEATS), or ensemble. Select by backtest performance, not training error.

**Completion criterion:** model family selected with justification.

### 4. Backtest
Rolling / expanding window; respect time order; evaluate MAPE, sMAPE, MASE, RMSE, MAE.

**Completion criterion:** backtest complete; metrics on holdout.

### 5. Prediction intervals
Provide 80% / 95% intervals — not just point forecasts.

**Completion criterion:** intervals in plot.

### 6. Recommendation
Best model with evidence; limitations (exogenous availability, structural breaks, seasonality change); re-training frequency.

**Completion criterion:** recommendation with conditions.