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.