asgard-ai-platform ARIMA Time Series Model
Overview
ARIMA(p,d,q) combines autoregression (AR), differencing (I), and moving average (MA) for time series forecasting. Seasonal variant: SARIMA(p,d,q)(P,D,Q,s). Requires stationary data (achieved through differencing). Best for univariate series with clear trend/seasonality patterns.
When to Use
Trigger conditions:
- Forecasting univariate time series (sales, demand, traffic)
- Data has clear trend and/or seasonal patterns
- Need interpretable model with statistical properties
When NOT to use:
- For multivariate forecasting with many external features (use ML models)
- For very long-range forecasts (ARIMA confidence intervals widen rapidly)
- For irregular/event-driven data (use causal models)
Algorithm
IRON LAW: ARIMA Requires STATIONARY Data
Non-stationary data (trend, changing variance) violates ARIMA assumptions.
Test stationarity with ADF test (p < 0.05 = stationary).
If non-stationary: difference the series (d=1 usually suffices).
If still non-stationary after d=2, ARIMA may not be appropriate.
Phase 1: Input Validation
Check: regular time intervals, no missing values (impute if needed), minimum 50 observations (ideally 2+ full seasonal cycles). Test stationarity with ADF test. Gate: Data is regular, sufficient length, stationarity assessed.
Phase 2: Core Algorithm
- Stationarity: ADF test. If p > 0.05, difference (d=1). Retest.
- Parameter selection: Examine ACF/PACF plots. Or use auto_arima (AIC-based grid search).
- p (AR terms): PACF cutoff lag
- q (MA terms): ACF cutoff lag
- d: number of differences needed
- Fit model: Maximum likelihood estimation
- Forecast: Generate predictions with confidence intervals
Phase 3: Verification
Check residuals: should be white noise (no autocorrelation). Ljung-Box test (p > 0.05 = no autocorrelation). Residuals normally distributed. Gate: Residuals pass Ljung-Box test, no remaining patterns.
Phase 4: Output
Return forecasts with confidence intervals.
Output Format
{
"forecasts": [{"period": "2025-04", "forecast": 1250, "lower_95": 1100, "upper_95": 1400}],
"model": {"order": [1,1,1], "seasonal_order": [1,1,1,12], "aic": 520.3},
"metadata": {"training_periods": 60, "forecast_horizon": 12}
}
Examples
Sample I/O
Input: 12 monthly observations with upward trend: [10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30, 32]
Step 1: First difference = [2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2] (constant → stationary, d=1 sufficient)
Step 2: ARIMA(0,1,0) random walk with drift μ=2 is the simplest fitting model.
Expected forecast (ARIMA(0,1,0) with drift=2):
- Period 13: 32 + 2 = 34
- Period 14: 32 + 4 = 36
- Period 15: 32 + 6 = 38
Verify: differenced series is constant (2) → no AR/MA terms needed. Residuals are exactly 0 → perfect fit (toy example). On real data, residuals should pass Ljung-Box (p > 0.05).
Edge Cases
| Input | Expected | Why |
|---|---|---|
| No trend, no seasonality | ARIMA(p,0,q) | No differencing needed |
| Strong trend only | ARIMA(p,1,q) | Single difference removes linear trend |
| Multiple seasonalities | ARIMA may struggle | Consider Prophet or TBATS instead |
Gotchas
- Over-differencing: d=2 when d=1 suffices introduces unnecessary noise. Check if first difference is stationary before differencing again.
- Auto-ARIMA isn't magic: AIC-based selection can pick overfit models. Always check residual diagnostics regardless of auto selection.
- Confidence intervals widen fast: Multi-step forecasts accumulate uncertainty. Don't trust point forecasts beyond 2-3 seasonal cycles.
- Calendar effects: Business days, holidays, and leap years affect monthly/weekly data. ARIMA doesn't handle these natively — add regressors or use Prophet.
- Structural breaks: ARIMA assumes the data-generating process is stable. COVID, market shocks, or policy changes break this assumption.
References
- For ACF/PACF interpretation guide, see
references/acf-pacf.md - For SARIMA seasonal parameter selection, see
references/seasonal-arima.md