A forecast is one estimate. Its benchmarks, competing models and historical errors help us assess how much weight to place on it.
Angavu evaluates different forecasting approaches. Which models appear in a published comparison depends on the indicator, available data and evaluation evidence.
01 / Benchmarks
Random walk, integrated moving average (IMA(1,1)) and direct autoregression provide benchmark forecasts. More elaborate models are evaluated against their forecasting performance.
02 / Time series
Autoregressive integrated moving average (ARIMA) models are evaluated using fixed, recursive and rolling estimation windows. Their forecasting performance is compared with the benchmarks.
03 / Inflation
Inflation-specific models examine the relationship between inflation, its past values and economic slack. Structural Phillips-curve models are evaluated only when the necessary expectations and slack information would have been available at the forecast origin.
NMSFE compares a model's mean squared forecast error with the corresponding benchmark. Values below 1 indicate lower forecast error than the benchmark.
Evaluation / 01
Angavu compares models using their forecasting errors over a historical validation period. A model with a low error may still require stability checks or review before being used in a published forecast. Where sufficient later data are available, separate tests examine whether the selected model continues to perform well and whether forecast adjustments improve its accuracy.
Release / 02
Each forecast identifies the period being predicted, the model used, its forecast origin and the information cutoff. Forecasts reconstructed later are distinguished from forecasts actually produced and frozen at the time.
02 / Research record
Follow the selected forecast through model comparisons, historical revisions, diagnostic results and its publication record.
Each row represents a target period and horizon in the current forecast curve. Unavailable forecasts remain identified rather than estimated.
Target period
Compare forecasts for the same target period across different origins. Where outcomes are available, observed values and forecast errors show how those estimates performed. The headline remains the current-origin forecast.
The forecast origin is the period from which the forecast is made.
| Origin | Horizon | Model | Forecast | Outcome | Error |
|---|---|---|---|---|---|
| 2025Q4 | h4 quarters | ARIMA(1,0,1) | 3.14% | — | — |
| 2026Q2 | h2 quarters | ARIMA(1,0,1) | 4.60% | — | — |
How the selected forecast was formed
Competition
The selection trace records evaluated models, including models not retained for publication. The observed value at the forecast origin is dated 2026Q2, with information available through 2026-10-12.
| Model | Estimation approach / model family | MSFE | NMSFE | Selection |
|---|---|---|---|---|
| ARIMA(1,0,1) | recursive · other:unknown | 6.792 | 0.937 | Selected |
| ARIMA(1,0,1) | rolling · other:unknown | 6.811 | 0.942 | Published alternative |
| ARIMA(1,0,1) | fixed · other:unknown | 7.146 | 0.931 | Published alternative |
| IMA(1,1) | rolling · Benchmark | 7.229 | 1.000 | Retained for evaluation |
| IMA(1,1) | recursive · Benchmark | 7.246 | 1.000 | Retained for evaluation |
| IMA(1,1) | fixed · Benchmark | 7.678 | 1.000 | Retained for evaluation |
| Parameter | Estimate | Standard error (SE) |
|---|---|---|
| const | 3.7087 | 0.8920 |
| ar.L1 | 0.9083 | 0.0270 |
| ma.L1 | -0.3989 | 0.0676 |
| sigma2 | 3.8896 | 0.1623 |
Validation & stability
Reported p-values are exploratory and are not adjusted for model selection or multiple testing.
Independent holdout evaluation
No model-based prediction interval is reported. Historical forecast errors, disagreement between models, stability diagnostics and independent evaluation, where available, inform interpretation. None should be treated as a prediction interval.
Produced and frozen at its forecast origin using the information available at that time.
The observation cutoff is the latest input observation included in the model. The information cutoff identifies the latest information permitted for the forecast. Additional research evidence preserves how the published result was produced.