The forecasting lifecycle

You’ve seen how data flows and where config files live. This page connects them in motion: what happens, in order, when a forecast actually runs — and which config file is in charge at each moment.

The lifecycle at a glance

A forecast run is a workflow executing its modules in sequence. A typical run looks like this:

  ┌─ 1 ─┐   ┌─ 2 ─┐   ┌─ 3 ─┐   ┌─ 4 ─┐   ┌─ 5 ─┐   ┌─ 6 ─┐   ┌─ 7 ─┐
  │ set │──▶│import│─▶│ pre- │─▶│ run  │─▶│ post-│─▶│ check│─▶│export│
  │ T0  │   │ obs  │  │ proc │  │ model│  │ proc │  │ thr. │  │ /show│
  └─────┘   └──────┘  └──────┘  └──────┘  └──────┘  └──────┘  └──────┘

Each step maps to configuration you now recognize:

Step	What happens	Governed by
1. Set T0	Fix the forecast time zero — “now” for this run. Everything before T0 is history; everything after is forecast.	The task/run definition (and system time settings).
2. Import observations	Pull in the latest observed data through hops ①–③.	Import modules + ID mapping.
3. Pre-processing	Fill gaps, convert stage→flow via rating curves, aggregate, disaggregate.	Transformation modules.
4. Run the model	Hand inputs to an external model (HBV, Sacramento, a hydraulic model…), run it, read results back.	The General Adapter module.
5. Post-processing	Combine members, apply error correction, derive secondary series.	More transformation modules.
6. Check thresholds	Compare results against warning levels; raise events where exceeded.	Threshold configs.
7. Export / display	Write results to files/services and surface them in the UI.	Export modules + display configs.

All seven steps are listed, in order, in a single workflow file. The workflow is the script; the modules are its lines.

T0: the concept everything pivots on

T0 (time zero) is the single most important idea in a FEWS run. It’s the instant the forecast is anchored to:

        ◀──────────── history ────────────┼──────── forecast ────────▶
                                          T0
     observed data lives here          "now"        model output lives here
     (imported, hop ①–③)                            (produced by hop ④)

• Everything left of T0 is where FEWS expects observed data — imports fill this region.
• Everything right of T0 is where the model writes its predictions.
• A single time series can straddle T0: observed up to now, forecast beyond.

When people say a forecast “runs at T0 = 06:00,” they mean the whole pipeline is anchored so that 06:00 is the boundary between imported history and simulated future. Get T0 wrong and the model gets the wrong initial state — a classic source of subtly-off forecasts.

Cold start vs. warm start

Step 4 needs the model’s state (soil moisture, storages, levels) to begin from. Where that comes from is a lifecycle choice worth knowing early:

• Cold start — begin from a fixed, pre-defined state file (ColdStateFiles). Simple, but the state may not reflect recent conditions.
• Warm start — begin from the state saved by a previous run. More accurate, because the model carries real recent history forward.

Operational systems run warm; you’ll often use cold starts while developing and testing.

Where “what-if” and reruns fit

Because a run is just a workflow anchored at a T0, you get powerful things almost for free:

• Re-run history by setting T0 to a past time — useful for testing against events you already know the outcome of.
• What-if scenarios by running the same workflow with modified inputs and comparing the resulting forecast series side by side.

Both reuse the exact same pipeline; only the inputs or the T0 change.

What to take away

• A forecast run is a workflow executing modules in a fixed order: set T0 → import → pre-process → model → post-process → thresholds → export.
• T0 divides imported history from simulated future; it anchors the whole run.
• The model starts from a state — cold (fixed file) or warm (previous run’s output).
• Reruns and what-ifs are the same pipeline with a different T0 or different inputs.

Next: keep the Glossary handy as you move into the task guides.