Chase forecasting
Chase forecasting model comparison
HRRR says storms fire at 5. GFS says nothing happens. NAM disagrees. Here is which to trust when.
The models to know
HRRR
3 km CONUS. Convection-allowing. Hourly runs. Best for storm mode within 18 hrs.
RAP
13 km CONUS. Hourly. Best for evolving synoptic setup.
NAM (12 km)
4 daily runs. Good for outlook 24-48 hrs.
NAM 3km nest
CONUS 3 km subset. Storm-scale forecasts.
GFS
Global 13 km. 4x daily. Best beyond 48 hrs.
ECMWF
European. Best global model overall for 3-10 day.
GEFS
31-member GFS ensemble. Uncertainty quantification.
EPS
ECMWF ensemble. Similar.
SREF
Short-range ensemble. Storm-mode uncertainty.
CMC/GEM
Canadian. Sometimes best for northern US.
AI models (GraphCast, Aurora, Pangu)
Best medium-range track. Not chase-ready yet.
The forecast timeline
- 10+ days out: GFS + ECMWF. Very broad pattern.
- 5-10 days: ECMWF + GEFS/EPS. Pattern recognition.
- 3-5 days: GFS + ECMWF + Canadian.
- 2-3 days: All above + NAM. Convergence check.
- 24-48 hrs: NAM 12km + 3km nest, HRRR, RAP.
- 12-24 hrs: HRRR every hour. Storm-mode focus.
- 0-12 hrs: HRRR + observed reality.
- 0-6 hrs: HRRR + radar + own eyes.
The model biases
HRRR
Sometimes fires storms too early. Sometimes underforecast wedges.
RAP
Reliable but coarser than HRRR for exact storm placement.
NAM
Historically over-forecast CAPE. Under-forecast shear.
GFS
Sometimes too aggressive with dryline placement.
ECMWF
Excellent overall but slow to update.
AI models
Under-forecast rare extreme events (training data bias).
Reading the ensemble
- Look at spread across all members.
- Tight cluster = high confidence.
- Wide spread = uncertain.
- Outliers may be right — don't dismiss.
- For chase target, look at plume of possible storm tracks.
- For go/no-go, ensemble mean is a start.
The specific composite indices
CAPE
Instability. Storm potential.
CIN
Cap. Higher = harder to break.
0-6 km bulk shear
Storm organization.
0-1 km SRH
Low-level rotation potential.
STP
Significant Tornado Parameter.
SCP
Supercell Composite Parameter.
EHI
Energy Helicity Index.
LCL height
Cloud base height. Lower = better tornado.
Effective SRH
Modern preferred over 0-1 km.
The chase-day model workflow
- 7 AM: HRRR 12Z run. Check storm mode.
- 8 AM: SPC Day 1 Outlook update.
- 9 AM: RAP + HRRR check.
- 10 AM: Confirm target zone.
- 11 AM: HRRR 15Z run. Refine.
- 12 PM: Position toward target.
- 2 PM: HRRR + observed radar.
- 4 PM: Adjust target based on cumulus development.
- 6 PM: All models secondary — visual and radar primary.
When models disagree
- Check SPC outlook first — SPC is human-integrated.
- Look at which model has been closer this week.
- Consider forecast confidence in the discussion.
- Weight by track record for this synoptic pattern.
- Give ECMWF benefit of the doubt for medium-range.
- Give HRRR benefit for short-range.
- Sometimes: split the difference.
- Sometimes: pick one and commit.
The pitfalls
- Trusting model output beyond model skill window.
- Focusing on point forecast, ignoring uncertainty.
- Chase-worthy CAPE + shear ≠ chase-worthy day.
- Model bias in specific regions.
- Time-lapse addiction: watching runs converge is not planning.
- Ignoring warm layer aloft or high LCL.
- Overinterpreting single-run features.
For beginners
- Start with SPC outlook.
- Read the Convective Discussion.
- Use HRRR the day of.
- Learn one model at a time.
- Compare forecast with reality afterward.
- Log lessons.
- Add complexity slowly.
For experienced chasers
- Multi-model ensemble in your head.
- Weight each per pattern.
- Trust your read of the atmosphere.
- Models are inputs, not outputs.
- The last hour of visual observation beats model.
- Chase what you see.