GEFS Ensemble - Tennessee Weather Network

--

All six products advance together through the shared lead times - one clock, six maps: synoptic pattern, temperature, moisture, jet, instability, winter precip. Thin annotations follow each map.

📆 Forecast verification lead-time picker · all leads vs the same observed hour · valid Fri 8 AM ET

GEFS mslp F072 forecast — MSLP + 500 mb heights · F072 from init AM ET · valid AM ET

mslp observed analysis — MSLP + 500 mb heights · F072 from init AM ET · valid AM ET

GEFS t2m F072 forecast — Temperature (2 m) · F072 from init AM ET · valid AM ET

t2m observed analysis — Temperature (2 m) · F072 from init AM ET · valid AM ET

GEFS cape F072 forecast — CAPE (instability) · F072 from init AM ET · valid AM ET

cape observed analysis — CAPE (instability) · F072 from init AM ET · valid AM ET

GEFS pwat F072 forecast — PWAT (moisture) · F072 from init AM ET · valid AM ET

pwat observed analysis — PWAT (moisture) · F072 from init AM ET · valid AM ET

Skill vs lead time - last 90 days (r & ACC)

Pattern correlation by valid day and lead

Anomaly correlation by valid day and lead

Day 3: r 0.83–1.00 over 90 daysDay 5: r 0.68–0.99 over 90 daysDay 7: r 0.50–0.99 over 90 daysDay 10: r 0.04–0.99 over 90 days

Pattern correlation r vs the GFS analysis, one point per valid day, one color per lead - the gap between the lines is the skill lost with lead time; the slope is day-to-day pattern variability. Scores are recomputed nightly from the archive and kept 90 days. ACC (second chart, red 0-line) is the anomaly correlation - fcst and obs vs the same-day climatology (mean of the last 5 years of GFS analyses): positive means the forecast beat climatology, the standard skill test, and points below the line are busts. S/S is the ensemble spread-skill ratio (spread / actual RMSE): ~1.0 means the 31 members know when they are uncertain, <1 means overconfident, and the parenthesised r links spread to where the error actually landed.

Each pair sets a lead's ensemble-mean forecast beside the observed analysis for the SAME valid hour - pick a lead to see how much skill the forecast loses with range (the grid frames above only survive 48 h, so these are re-fetches from NOAA's archive). Same levels and palette as the grid above, so the differences you see are the forecast's, not the map's. Truth: GFS 0.25-deg analysis (noaa-gfs-bdp-pds, no keys).

New to forecast verification? Start here

Every forecast is easy to check in hindsight - this card does it with numbers. The idea: take a real forecast from 3, 5, 7, or 10 days out (kept in NOAA's archive), wait for the weather to happen, then compare the two maps and score them. A score near the top means the forecast put the right weather in the right place.

Score	Plain-English question it answers	Good
RMS	On average, how far off was the number? (hPa, degrees, ...)	lower is better
r	Do the map's highs and lows line up with what happened?	1 = perfect
ACC	Did it beat just predicting "normal weather for the date"? (vs a 5-year climatology)	above 0; below 0 = worse than climatology
S/S	Did the 31 members disagree as much as the forecast turned out to be wrong? (their spread vs the actual error)	near 1.0; <1 = overconfident
bias	Did it systematically run too warm/wet (+) or too cool/dry (−)?	near 0

Rules of thumb from our own 77-day record (MSLP): day-3 forecasts are near-perfect (ACC 0.96); day-5 are good (0.88); day-7 are useful but rougher (0.72); day-10 average 0.55 - and in our record they dipped below 0 once, meaning on that day flipping a coin with climatology would have done better. Temperature holds up much better than rain-based fields at long range: pattern errors grow fastest in CAPE and PWAT, so trust a day-10 temperature outlook more than a day-10 storm forecast.

How to read a pair of maps: start with the forecast (left/top image), form a picture - where is the low, where is the warm tongue - then look at the observed map and ask what moved. Then check the badge: high ACC but visible differences means the big picture was right and details moved; ACC near 0 means the whole pattern was wrong. The S/S number tells you whether the ensemble was honestly uncertain about it.

Both chart panels show these scores for every valid day of the last ~3 months, one colored line per lead: the vertical gap between lines is skill lost with lead time; the red 0-line on the ACC chart is the bust line. Data: NOAA GEFS + GFS, no keys. Scores are computed nightly by this site from the public archive.

📅 Day high / low TMAX | TMIN ensemble means · East TN town table · init Fri 8 AM ET

--

🧭 Deep dives

Full-screen animated panels with ensemble spread and (on the jet) 16-member axis spaghetti live on the Models page GEFS card. Storm tracks from these same members are on the NHC Tropical page.