What this page is for. It is for the synoptic scale: reading the pattern across the whole ensemble, where the members agree, where they split, and which scenarios lie behind the split. That is why its grids are 0.25° near land and 1° elsewhere, and why every map is a statistic of 50 members rather than one model run. It is not for the smallest detail, such as the deepest snow at one point, one run's heaviest shower or a valley's lowest temperature: members cannot place that detail, and fields that only make sense at that scale (snow depth among them) are left out on purpose. Use a high-resolution deterministic model for those.
The explorer shows the 50-member ECMWF ensemble (ENS) over the northern hemisphere, for the two newest runs. It is built for forecasters: every map is worked out on request from every member, so you choose the field, the period, the statistic and the threshold.
- The data
- The map and the grids
- Layers
- Worked-out fields
- Displays (keys 1–5)
- Probabilities and neighbourhoods
- Periods
- Boxes, thirds and x–y
- Plumes
- Event timing
- Fronts
- Cyclones
- Below the ground
- Setups and colour scales
- Keys
- Saving pictures and animations
- Sharing a view
- Hovmöllers
- Run-Run
- Clustering
The data
Each 00Z and 12Z run is downloaded from ECMWF's open data (CC-BY-4.0) and held in one Icechunk store. The three newest runs are kept. Fields are 6-hourly. The Hovmöller tab has its own, much smaller archive that goes back 60 days (below).
| Grid | Area | Steps | Fields |
|---|---|---|---|
| 1° | 0–90°N, every longitude | T+0 to T+240 | Every field |
| 0.25° | 12–70°N, 130°W–60°E, over land and within 250 km of it (MSLP everywhere in the area) | T+0 to T+120 | MSLP, 2 m temperature and dew point, gust, 100 m wind, precipitation, snowfall, MUCAPE, 1000–850 hPa thickness, 850 hPa temperature, humidity and wind (so θw and 850 hPa wind speed) |
The ensemble mean and spread of every field are stored as each run comes in, so those maps are instant; probabilities, percentiles, periods and thirds are worked out from all the members when you ask.
The map and the grids
The map is equirectangular (Lat–lon), true to scale at 45°N, and repeats round the dateline, or Polar stereographic, chosen under Map. The polar map looks down on the pole, has no break at the dateline and keeps shapes and angles locally, so the flow round the hemisphere (jets, troughs, the paths of cyclones) reads as one. Its upright longitude is the meridian that runs straight up the screen to the pole: set it to the region you care about (30°W for the Atlantic, 0° or 20°E for Europe), so that features upstream sit to its left and the flow over it runs left to right. Changing projection or upright longitude keeps the middle of the map. Wind barbs are turned to each projection's north. The side panel's thirds map and the member stamps stay lat–lon.
The colour scales and keys of everything on the map sit in its lower right corner; fold them away with the arrow on their heading (your browser remembers it). The full legend, with notes, is at the foot of the left panel. Drag the map to move it; click it for plumes at that point.
Drag the left edge of the side panel to make it wider or narrower (your browser remembers it); double-click the edge to put it back.
Every field is drawn on one 0.25° grid. Where only 1° values exist they are interpolated (bilinear). Inside the 0.25° area the 0.25° values are blended in: fully where they are kept, fading to the 1° values over the outermost 1.5° (6 points) of what is kept, so there is no seam where one grid meets the other. The ensemble statistic is worked out on each grid's own points and then blended. The dashed outline shows the 0.25° area. The badge by the time bar says which grids the map is using: after T+120 everything is 1°.
Relief, coastlines, borders, rivers and place names come from the ERA5 climate explorer's basemap (ETOPO 2022 and Natural Earth). Turn them on and off under Map.
Layers
A map is a stack of layers; the top of the list draws on top. Click a layer for its settings. ⛓ makes a layer follow the field and period of the layer above (barbs follow its level), so changing the top layer changes both.
| Type | Shows |
|---|---|
| Shaded | Colour fill of any display below |
| Shaded + isolines | The same, with isolines of the shaded map on top (every 10% for a probability) |
| Isolines | Contours of the mean, a percentile, or one member |
| Wind | Ensemble-mean wind in knots at 10 m, 850, 700, 500, 250 or 10 hPa, or the 850–200 hPa steering flow, drawn as barbs, arrows or streamlines; the whole wind or only its divergent part (see below) |
| Numbers | Values printed at evenly spaced points |
| Fronts | Automated 850 hPa fronts: one member's, every member's, or how often members have one |
| Cyclones | Every member's cyclone centres, coloured by stage |
Isolines take a colour, a width and a style (solid, dashed or dotted), and an interval if the field's own does not suit. Every wind (gusts, speeds, components and barbs) is in knots.
Arrows point the way the wind blows, as long and as thick as it is strong: the legend gives the speed of a full-length arrow (about the strongest 5% over the map, rounded). Hide arrows below leaves out the weak winds so the strong ones stand out (5 kt by default for the divergent wind, where it picks out the outflow over rising air). Streamlines follow the wind's direction only, with an arrowhead halfway along each; read the speed from shading beneath them. Both are drawn from the ensemble mean.
Label extremes (isolines and shading) marks highs and lows (H and L, in red and blue, with the value) or only the maxima or minima, such as the core of a jet ("Max 180"). A centre must be the highest or lowest point for about 900 km and stand out by at least one isoline interval; where two of a kind are close, the more prominent is kept, and labels that would overlap on screen are dropped, weakest first. MSLP isolines label highs and lows from the start.
Worked-out fields
These are worked out from the stored fields member by member, so every display (probabilities, percentiles, thirds) works on them.
| Field | How |
|---|---|
| 2 m dew-point depression | 2 m temperature − 2 m dew point: small where the air near the ground is nearly saturated (fog, low cloud) |
| 10 m–700 hPa shear (about 0–3 km) | The size of the 700 hPa wind minus the 10 m wind, as vectors |
| 10 m–500 hPa shear (about 0–6 km) | The same with the 500 hPa wind: 40 kt or more favours organised convective storms |
| Relative vorticity | ∂v/∂x − ∂u/∂y on the sphere, at every level with a wind, in 10⁻⁵ s⁻¹; positive is cyclonic |
| 250 hPa velocity potential | χ, whose slope is the divergent wind. Negative where the upper air spreads out over rising air (the tropical convection that drives the MJO, or ahead of a trough) |
| Divergent wind | The irrotational part of the wind: the slope of χ, so it blows from low χ to high. Small (a few knots) next to the whole wind, but it shows where air rises and sinks. Offered on wind layers as "Wind: divergent" |
Vorticity, χ and the divergent wind need each member's whole map, so they use the 1° grid only. χ is solved over the northern hemisphere taking the southern as its mirror image, as the store holds 0–90°N only. Its zonal mean (and so the zonal-mean divergent wind, as in the Hadley circulation's outflow) is exact: it comes straight from the zonal-mean northward wind, which is all divergent. The rest is right in mid and high latitudes (within about 15–20% of a whole-globe solution) but only roughly so south of 30°N, where the mirror matters. The 10 m and 700 hPa winds (and so the shear) are stored only by runs that came in after they were added; an older run says so if you ask for them.
Displays (keys 1–5)
| Key | Display | What it is |
|---|---|---|
| 1 | Mean & spread | Spread (standard deviation of the members) shaded, the mean as isolines |
| 2 | Mean | Ensemble mean |
| 3 | Probability | % of members at or beyond a threshold (see below) |
| 4 | Percentile | A percentile of the members at each point; 50 is the median, 90 a reasonable worst case for high values |
| 5 | Members | Postage stamps of every member over the part of the map on screen; click one to see it full size, then [ ] to step through members |
For fronts the keys pick one member (1), every member (2) or the frequency (3). Box thirds are in the Box tab.
Probabilities and neighbourhoods
A probability is the share of members at or above (or at or below) the threshold at each point. With a neighbourhood it is averaged over a disc of that radius around each point (true distance, so the disc covers more longitude further north): the chance of the event somewhere nearby rather than exactly here. Use one for showery precipitation or gusts, where the members cannot place the detail.
+ and / or adds a second condition, on any field with its own threshold and period. With and the map is the share of members where both happen in the same member, such as 24 h precipitation of 1 mm or more and a 2 m dew point of 1 °C or less; with or, where either (or both) does. That is not the product of the two probabilities: it counts the members, so it keeps how the two go together in each one.
Against the climate. For 24 h precipitation and the 24 h highest gust, Threshold can be the ERA5 climate instead of a fixed value: the 90th, 95th or 99th percentile, or the highest in 30 years, for that place and time of year. The map is then the share of members at or above what is unusual there, so 20 mm in Madrid and in Fort William are no longer the same thing. The climate is ERA5 from 1996 to 2025, every year's days within 15 days of the date, so about 930 days (or 3,700 overlapping 24 hour totals) behind each percentile: well sampled up to the 99th, while the 30-year highest is a single case. It is on the 1° grid, the members' 1° values are compared with it, and it covers land only, up to 70°N: the sea is blank. ERA5 is coarser (about 31 km) than the ensemble (9 km), so heavy precipitation and strong gusts get past ERA5's percentiles more easily than they would past a climate of the ensemble itself; and ERA5's day for gusts is the local day, the members' 24 hours end at the valid time. Read it as "unusual for here", not as an exact return period.
Periods
A period combines each member's steps over the hours before the valid time, before any statistic is taken: the probability of a 24 h maximum gust over 50 kt is the share of members whose highest gust in those 24 hours is over 50 kt. Precipitation over a period is each member's total. With no period, precipitation is the total since T+0. The gust is ECMWF's highest gust since the previous output (1 h to T+90, 3 h to T+144, 6 h after), so a period's maximum gust covers the whole period. For other fields only the 6-hourly steps are used, so peaks between them are missed.
Snow. Snowfall is ECMWF's own snowfall, as the depth of water it melts to (about 1 mm is 1 cm of fresh snow, more for dry powder, less for wet snow). It is part of the precipitation, so a period's snowfall is never more than its precipitation. Snow share of precipitation is each member's snowfall over the period as a share of its precipitation: 0 is all rain, 100 all snow, and between is a mix, or rain turning to snow (or back) within the period. A member with under 0.5 mm in the period has no share, so the mean and percentiles are of the members with precipitation; a probability counts the dry members as not meeting it, so "snow share ≥ 50%" is the chance of 0.5 mm or more falling mostly as snow. For "at least 5 mm, mostly as snow" add a second condition on precipitation. Snowfall is stored by runs from 12 November 2026; an older run says so if you ask for it. Snow lying on the ground (and its melting) isn't shown: the open data rounds snow depth to about 3 cm of water (30 cm of fresh snow), too coarse to show a fall or a thaw.
Boxes, thirds and x–y
Draw a box with ▭ or b and drag on the map (drag inside it to move it, a corner to resize). The Box tab works out a statistic over the box for every member: the area-weighted mean, the max, the min, or the % of the box beyond a threshold. That ranks the members into thirds. Pin it to a time (📌) to rank by one event while you step through others.
- Thirds: a small map, centred on the box, of how the members in one third differ from all of them, for any field at the map's time or a fixed one. Shading is the third's mean minus the all-member mean, shown only where it is larger than one standard error (the all-member spread ÷ √ the number in the third); solid contours are the third's mean, dotted ones the spread within it. One colour scale serves all three thirds, so they compare directly. Hover for values. The small map is drawn in the main map's projection: on the polar map it is fitted to the box (a ring round the pole is shown whole); on the latitude–longitude map it widens to fit a box wider than 240°.
- Through time plots each member's box statistic at every step, coloured by third.
In the polar projection a box can go all the way round: drag it past about 300° of longitude and it becomes a ring (or, up to the pole, a cap), and the box statistics are over the whole band, area-weighted, which is as valid as any box. Drag its inner or outer edge to change the latitudes.
The x–y tab plots one box statistic against another, one point per member, coloured by third, with the correlation r, a bootstrap range, the 5% critical value and r² (the share of how the members differ in y that goes with x). It answers questions such as whether the members with the deepest low (box minimum of MSLP) are the ones with the strongest gusts (box maximum of the gust): often they are not. "Every grid point" instead bins every point in the box for every member: the physical relationship between two fields, with no significance test (neighbouring points are not independent).
Plumes
Click the map for plumes of the fields on the map (barbs give their wind speed) at that point, through the run: every member, the 10–90% range shaded, the median bold, the mean dashed, the min and max dotted, and the other run's 10–90% range and median in grey, placed by valid time. The table gives the numbers at the map's time. The shaded stretch on the right is after T+120, where the point uses the 1° fields. Click a plume to go to that time.
Event timing
The Timing tab asks, of each member, when a condition such as 100 m wind ≥ 50 kt starts and ends, at the point clicked or over the box (anywhere in it, its mean, or over at least a share of it). A member's start is the first 6-hourly step meeting the condition in the window, and its end the last. At the top of the chart are the median and middle 80% of the starts (blue) and of the ends (orange). Under them, the share of all the members that have started, and that have ended, by each time: the steeper a curve, the more certain that time, and the shading between the two is the share in the event. Then a bar per member from its start to its end, sorted by start, end or length. If the starts bunch together more than the ends, the start is more certain, and the line under the table says which, with the numbers.
The window. Drag across the chart (or pick From and To) to look at one event when a run has two: a member's event is its first in the window, or its longest or strongest. Outside the window is hatched. Gaps allowed stops a one-step dip below the threshold splitting an event in two. A member already meeting the condition when the window opens started at or before it, so its start counts as that (≤ in the table), and one still meeting it when it closes ends at or after it (≥); its bar has an arrow at that edge. They are counted, not left out, because they are the earliest starts and latest ends: leaving them out would make the times look more certain than they are. A spread that reaches them is only a lower bound (≥), and a cut-off event's length is unknown.
Estimate times between steps. With this on, a start is when the value crossed the threshold on a straight line between the step before and the first step meeting it (and an end likewise), rather than the step itself. These are estimates, and are labelled as such: the real crossing need not lie on that line.
Bear in mind. The steps are 6 hours apart. Wind speeds are snapshots, so a spell between two steps can be missed; the gust is the highest since the previous output, but only over the last 1 h to T+90 and 3 h to T+144. After T+120 a point uses the 1° fields, which are smoother (usually weaker winds), so an end at T+120 may be the grid changing rather than the weather. With fewer than about 10 members having the event, the spread of their times is not reliable.
Fronts
Fronts are found in each member at 850 hPa, on the 0.25° grid to T+120, with Hewson's thermal front parameter on the wet-bulb potential temperature θw. A front is kept only where a 100 m wind shift or an MSLP trough backs it. No fronts are drawn where 850 hPa is below the ground in most members. Each point is typed cold, warm or quasi-stationary from the wind across it, and each line is smoothed a little (over about 40 km) to take out the steps of the grid. The thresholds are set a little below Hewson's, so weaker fronts show too. For one member, triangles are drawn on the warm side of cold fronts and semicircles on the cold side of warm fronts (the way each moves), alternating on quasi-stationary ones, which are drawn red and blue in turn. Frequency is the share of members with a front within 100 km.
Troughs and occlusions (a tick box on the fronts layer) carry the analysis on where the fronts stop. A front needs a θw gradient; where that fades, as when a cold front runs into the occluded part of a low, the trough often goes on as a strip of cyclonic 850 hPa vorticity in an MSLP trough. Those strips' axes are drawn as lines, worked out from the 1° vorticity, θw and MSLP, so they reach 85°N, north of the fronts' 0.25° area (which stops at 70°N). A line must be at least 400 km long and fairly straight, mostly in an MSLP trough, and more than 180 km from a front (the vorticity strip just behind a front belongs to that front). Most are drawn as plain troughs (black, white in dark mode). One is an occlusion (purple, triangles and semicircles in turn on the side it moves to) only if it carries a front on (an end within 300 km of one), reaches to within 500 km of one of that member's cyclone centres, and mostly follows a ridge of θw, the tongue of warm air being lifted. Occlusions are found only where cyclones are tracked (20–80°N, 90°W–50°E). In Frequency, dashed lines give the share of members with a trough or occlusion within 100 km (25, 50 and 75%). They are worked out when asked for, so the first look at a time takes a moment longer.
Cyclones
Cyclones are tracked in every member from MSLP minima to T+240 over 20–80°N, 90°W–50°E (refined on the 0.25° grid to T+120). Each centre's stage comes from its tilt: how far, and which way, the nearest closed 500 hPa low is from the surface centre. A baroclinic low deepens while its upper low lies well to the west, and the two stack up as it matures and occludes.
| Stage | 500 hPa low |
|---|---|
| Open wave | No closed low within 1000 km |
| Developing | 500 km or more to the west |
| Mature | 200–500 km to the west |
| Occluded / stacked | Within 200 km |
| Upper low ahead | 200 km or more to the east |
Click a centre to follow that cyclone: the nearest centre within 500 km in every member at that time, followed back and forward along its track, in both runs. The Cyclone tab follows each member's cyclone through time: its central pressure (dots by stage), the 24 h change in it scaled to 60°N (× sin 60° ÷ sin latitude; below −24 hPa it is deepening explosively, a "bomb", after Sanders and Gyakum), or the strongest gust within 500 km of the centre. The x–y chart plots two of these against each other, a point per member, with the correlation: the lowest pressure against the strongest gust, for instance, asks whether the deepest members are the windiest. The table gives each member's lowest pressure and largest 24 h fall.
The centres can also be coloured by central pressure (920–1010 hPa), by the strongest 100 m wind, or by the strongest gust within 500 km of the centre in that member (keys 1–4). The scales are fixed, so a colour means the same from run to run, and the deepest or strongest are drawn on top, so the outliers stand out. The winds are from the 1° grid; 500 km can take in a neighbouring system's wind, and over land the 100 m wind is held back by friction. The gust is ECMWF's highest since the previous output, so it is noisier than the 100 m wind.
Below the ground
Over high ground the 850 hPa (and 1000 hPa) surface lies below the ground, and the model's values there are extrapolated. Pressure-level fields are hatched where the level is below the ground in most members, the plumes shade those times, and the hover note gives the share of members.
Setups and colour scales
Save as… keeps the layers (and, if ticked, the map's area) as a named setup. Setups and colour scales are kept in this browser, for you alone. The admin can also save them for everyone: shared ones are marked ★ and listed under "Shared with everyone". The Hovmöller tab has setups of its own, saved the same way.
Copy link copies the page's address, which holds everything on screen (layers, colours, time and area), so whoever opens it sees the same map. It is the quick way to show someone a map.
Download… saves the chosen setup (with any colour scales of yours it uses), or everything kept in this browser (map setups, colour scales and Hovmöller setups), as a small file. Import… adds the setups in such a file to this browser: use it to move your setups to another browser or computer, to keep a copy, or to pass setups on. Nothing is sent to the server. Layers or fields this explorer cannot show are left out, and it says which; where a name is already in use you can keep both (the new one gets "(2)") or replace yours. Clearing the browser's site data deletes the setups kept in it, so download them first.
Open with this in this browser makes a setup the one this page starts with when it is opened without a link. Otherwise it starts with the setup the admin has chosen for everyone (Make everyone's default, shown to the admin on a shared setup), and at first that is "Surface overview" (the maps) and "Waves round the hemisphere" (the Hovmöllers).
The ? beside a panel's or card's title opens a short note on it; click again, or anywhere else, to close it.
Keys
| ← → | Previous / next step; with Shift, a day |
| Home End | First / last step |
| Space | Play |
| 1–5 | Display of the open layer (else the top shaded one) |
| [ ] | Previous / next member, showing one member |
| r | Switch run, keeping the valid time |
| l | Link the open layer to the one above |
| b | Draw a box |
| Esc | Stop drawing; close the box, cyclone or point tab |
| ? | This page |
Saving pictures and animations
⤓ on the map opens Export. Save PNG saves the map as on screen. Its title is the saved setup's name ("(modified)" once you change something), or what the map shows if it is not a setup; below it are the valid time, T+ and run. Under the map are the colour scales and, on the right, what the map shows and the credits. Make GIF steps the map through a range of times (every step, or every 2nd or 4th) and saves them as an animated GIF at screen size (at most 960 pixels wide), with the frame time you pick; one palette serves every frame, so a colour means the same throughout. The map can't be used while it runs.
In the side panel, point at any chart (a plume, the box strip, the thirds map, the x–y plot, through time, a cyclone's pressure) and click ⤓ PNG: it saves the chart with its heading, where and when it is for, and, for the x–y plot, its key and correlation statistics.
Sharing a view
The address holds everything on screen (run, time, layers, box, point, cyclone, projection, map area), so copying it shares the view. "latest" in it follows the newest run.
Hovmöllers
The Hovmöller tab shows a latitude band's mean at each longitude through time, oldest at the top, so troughs and ridges show as sloping streaks: down to the right as they move east. Choose the band in 5° steps from 0 to 90°N and the longitudes in 10° steps; if west is east of east, the plot crosses the dateline. A map of the band sits underneath, on the same longitudes.
Analysis, then a forecast. Above the black line are ECMWF's analyses at 00Z and 12Z for the days before the run (the T+0 of its deterministic run, HRES, which starts from the analysis). Below it is the chosen run's ensemble mean every 6 hours to T+240, with the spread shaded grey. Pick an earlier run to see what it expected.
Departure from the time mean takes each longitude's mean over the whole plot off its values. Plain heights or pressures are dominated by the waves that stay put (the troughs over eastern Canada and Asia, the ridges over the Rockies and Europe), which hide the systems moving through. It is not an anomaly from climate: change the dates or the band and the baseline changes with them. Red is above the mean, blue below, green the zero line.
Spread shaded draws the values as contours over the grey spread. Untick it and the values themselves are shaded in colour instead, which draws the eye to the waves: 11 classes of a ColorBrewer diverging scale (red–blue, red–yellow–blue, spectral, brown–teal and others), the middle one around zero, with the interval chosen from the largest values unless you set one. Flip swaps the ends, for instance to make moist teal and dry brown. The spread is the band mean of each point's spread, so it is larger than the spread of the band mean would be. Daily means average each UTC day, which takes out the daily cycle of the near-surface fields.
Smooth shading interpolates between the cells (1° of longitude by 6 or 12 hours) as the contours do, so the colour bands have smooth edges; untick it to see the cells themselves. Height of a day keeps a day the same height however many days back you go (compact, normal or tall), and the plot scrolls when it is taller than the window, opening on the forecast's start; fit the window squeezes it all in instead. Save PNG saves the whole plot.
Forecast bias at a lead time. Each row is one run's ensemble mean at that lead time minus the analysis at the time it was for, so a column that stays red is a place the forecasts keep running too high, or too warm. The chart below gives, at each longitude over the period, the mean error, the RMS error and the spread: the spread should be about the size of the RMS error once the mean error is small. The rows fill in as the archive grows: a T+120 bias needs runs at least five days old.
| Fields | Levels |
|---|---|
| MSLP, total column water vapour | surface, column |
| Temperature, geopotential height | 850, 700, 500, 250 hPa (and height at 10 hPa) |
| u and v wind | 850, 700, 500, 250, 10 hPa |
| Divergent (irrotational) wind u and v; velocity potential χ | 850, 700, 500, 250 hPa; χ at 250 hPa |
| Rossby wave source | 250 hPa, 20–80°N |
Rossby wave source is where the divergent wind stirs up the vorticity enough to start Rossby waves: S = −ζaD − vχ·∇ζa, the absolute vorticity squeezed or stretched by the divergence, and carried across its own gradient by the divergent wind (Sardeshmukh and Hoskins 1988). Positive is a source of cyclonic (trough) and negative of anticyclonic (ridge) vorticity, and the strongest are usually where the outflow over tropical storms or convection meets the subtropical jet. It takes derivatives twice, so each member's is smoothed round each latitude (waves 0 to 20 kept, about the T21 of spherical-harmonic studies) before the band means. With only the northern hemisphere the divergent wind is rough near the equator, so it is shown from 20 to 80°N only. Checked against a whole-globe calculation (windspharm), the band means are within about 10–20% from 25°N to 75°N, but rougher at 20–25°N and 75–80°N. Runs and analyses archived before it was added (12 November 2026) don't have it, so those rows are blank.
The band means come from the 1° fields, weighted by the cosine of latitude. At 850 hPa, points below the ground in half the members or more (or in the analysis) are left out, so over the Rockies and Asia an 850 hPa band mean is of what is left. The divergent wind is solved over the northern hemisphere only, so it is rough near the equator (and the MJO, mostly south of 15°N, is beyond it).
The archive keeps, for each run, the ensemble mean and spread of every field in 5° bands at every step, and the analysis: about 30 MB a run. Runs older than 60 days are deleted. The analyses for the 60 days before the archive began were fetched from ECMWF's open data on Amazon S3; the forecasts start with the archive.
Saved setups. A Hovmöller setup keeps the field, the band, the longitudes, the mode and the display, and always opens on the latest run. Shared ones to start from: waves round the hemisphere (500 hPa height, 35–60°N), Rossby wave packets (250 hPa v wind, which shows them best without the time mean taken off), the Atlantic storm track (MSLP), Pacific moisture (total column water vapour) and the Rockies' 700 hPa temperature bias at day 5 (which fills in once the archive holds a week of runs with 700 hPa temperature). Links, downloads and imports work as on the maps (see Setups and colour scales).
Run-Run
The Run-Run tab shows what changed between two runs. It reads a slim archive of the last 10 runs of 00Z and 12Z: every member of a dozen fields on a 1° grid, to T+240. Runs are always compared at the same valid time, so the older run is at a longer lead time, and has more spread for that reason alone.
New and Old show the ensemble mean of either run. Overlay draws both runs as contours, the old run dashed, over the change in the mean, shaded as in the view below. A field without contours, such as precipitation, shows only the change.
Swipe across the map works with every view: drag the handle, or select it and use the arrow keys. With New or Old it puts the old run to the left and the new run to the right, in the same colours and contours. With a change in a probability, each side shows its own run's chance of the threshold. With the other views, the view is on the left and the new run's mean on the right, to show what the change sits in.
⤓ on the map saves it as a PNG, labelled with the field, the valid time, both runs and their lead times, and the area shown, with its colour scales below. ⤓ PNG on each chart in the Point box saves that chart, labelled with its latitude and longitude, the times it is valid for and the runs.
Change in the mean is the new mean minus the old, tested at each point with Welch's t-test on the 50 members of each run. Shift in spreads is the same change divided by the two runs' pooled spread, so it says whether a change is large for that place and lead time, and reads the same in any field; the same test decides what is shaded. Change in a probability is the new chance minus the old, in percentage points, tested with a two-proportion test. That test is rough when very few members are beyond the threshold.
Surprise shows where the new run's median falls among the old run's members, shaded where it is outside their 5th to 95th percentiles: the new run's central forecast is something the old run gave little room for. Blue is below the old members, red above, and the darkest colours are beyond almost all of them. A median is hard to move that far: against the previous run little is shaded after the first two days, and comparing with a run a few days older shows more. Ensembles have too little spread at short lead times, so some shading in the first day or two is to be expected.
Which changes are shaded. With 50 members the mean wobbles by about 0.2 of the spread by chance, and a 50% probability by about 14 points, so some differences mean nothing. Significant across the map, the default, uses a false discovery rate of 5%: of the points shaded, about 5% are expected to be false alarms. Significant at each point is looser: about 5% of all points pass by chance alone, however alike the runs are. Every difference turns the test off. The legend says how much of the map passes. Significant doesn't mean large: look at the size too.
Map. The map can be latitude–longitude or north polar stereographic, turned to any longitude. Both work with every view; the polar map has no break at the dateline, so the flow round the hemisphere reads as one.
Point. Click the map for both runs at that point. The top panel has each run's median at each valid time, the old run dashed, with the middle half of its members shaded. Below, on scales of their own: the new median minus the old, solid where the means differ significantly at this point (tested at each time, with no check across the times), and the new spread minus the old, as standard deviations. The older run is further ahead, so the new spread is usually the smaller; a bar above zero is the unusual one. Below these, dprog/dt follows one valid time through every run that reaches it, oldest on the left. A steady drift suggests the forecast is settling on something; a zig-zag suggests it is flip-flopping. Each run has a shorter lead time than the one before, so the spread should narrow from left to right. A point uses its nearest 1° value, so it won't match a 0.25° plume on the maps exactly.
Biggest changes ranks each 10° box at each valid time by the distance between the two runs' member box means (the Wasserstein distance: how far the members would have to move to turn one set into the other). This also picks up a change in spread or shape, which a change in the mean misses. Only boxes whose means differ significantly are listed (the same test as the maps, with the false discovery check across the boxes at each valid time), and only the largest change among those within 20° and 24 hours of each other. Averaging over a box removes noise, so a box can pass even where the map shades few points or none inside it. Size ranks in the field's units, which favours long lead times, where everything moves more. Size ÷ spread ranks against the spread, which favours short lead times, where a small change stands out. From and To keep the list to a window of the new run's lead times: for an event at T+84, say, T+12 to T+72 shows what changed in the run-up.
Fields. Winds are in knots. Strongest 100 m wind and heaviest 6 hour precipitation are the largest value at any 0.25° point in each 1° cell, so a narrow feature isn't averaged away. Unlike the maps, the archive has no record of where 850 hPa is below the ground, so 850 hPa fields aren't hatched there.
Questions about the forecast belong to the forecaster; this explorer shows what the ensemble does, not what will happen.
Clustering
The Clustering tab groups a run's 50 members by how alike their patterns are, over a region and a window of lead times. Each group is a scenario. The regions are the British Isles, Iberia, the Eastern Mediterranean and Levant, and Ukraine, on the run archive's 1° grid. A member's pattern is its field in the box at every 12 hours of the window, taken together, so a scenario means the same thing at every time you step through.
Scenarios are not separate outcomes. Grouping always gives groups. Across the 10 runs tested, 28 of 480 cases in these four regions stood out from a smooth spread of members, when chance alone would give about 19. Most of the time the members form one continuous spread, and the scenarios are a handy way to divide it and show its range. The member counts then depend on how the spread was divided, so don't read them as the chances of distinct outcomes.
The statement above the maps compares the grouping with 49 smooth spreads that have the same variance in each pattern and no groups. It says "A possible split" when the best grouping's silhouette (how tight and well apart the groups are) beats at least 48 of them and its smallest group has at least 5 members. About 1 case in 25 will say so by chance. It says "No clear split" otherwise, which isn't the same as no split: on made-up members with a typical region's variances, it found about 9 in 10 splits 4 standard deviations of the leading pattern apart, but only about 3 in 10 at 3.
You choose the number of scenarios, because when the spread is smooth there is no right number. The percentage in the statement is how much of the differences between members those scenarios hold.
Show chooses what the maps draw: any archive field, as each scenario's mean or its difference from the mean of all 50 members, so you can group on 500 hPa height and see what each scenario brings in precipitation or wind. Differences smaller than the first colour step are left clear. Contours add a second field on each map, as that scenario's mean: shade 24 hour precipitation and contour pressure, for example. Precipitation has no contours, so it can only be the shaded field. The coloured box is the region. Times outside the window still show the same members, but they were not grouped on those times.
Click a map for every member at that point in its scenario's colour, with each scenario's mean drawn thick and the window shaded. A scenario can look alike across the box and still differ at one point.
Run to run places the members of every run that reaches the same valid times in this run's scenarios, by whichever scenario each is nearest. A share growing across several runs says the forecast has been moving that way. Older runs are at longer lead times, so they have more spread for that reason alone and tend to place more members in the outer scenarios. The scenarios are also fitted to the chosen run's own members, so its bar is the best fit it can be, and a jump between it and the run before partly reflects that. Look for a trend across the older runs too.
⤓ PNG on each map, chart or the whole set saves it labelled with its region's latitudes and longitudes, or the point's, the valid time, the run and the statement above the maps.
Make picture
Make picture draws the scenarios on screen side by side as one PNG, using the tab's run, region, grouping field, lead times and number of scenarios, and the precipitation amount, accumulation period and gust you set above the button. It takes a few seconds, and nobody reviews the result, so treat it as a draft.
Headings give each scenario's letter and member count. These counts are not probabilities: they depend on how the spread was divided, and the line at the foot of the picture says whether the groups stood out from a smooth spread at all.
Blue areas are where at least half the scenario's members reach the precipitation amount over the chosen period at some time within the lead times: 15 mm in 24 hours unless you change it. The periods are 6, 12, 24 and 48 hours. Dashed outlines are the same for the gust, 40 kt unless you change it. Both are 1° cell averages, so hills aren't resolved and a gust at any one place can be well above the outline's value. A scenario of 8 members needs only 4 of them, and the picture says when shapes are rough. Gusts are read from the main store, which holds only the latest 3 runs, so an older run's picture leaves them out and says so. Precipitation is rain and snow together, with snow counted as the water it melts to: the data held has no snowfall field, so snow can't be shown apart, and the title says precipitation whether it falls as rain or snow.
The "L" and track are the deepest low in view in the scenario's mean pressure, followed through the window, whatever field the members were grouped on. Averaging members weakens a low and can place it between their tracks, and a low too near the edge of the view, or above 1015 hPa, isn't drawn. The text under each panel is worked out from the same numbers by fixed rules, with positions given only as parts of the region's box.